diff --git a/md/dev/JhzzvJnL9t/JhzzvJnL9t.md b/md/dev/JhzzvJnL9t/JhzzvJnL9t.md new file mode 100644 index 0000000000000000000000000000000000000000..41c3cb517ca6df4906ecb4c451d7a531659663dc --- /dev/null +++ b/md/dev/JhzzvJnL9t/JhzzvJnL9t.md @@ -0,0 +1,395 @@ +# Out-of-Domain Intent Detection Considering Multi-turn Dialogue Contexts + +Anonymous EMNLP submission + +# Abstract + +Out-of-Domain (OOD) intent detection is vital for practical dialogue systems, and it usually requires considering multi-turn dialogue contexts. However, most previous OOD intent detection approaches are limited to single dialogue turns. In this paper, we introduce a contextaware OOD intent detection (Caro) framework to model multi-turn contexts in OOD intent detection tasks. Specifically, we follow the information bottleneck principle to extract robust representations from multi-turn dialogue contexts. Two different views are constructed for each input sample and the superfluous information not related to intent detection is removed using a multi-view information bottleneck loss. Moreover, we also explore utilizing unlabeled data in Caro. A two-stage training process is introduced to mine OOD samples from these unlabeled data, and these OOD samples are used to train the resulting model with a bootstrapping approach. Comprehensive experiments demonstrate that Caro establishes state-of-theart performances on multi-turn OOD detection tasks by improving the F1-OOD score of over $2 9 \%$ compared to the previous best method. + +# 1 Introduction + +Intent detection is vital for dialogue systems (Chen et al., 2017). Recently, promising results have been reported for intent detection under the closed-world assumption (Shu et al., 2017), i.e., the training and testing distributions are assumed to be identical, and all testing intents are seen in the training process. However, this assumption may not be valid in practice (Dietterich, 2017), where a deployed system usually confronts an open-world (Fei and Liu, 2016; Scheirer et al., 2012), i.e., the testing distribution is subject to change and Out-of-Domain (OOD) intents that are not seen in the training process may emerge in testing. It is necessary to equip intent detection modules with OOD detection abilities to accurately classify seen In-Domain (IND) + +intents while rejecting unseen OOD intents (Yan et al., 2020a). + +Various methods are proposed to tackle the issue of OOD detection on classification problems (Geng et al., 2020). Existing approaches include using thresholds (Zhou et al., 2021) or $( k + 1 )$ -way classifiers ( $k$ is the number of IND classes) (Zhan et al., 2021). Promising results are reported to apply these OOD detection methods on intent detection modules (Zhou et al., 2022). However, most existing OOD intent detection studies only focus on singleturn inputs (Yan et al., 2020a; Lee and Shalyminov, 2019), i.e., only the most recently issued utterance is taken as the input. In real applications, completing a task usually necessitates multiple turns of conversations (Weld et al., 2021). Therefore, it is important to explicitly model multi-turn contexts when building OOD intent detection modules since users’ intents generally depend on turns of conversations (Qin et al., 2021). + +However, it is non-trivial to directly extend previous methods to the multi-turn setting (Ghosal et al., 2021). Specifically, we usually experience long distance obstacles when modeling multi-turn dialogue contexts, i.e., some dialogues have extremely long histories filled with irrelevant noises for intent detection (Liu et al., 2021). It is challenging to directly apply previous OOD intent detection methods under this obstacle since the learned representations may contain superfluous information that is irrelevant for intent detection tasks (Federici et al., 2019). + +Another challenge for OOD detection in multiturn settings is the absence of OOD samples in the training phase (Zeng et al., 2021a). Specifically, it is hard to refine learned representations for OOD detection without seeing any OOD training samples (Shen et al., 2021), and it is expensive to construct OOD samples before training, especially when multi-turn contexts are considered (Chen and Yu, 2021). Fortunately, unlabeled data (i.e., a mixture of IND and OOD samples) provide a convenient way to access OOD samples since these unlabeled data are almost “free” to collect from a deployed system. However, few studies have explored utilizing unlabeled data for OOD detection in the multi-turn setting. + +In this study, we propose a novel context-aware OOD intent detection framework Caro to address the above challenges for OOD intent detection in multi-turn settings. Specifically, we follow the information bottleneck principle (Tishby et al., 2000) to tackle the long-distance obstacle exhibited in multi-turn contexts. Robust representations are extracted by retaining predictive information while discarding superfluous information unrelated to intent detection. This objective is achieved by optimizing an unsupervised multi-view information bottleneck loss, during which two views are built based on the global pooling approach and adaptive reception fields. A gating mechanism is introduced to adaptively aggregate these two views to obtain an assembled representation. Caro also introduces a two-stage self-training scheme to mine OOD samples from unlabeled data. Specifically, the first stage builds a preliminary OOD detector with OOD samples synthesized from IND data. The second stage uses this detector to select OOD samples from the unlabeled data and use these samples to further refine the OOD detector. We list our key contributions: + +1. We propose a novel framework Caro to address a challenging yet under-explored problem of OOD intent detection considering multi-turn dialogue contexts. + +2. Caro learns robust representations by building diverse views of inputs and optimizing an unsupervised multi-view loss following the information bottleneck principle. Moreover, Caro mines OOD samples from unlabeled data to further refine the OOD detector. + +3. We extensively evaluate Caro on multi-turn dialogue datasets. Caro obtains state-of-the-art results, outperforming the best baseline by a large margin $( 2 9 . 6 \%$ in the F1-OOD score). + +et al., 2021a; Zhou et al., 2022; Wu et al., 2022) and use these representations to develop density-based or distance-based OOD detectors (Lee et al., 2018; Tan et al., 2019; Liu et al., 2020; Podolskiy et al., 2021). Some works also try to build OOD detectors with generated pseudo OOD samples (Hendrycks et al., 2018; Shu et al., 2021; Zhan et al., 2021; Marek et al., 2021) or thresholds based approaches (Gal and Ghahramani, 2016; Lakshminarayanan et al., 2017; Ren et al., 2019; Gangal et al., 2020; Ryu et al., 2017). + +Some OOD detection methods also make use of unlabeled data. Existing approaches either focus on utilizing unlabeled IND data (Xu et al., 2021; Jin et al., 2022) or adopting a self-supervised learning framework to handle mixtures of IND and OOD samples (Zeng et al., 2021b). These approaches do not explicitly model multi-turn contexts. + +Modeling Multi-turn Dialogue Contexts is the foundation for various dialogue tasks (Li et al., 2020; Ghosal et al., 2021; Chen et al., 2021). However, few works focus on detecting OOD intents in the multi-turn setting. Lee and Shalyminov (2019) proposed to use counterfeit OOD turns extracted from multi-turn contexts to train the OOD detector, and Chen and Yu (2021) augmented seed OOD samples that span multiple turns to improve the OOD detection performance. Nevertheless, these approaches either suffer from the long distance obstacle or require expensive annotated OOD samples. In this study, we attempt to learn robust representation by explicitly identifying and discarding superfluous information. + +Representation Learning is also related to our work. Recent approaches for representation learning include optimizing a contrastive loss (Caron et al., 2020; Gao et al., 2021) or maximizing the mutual information between features and input samples (Poole et al., 2019). However, these approaches cannot tackle the long distance obstacle exhibited in multi-turn contexts. In this study, we follow the information bottleneck principle (Tishby et al., 2000; Federici et al., 2019) to remove superfluous information from long contexts. + +# 2 Related Work + +OOD Detection is a widely investigated machine learning problem (Geng et al., 2020). Recent approaches try to improve the OOD detection performance by learning more robust representations on IND data (Zhou et al., 2021; Yan et al., 2020b; Zeng + +# 3 Problem Setup + +We start by formulating the problem: Given $k$ IND intent classes $\mathcal { T } = \{ I _ { i } \} _ { i = 1 } ^ { k }$ , we denote all samples that do not belong to these $k$ classes as the $( k + 1 )$ th intent $I _ { k + 1 }$ . Our training data contain a set of labeled IND samples ${ \mathcal { D } } _ { I } = \{ \langle x _ { i } , y _ { i } \rangle \}$ and a set of unlabeled samples $\mathcal { D } _ { U } = \{ \langle \widetilde { \pmb { x } } _ { i } , \widetilde { \pmb { y } } _ { i } \rangle \}$ , where $y _ { i } \in \mathcal { I }$ and $\widetilde { y _ { i } } \in \mathcal { T } \cup \{ I _ { k + 1 } \}$ is the label of input sample $\mathbf { \nabla } _ { \mathbf { x } _ { i } }$ and $\widetilde { \pmb { x } } _ { i }$ , respectively. $\widetilde { y _ { i } }$ labels are not observed eduring training. Our testing data contain a mixture of IND and OOD samples ${ \mathcal { D } } _ { T } = \{ \langle \widetilde { \pmb { x } } _ { i } , \widetilde { \ y } _ { i } \rangle \}$ , where $\widetilde { y _ { i } } \in \mathcal { T } \cup \left\{ I _ { k + 1 } \right\}$ . For a testing input $\widetilde { \pmb x }$ , our OOD intent detector aims to classify the intent label of $\widetilde { \pmb x }$ if it belongs to an IND intent or reject $\widetilde { \pmb x }$ if it belongs to the OOD intent $I _ { k + 1 }$ . We also assume a validation set ${ \mathcal { D } } _ { V }$ that only contains IND samples is available. Moreover, each input sample $_ { \pmb { x } }$ from $\mathcal { D } _ { I } , \mathcal { D } _ { U } , \mathcal { D } _ { V }$ , and $\mathcal { D } _ { T }$ consists of an utterance $\textbf { \em u }$ and a multi-turn dialogue history $\pmb { h } = \pmb { u } _ { 1 } , \ldots , \pmb { u } _ { t }$ $( t \geq 0 )$ prior of $\textbf { \em u }$ : ${ \pmb x } = \langle { \pmb h } , { \pmb u } \rangle$ . $\mathbf { \Delta } \mathbf { u } _ { i }$ is the utterance issued in each dialogue turn. + +# 4 Method + +Care tackles the OOD intent detection problem by training a $( k + 1 )$ -way classifier $F$ on $\mathcal { D } _ { I } \cup \mathcal { D } _ { U }$ Specifically, samples classified into the $( k + 1 )$ -th intent $I _ { k + 1 }$ are considered as OOD samples. There are mainly two challenges to be addressed in Caro: (1) How to alleviate the long distance obstacle and learn robust representations from multi-turn dialogue contexts; (2) How to effectively leverage unlabeled data for OOD intent detection. These two issues are tackled with two key ingredients in Caro (see Figure 1): 1. A multi-view information bottleneck method (Section 4.1); 2. A two-stage self-training scheme (Section 4.2). + +# 4.1 Multi-View Information Bottleneck + +The major challenge for learning robust representations from multi-turn dialogue contexts is the long distance obstacle, i.e., information that is irrelevant for intent detection may degenerate the extracted representation if the dialogue history $h$ becomes too long. In this study, we follow the information bottleneck principle (Tishby et al., 2000) to alleviate this issue, i.e., only the task-relevant information is retained in the extracted representations while all the superficial information is discarded. Specifically, we adopt a more general unsupervised multi-view setting for the information bottleneck method (Federici et al., 2019). For each input sample $\mathbf { \nabla } _ { \mathbf { x } _ { i } }$ , two semantic invariant views are constructed: $v _ { 1 } ( \pmb { x } _ { i } )$ , $v _ { 2 } ( \pmb { x } _ { i } )$ . These two views preserve the same task-relevant information (Zhao et al., 2017). The mutual information between $v _ { 1 } ( \pmb { x } _ { i } )$ and $v _ { 2 } ( \pmb { x } _ { i } )$ are maximized while the information not shared between $v _ { 1 } ( \pmb { x } _ { i } )$ and $v _ { 2 } ( \pmb { x } _ { i } )$ are eliminated. + +To achieve this goal, we adopt the multi-view information bottleneck loss introduced by Federici et al. (2019). + +Constructing Multiple Views for an input sample $_ { \pmb { x } }$ is the key to the success of the unsupervised information bottleneck method. In this study, we construct these two views $v _ { 1 } ( \pmb { x } _ { i } )$ , $v _ { 2 } ( \pmb { x } _ { i } )$ by adjusting the receptive fields of the final representation. This scheme is inspired by the observation in the neuroscience community that human brains process information with multiple receptive fields (Sceniak et al., 1999), i.e., the receptive field size for neurons is adapted based on input stimuli (Spillmann et al., 2015) so that different regions of inputs are emphasized (Pettet and Gilbert, 1992). This phenomenon has been demonstrated to be effective in modeling more robust features (Pandey et al., 2022) and inspired numerous successful neural models (Wang et al., 2021; Wei et al., 2017). + +Specifically, for each input sample ${ \pmb x } = \langle { \pmb h } , { \pmb u } \rangle$ , we first concatenate all utterances in $_ { \pmb { x } }$ and then use a pre-trained BERT model $E$ (Devlin et al., 2018) to encode the sequence of concatenated tokens into a sequence of embedding vectors $E ( { \pmb x } ) =$ $[ e _ { 1 } , \cdots e _ { n } ]$ , where $\boldsymbol { e } _ { i } \in \mathbb { R } ^ { m }$ . The following two strategies are used to construct two different views: + +1. Global Pooling builds view $v _ { 1 } ( \pmb { x } )$ with a mean-pooling layer on top of $[ e _ { 1 } , \cdots e _ { n } ]$ , $v _ { 1 } ( \pmb { x } )$ assumes each token embedding is equally weighted: + +$$ +v _ { 1 } ( { \pmb x } ) = \sum _ { i = 1 } ^ { n } e _ { i } / n +$$ + +2. Adaptive Reception Field builds view $v _ { 2 } ( { \pmb x } )$ by adapting the synaptic weight of each token embedding based on the input $_ { \pmb { x } }$ : + +$$ +\begin{array} { r } { v _ { 2 } ( \pmb { x } ) = \displaystyle \sum _ { i = 1 } ^ { n } \frac { \exp ( \alpha _ { i } ) } { \sum _ { j = 1 } ^ { n } \exp ( \alpha _ { j } ) } \cdot \pmb { e } _ { i } } \\ { \alpha _ { i } = \sigma ( \pmb { w } _ { i } \cdot \mathrm { R e L U } ( \pmb { W } _ { 1 } \cdot \pmb { s } ) ) , } \end{array} +$$ + +where $\textbf { \textit { s } } \in \mathbb { R } ^ { n m }$ is the concatenation of all $n$ embeddings $[ e _ { 1 } , \cdots e _ { n } ]$ . $\sigma$ is the Sigmoid activation function. ${ \pmb w } _ { i } \in \mathbb { R } ^ { 1 \times r _ { 1 } } ( i = 1 , \cdot \cdot \cdot n )$ and $\pmb { W } _ { 1 } \in \mathbb { R } ^ { r _ { 1 } \times n m }$ are learnable parameters. $r _ { 1 }$ is the size of the intermediate layer. Moreover, to enhance the generalization ability, we set a small value for $r _ { 1 }$ in our implementation to form a bottleneck structure in the weighting function (Hu et al., 2017). + + +Figure 1: Framework of Caro. For each input sample ${ \pmb x } = \langle { \pmb h } , { \pmb u } \rangle$ , two views $v _ { 1 } ( \pmb { x } )$ and $v _ { 2 } ( { \pmb x } )$ are obtained and a multi-view information bottleneck loss $\mathcal { L } _ { I B }$ is optimized to learn robust representations. A two-stage training process is introduced to mine OOD samples $\mathcal { D } _ { O }$ from unlabeled data $\mathcal { D } _ { U }$ , and optimize the cross entropy loss $\mathcal { L } _ { C E }$ with $\mathcal { D } _ { O } \cup \mathcal { D } _ { I }$ + +Optimizing Information Bottleneck is performed in an unsupervised setting based on the two views of each sample. Specifically, we assume the representation $z _ { i }$ of each view $v _ { i } ( { \pmb x } )$ , $( i = 1 , 2 )$ ) follows a distribution that is parameterized by an encoder $p ( z | v _ { i } )$ , where $v _ { i }$ is short for ${ v } _ { i } ( { \pmb x } )$ for abbreviation. To facilitate the computation, we model $p$ as factorized Gaussian distributions, i.e., $p ( z | v _ { i } ) = \mathcal { N } [ \mu ( v _ { i } ) , \Sigma ( v _ { i } ) ]$ , in which $\mu ( v _ { i } )$ and $\Sigma ( v _ { i } )$ are two neural networks that produce the mean and deviation, respectively. The following information bottleneck loss (Federici et al., 2019) is optimized to remove superfluous information in $v _ { 1 } ( \pmb { x } )$ and $v _ { 2 } ( { \pmb x } )$ : + +$$ +\begin{array} { r } { \mathcal { L } _ { I B } = - I ( z _ { 1 } ; z _ { 2 } ) + \displaystyle \frac { 1 } { 2 } ( D _ { K L } [ p ( z | \pmb { v } _ { 1 } ) | | p ( z | \pmb { v } _ { 2 } ) ] } \\ { + D _ { K L } [ p ( z | \pmb { v } _ { 2 } ) | | p ( z | \pmb { v } _ { 1 } ) ] ) , } \end{array} +$$ + +where $I$ calculates the mutual information of two random variables, and $D _ { K L }$ calculates the KL divergence between two distributions. + +# 4.2 Two-stage Self-training + +Although robust representations can be obtained with the help of the information bottleneck loss $\mathcal { L } _ { I B }$ from Section 4.1, we still lack the annotations for OOD samples to train the $( k + 1 )$ -way classifier $F$ for OOD detection. In this study, we tackle this issue with a two-stage self-training process, which mines OOD samples from the unlabeled data $\mathcal { D } _ { U }$ with a bootstrapping approach. Moreover, for each input sample $_ { \pmb { x } }$ , we also aggregate its two views $v _ { 1 } ( \pmb { x } )$ and $v _ { 2 } ( { \pmb x } )$ with a dynamic gate to obtain assembled representations in training. + +Stage One synthesizes pseudo OOD samples $\mathcal { D } _ { P }$ by mixing up IND features. Specifically, samples from $\mathcal { D } _ { I }$ are first mapped into IND representation vectors, and pseudo OOD samples are obtained as convex combinations of these vectors (Zhan et al., 2021). A preliminary OOD detector $F$ is trained using the classical cross-entropy loss $\mathcal { L } _ { C E }$ on these synthesized pseudo OOD samples and labeled IND samples $\mathcal { D } _ { I }$ . This stage endows $F$ with a preliminary ability to predict the intent distribution of each input sample. + +Stage Two predicts a pseudo label for each sample $\mathbf { \boldsymbol { x } } \in \mathcal { D } _ { U }$ using $F$ , and then collects samples that are assigned with the OOD label $I _ { k + 1 }$ as a set of mined OOD samples $\mathcal { D } _ { O }$ . With the help of $\mathcal { D } _ { O }$ we further train the classifier $F$ on the following loss: + +$$ +\mathcal { L } = \underset { \pmb { x } \in \mathcal { D } _ { I } \cup \mathcal { D } _ { O } } { \mathbb { E } } \mathcal { L } _ { C E } + \lambda \underset { \pmb { x } \in \mathcal { D } _ { U } } { \mathbb { E } } \mathcal { L } _ { I B } +$$ + +where $\lambda$ is a scalar hyper-parameter to control the weight of the information bottleneck loss. + +Multi-view Aggregation is performed to obtain assembled representations for input samples. Specifically, whenever we need to extract the representation $v ( { \pmb x } )$ for an input sample $_ { \pmb { x } }$ in the training process, we use the following aggregation approach: + +$$ +\begin{array} { c } { { v ( { \pmb x } ) = \beta \otimes v _ { 1 } ( { \pmb x } ) + ( 1 - \beta ) \otimes v _ { 2 } ( { \pmb x } ) } } \\ { { \beta = \sigma ( \mathbf { W } _ { 3 } \cdot \mathrm { R e L U } ( \mathbf { W } _ { 2 } \cdot ( v _ { 1 } ( { \pmb x } ) + v _ { 2 } ( { \pmb x } ) ) ) ) } } \end{array} +$$ + +where $\otimes$ represents the element-wise product, $\mathbf { W } _ { 2 } \in \mathbb { R } ^ { r _ { 2 } \times m }$ and $\mathbf { W } _ { 3 } \in \mathbb { R } ^ { m \times r _ { 2 } }$ are learnable parameters. $r _ { 2 }$ is the size of the intermediate layer. + +The training of Caro is given in Algorithm 1. + +# 5 Experiments + +# 5.1 Datasets + +We perform experiments on two variants of the STAR dataset (Mosig et al., 2020), i.e., STAR-Full and STAR-Small. Specifically, STAR is a taskoriented dialogue dataset that has 150 intents. It is designed to model long context dependence, and provides explicit annotations of OOD intents. Following Chen and Yu (2021), we regard samples + +Input: IND data $\mathcal { D } _ { I }$ , unlabeled data $\mathcal { D } _ { U }$ . Output: A trained OOD detector $F$ . // Stage 1 +1 Synthesize pseudo OOD samples $\mathcal { D } _ { P }$ by mixing up IND representations. +2 Train $F$ using the cross-entropy loss $\mathcal { L } _ { C E }$ on $\mathcal { D } _ { I } \cup \mathcal { D } _ { P }$ . // Stage 2 +3 Mine OOD samples $\mathcal { D } _ { O }$ from $\mathcal { D } _ { U }$ using $F$ . +4 Train $F$ using $\mathcal { L }$ (Eq. 4) on $\mathcal { D } _ { I }$ , $\mathcal { D } _ { O }$ , and $\mathcal { D } _ { U }$ + +Algorithm 1: The training process of Caro +Table 1: Dataset statistics. + +
| Train | Valid Test Dv DT | #Avg.( Context Turns | ||
| D1 | Du | |||
| STAR-Full | 15.4K | 7.9K | 2.8K 2.9K | 6.13 |
| STAR-Small | 7.7K | 3.9K | 2.8K 2.9K | 6.12 |
| Model | STAR-Full | STAR-Small | |||||
| Oracle | F1-All 50.1 | F1-OOD 64.46 | F1-IND 50 | F1-All 46.54 | F1-OOD 58.23 | F1-IND 46.46 | |
| D1 | MSP | 40.83 | 19.74 | 40.97 | 37.17 | 18.1 | 37.31 |
| MSP w/o h | 17.29 | 14.12 | 17.31 | 17.12 | 13.49 | 17.14 | |
| SEG | 17.45 | 6.85 | 17.53 | 11.66 | 7.39 | 11.69 | |
| SEG w/o h | 0.06 | 2.77 | 0.04 | 0.05 | 2.27 | 0.04 | |
| DOC | 26.53 | 16.80 | 26.60 | 3.47 | 11.78 | 3.41 | |
| DOC w/o h | 11.31 | 14.16 | 11.29 | 0.08 | 11.04 | 0 | |
| ADB | 44.64 | 20.56 | 44.80 | 41.36 | 18.23 | 41.51 | |
| ADB w/o h | 23.27 | 17.63 | 23.30 | 20.08 | 21.27 | 20.07 | |
| DAADB | 37.27 | 22.87 | 37.37 | 34.81 | 20.43 | 34.91 | |
| DAADB w/o h | 17.87 | 15.15 | 17.88 | 16.34 | 17.03 | 16.33 | |
| Outlier | 43.84 | 19.53 | 44.01 | 39.51 | 19.92 | 39.64 | |
| Outlier w/o h | 23.35 | 16.75 | 23.39 | 19.56 | 15.42 | 19.59 | |
| CDA | 43.76 | 5.26 | 44.03 | 40.02 | 10.48 | 40.22 | |
| D1+Du | ASS+MSP | 41.97 | 25.15 | 42.08 | 40.85 | 19.47 | 40.99 |
| ASS+LOF | 39.87 | 17.65 | 40.02 | 39.54 | 18.49 | 39.68 | |
| ASS+GDA | 43.73 | 21.24 | 43.88 | 40.86 | 16.72 | 41.02 | |
| Caro (ours) | 48.75(±1.0) | 54.75(±3.2) | 48.71(±1.0) | 45.02(±1.1) | 46.78(±1.8) | 45.01(±1.1) | |
| Model | STAR-Full F1-Al1 F1-00D F1-IND | STAR-Small F1-All F1-OOD F1-IND | ||||
| Caro | 48.75 | 54.75 | 48.71 | 45.02 | 46.78 | 45.01 |
| w/o Du | 45.97 | 21.45 | 46.14 | 42.24 | 23.23 | 42.37 |
| w/o MV | 47.71 | 53.35 | 47.67 | 44.42 | 38.89 | 44.46 |
| w/o VA | 47.34 | 50.85 | 47.32 | 44.14 | 43.88 | 44.15 |
| w/o IB | 48.23 | 49.37 | 48.22 | 44.14 | 37.06 | 44.19 |
| Model | STAR-Full F1-All F1-OOD F1-IND | STAR-Small F1-All F1-OOD F1-IND | ||||
| Caro | 48.75 | 54.75 | 48.71 | 45.02 | 46.78 | 45.01 |
| InfoMax | 47.27 | 49.92 | 47.25 | 44.27 | 36.66 | 44.32 |
| MVI | 48.46 | 51.99 | 48.44 | 44.70 | 36.16 | 44.76 |
| CL | 48.18 | 52.54 | 48.15 | 44.59 | 35.31 | 44.65 |
| SimCSE | 47.73 | 47.74 | 47.73 | 44.30 | 27.02 | 44.42 |
| Context Len | F1-All | F1-00D | F1-IND | |
| Long | w/o IB | 44.14 | 37.06 | 44.19 |
| wIB | 45.02 (+0.88) | 46.78 (+9.72) | 45.01 (+0.82) | |
| Short | w/o IB | 43.61 | 40.68 | 43.63 |
| wIB | 43.70 (+0.09) | 43.32 (+2.64) | 43.70 (+0.07) | |
| Methods | #Para. | Training Time | Testing Time |
| Outlier | 111.47 M | 7.26 min | 14.46 s |
| Caro | 116.80 M | 8.75 min | 14.53 s |
| DownSample-Rate | F1-All | F1-OOD | F1-IND |
| 100%75%50%25% | 45.02 | 46.78 | 45.01 |
| 44.40 | 37.77 | 44.44 | |
| 45.04 | 30.13 | 45.15 | |
| 44.47 | 20.76 | 44.62 |
| 入 | F1-All | F1-0OD | F1-IND |
| 0.3 | 47.57 | 55.68 | 47.51 |
| 0.4 | 48.10 | 51.44 | 48.08 |
| 0.5 | 48.75 | 54.75 | 48.71 |
| 0.6 | 47.15 | 50.84 | 47.12 |
| 0.7 | 47.83 | 50.60 | 47.81 |
| Dialogue | Prediction of “Outlier” | Prediction of “Caro” | |
| Customer | Agent | ||
| Somebody has transferred $5oo from my account. | bank_ask_pin | OOD | |
| yes please | Would you like to file a fraud report? Could I get your full name,please? | ||
| Jane Doe | |||
| Can you tell me your account number, please? | |||
| What am I doing on this chat? Hey there! Can you tell me what the weather will be tomorrow please? | weather _inform | OOD | |
| oh dang when's the next time it will be sunny? | It will be Raining all day on Saturday in Los Angeles,with temperatures of around 19 degrees celsius. | _forecast | |
| It will be Sunny all day on Friday in Los Angeles,with temperatures of around 11 degrees celsius. | |||
| Nice! I think I'll head to La brea tar pits. I've never been there.have you seen those animals? Crazy | bank_inform | ||
| apartment for Thursday at 8pm my name is john | |||
| Ican't remember | Have you already paid the application fee for the apartment? | _cannot _authenticate | OOD |
| Iam sorry, but there is no viewing available at your preferred time. | |||
| Ihave a hot date by 8pmo Thursday | When would you like the viewing to start? | ||
| Model | IS↑ | FID↓ | Recall↑ | NFE↓ | Time (s)↓ |
| Denoising Diffusion GAN (ours), T=4 | 9.63 | 3.75 | 0.57 | 4 | 0.21 |
| DDPM (Ho et al., 2020) | 9.46 | 3.21 | 0.57 | 1000 | 80.5 |
| NCSN (Song & Ermon,2019) | 8.87 | 25.3 | 1 | 1000 | 107.9 |
| Adversarial DSM (Jolicoeur-Martineau et al., 2021b) | 1 | 6.10 | = | 1000 | - |
| Likelihood SDE (Song et al., 2021b) | = | 2.87 | - | - | - |
| Score SDE(VE) (Song et al.,2021c) | 9.89 | 2.20 | 0.59 | 2000 | 423.2 |
| Score SDE(VP) (Song et al.,2021c) | 9.68 | 2.41 | 0.59 | 2000 | 421.5 |
| Probability Flow (VP) (Song et al.,2021c) | 9.83 | 3.08 | 0.57 | 140 | 50.9 |
| LSGM(Vahdat et al.,2021) | 9.87 | 2.10 | 0.61 | 147 | 44.5 |
| DDIM, T=50 (Song et al., 2021a) | 8.78 | 4.67 | 0.53 | 50 | 4.01 |
| FastDDPM,T=50 (Kong& Ping,2021) | 8.98 | 3.41 | 0.56 | 50 | 4.01 |
| Recovery EBM (Gao et al.,2021) | 8.30 | 9.58 | - | 180 | - |
| Improved DDPM (Nichol & Dhariwal, 2021) | - | 2.90 | - | 4000 | |
| VDM (Kingma et al., 2021) | - | 4.00 | 1000 | ||
| UDM (Kim et al., 2021) | 10.1 | 2.33 | 2000 | ||
| D3PMs (Austin et al., 2021) | 8.56 | 7.34 | 1000 | ||
| Gotta Go Fast (Jolicoeur-Martineau et al., 2021a) | - | 2.44 | - | 180 | |
| DDPMDistillation (Luhman & Luhman, 2021) | 8.36 | 9.36 | 0.51 | 1 | |
| SNGAN (Miyato et al., 2018) | 8.22 | 21.7 | 0.44 | 1 | |
| SNGAN+DGflow (Ansari et al., 2021) | 9.35 | 9.62 | 0.48 | 25 | 1.98 |
| AutoGAN (Gong et al., 2019) | 8.60 | 12.4 | 0.46 | 1 | - |
| TransGAN (Jiang et al., 2021) | 9.02 | 9.26 | - | 1 | 1 |
| StyleGAN2 w/o ADA (Karras et al., 2020a) | 9.18 | 8.32 | 0.41 | 1 | 0.04 |
| StyleGAN2 w/ ADA (Karras et al., 2020a) | 9.83 | 2.92 | 0.49 | 1 | 0.04 |
| StyleGAN2 w/Diffaug (Zhao et al., 2020) | 9.40 | 5.79 | 0.42 | 1 | 0.04 |
| Glow (Kingma & Dhariwal, 2018) | 3.92 | 48.9 | - | 1 | - |
| PixelCNN (Oord et al.,2016b) | 4.60 | 65.9 | - | 1024 | - |
| NVAE(Vahdat&Kautz,2020) | 7.18 | 23.5 | 0.51 | 1 | 0.36 |
| IGEBM(Du & Mordatch,2019) | 6.02 | 40.6 | 1 | 60 | - |
| VAEBM (Xiao et al., 2021) | 8.43 | 12.2 | 0.53 | 16 | 8.79 |
| Model Variants | IS↑ | FID↓ | Recall个 |
| T=1 | 8.93 | 14.6 | 0.19 |
| T=2 | 9.80 | 4.08 | 0.54 |
| T=4 | 9.63 | 3.75 | 0.57 |
| T=8 | 9.43 | 4.36 | 0.56 |
| One-shot w/ aug | 8.96 | 13.2 | 0.25 |
| Direct denoising | 9.10 | 6.03 | 0.53 |
| Noise generation | 8.79 | 8.04 | 0.52 |
| No latent variable | 8.37 | 20.6 | 0.42 |
| Model | Modes↑ | KL↓ |
| VEEGAN (Srivastava et al.) | 762 | 2.173 |
| PacGAN (Lin et al.) | 992 | 0.277 |
| PresGAN (Dieng et al.) | 1000 | 0.115 |
| InclusiveGAN (Yu et al.) | 997 | 0.200 |
| StyleGAN2 (Karras et al.) | 940 | 0.424 |
| Adv. DSM (Jolicoeur-Martineau et al.) | 1000 | 1.49 |
| VAEBM (Xiao et al.) | 1000 | 0.087 |
| Denoising Diffusion GAN (ours) | 1000 | 0.071 |
| Model | FID↓ |
| Denoising Diffusion GAN (ours) | 7.64 |
| Score SDE (Song et al., 2021c) LSGM (Vahdat et al., 2021) UDM (Kim et al., 2021) | 7.23 7.22 7.16 |
| NVAE(Vahdat & Kautz, 2020) VAEBM (Xiao et al., 2021) | 29.7 20.4 |
| NCP-VAE (Aneja et al.,2021) | 24.8 |
| PGGAN (Karras et al., 2018) Adv.LAE (Pidhorskyi et al., 2020) | 8.03 |
| 19.2 | |
| VQ-GAN (Esser et al., 2021b) DC-AE (Parmar et al., 2021) | 10.2 15.8 |
| Model | FID↓ |
| Denoising Diffusion GAN (ours) | 5.25 |
| DDPM (Ho et al., 2020) | 7.89 |
| ImageBART (Esser et al., 2021a) Gotta Go Fast (Jolicoeur-Martineau et al.) | 7.32 25.67 |
| PGGAN (Karras et al., 2018)) | 6.42 |
| StyleGAN (Karras et al., 2019) | 4.21 |
| StyleGAN2 (Karras et al., 2020b) | 3.86 |
| CIPS (Anokhin et al., 2021) | 2.92 |
| CIFAR10 | CelebA-HQ | LSUNChurch | |
| # of ResNet blocks per scale | 2 | 2 | 2 |
| Initial #of channels | 128 | 64 | 64 |
| Channel multiplier for each scale | (1,2,2,2) | (1,1,2,2,4,4) | (1,1,2,2,4,4) |
| Scale of attention block | 16 | 16 | 16 |
| LatentDimension | 100 | 100 | 100 |
| # of latent mapping layers | 3 | 3 | 3 |
| Latent embedding dimension | 256 | 256 | 256 |
| CelebA-HQ and LSUN Church |
| 1 × 1 conv2d,128ResBlock down, 256ResBlock down, 512ResBlock down, 512 |
| ResBlock down, 512 |
| ResBlock down, 512 |
| ResBlock down, 512 |
| minibatch std layerGlobal Sum PoolingFC layer→ scalar |
| CIFAR-10 |
| 1 × 1 conv2d,128 ResBlock,128 |
| ResBlock down, 256 ResBlock down, 512 ResBlock down, 512 |
| minibatch std layer |
| Global Sum Pooling FC layer → scalar |
| CIFAR10 | CelebA-HQ | LSUN Church | |
| Initial learning rate for discriminator | 10-4 | 10-4 | 10-4 |
| Initial learning rate for generator | 1.6 × 10-4 | 1.6 × 10-4 | 2×10-4 |
| Adam optimizer β1 | 0.5 | 0.5 | 0.5 |
| Adam optimizer β2 | 0.9 | 0.9 | 0.9 |
| EMA | 0.9999 | 0.999 | 0.999 |
| Batch size | 256 | 32 | 64 |
| # of training iterations | 400k | 750k | 600k |
| # of GPUs | 4 | 8 | 8 |
| Model | 0 | SST-2 | MNLI | QNLI | MRPC | QQP | STSB | COLA | AVG |
| BERTBase,Subword | 110M | 92.7 | 84.4/- | 88.4 | 86.7/- | - | - | - | - |
| T5Base,Subword | 220M | 92.7 | 84.2/84.6 | 90.5 | 88.9/92.1 | 91.6/88.7 | 88.0 | 53.8 | 84.3 |
| Byte-level T5 Base | 200M | 91.6 | 82.5/82.7 | 88.7 | 87.3/91.0 | 90.9/87.7 | 84.3 | 45.1 | 81.5 |
| Byte-level T5+ConV Base | 205M | 89.8 | 81.1/82.5 | 89.2 | 83.6/89.2 | 90.7/87.7 | 85.0 | 47.1 | 81.2 |
| Byte-level T5+LASCBase | 205M | 90.0 | 80.0/80.8 | 87.1 | 82.8/88.1 | 89.0/85.4 | 83.7 | 25.3 | 77.0 |
| CHARFORMER Base | 203M | 91.6 | 82.6/82.7 | 89.0 | 87.3/91.1 | 91.2/88.1 | 85.3 | 42.6 | 81.4 |
| Byte-level T5s Base | 133M | 91.2 | 83.9/83.7 | 90.9 | 85.5/89.2 | 91.1/88.1 | 85.7 | 49.3 | 82.6 |
| CHARFORMERSBase | 134M | 91.5 | 83.7/84.4 | 91.0 | 87.5/91.4 | 91.4/88.5 | 87.3 | 51.8 | 83.6 |
| Model | Civil Comments | Wiki Comments |
| T5Base,Subword | 81.2/- | 91.5/- |
| Byte-level T5 Base | 82.8/78.7 | 93.2/75.4 |
| Byte-level T5+LASCBase | 82.9 / 78.2 | 93.0 / 75.0 |
| CHARFORMER Base | 83.0/78.8 | 92.7 / 79.7 |
| CHARFORMER SBase | 83.0 /78.9 | 93.5 /75.5 |
| Model | IMDb | News |
| T5 Base,Subword | 94.2 | 93.5 |
| Byte-level T5 Base | 91.5 | 93.6 |
| Byte-level T5+LASC Base | 91.1 | 93.5 |
| CHARFORMER Base | 91.5 | 94.0 |
| CHARFORMER SBase | 94.4 | 94.1 |
| In-Language | Translate-Train-All | Zero-Shot | ||||||
| Model | |0 | TyDiQA-GoldP | XQuAD | MLQA | XNLI | PAWS-X | XNLI | PAWS-X |
| mBERTBase (Subword) | 179M | 77.6/68.0 | -- | -- | - | 1 | 65.4 | 81.9 |
| mT5 Base (Subword) | 582M | 80.8/70.0 | 75.3/59.7 | 67.6/48.5 | 75.9 | 89.3 | 75.4 | 86.4 |
| Byte-level T5 Base | 200M | 75.6/65.4 | 68.6/54.3 | 61.8/44.4 | 69.4 | 87.1 | 57.4 | 80.9 |
| Byte-level T5+LASCBase | 205M | 70.6/59.7 | 66.8/52.1 | 58.8/41.1 | 67.9 | 84.8 | 55.2 | 79.0 |
| CHARFORMER Base | 203M | 75.9/65.6 | 70.2/55.9 | 62.6/44.9 | 71.1 | 87.2 | 57.6 | 81.6 |
| CHARFORMER S Base | 134M | 79.1/68.8 | 73.6/59.0 | 66.3/48.5 | 72.2 | 88.2 | 66.6 | 85.2 |
| CHARFORMER s Base,LongPT | 134M | 81.2/71.3 | 74.2/59.8 | 67.2/49.4 | 72.8 | 88.6 | 67.8 | 83.7 |
| Model | Batch Size | L | ds | |0 | Speed (steps/s) | FLOPS |
| mT5 Base (Subword) | 1024 | 1024 | - | 582M | 1.54 | 1.3 × 1015 |
| CHARFORMERS Base | 1024 | 2048 | 2 | 134M | 1.98 | 4.3 ×1014 |
| CHARFORMERs Base,LongPT | 2048 | 2048 | 2 | 134M | 1.01 | 4.3× 1014 |
| Model | L | ds | 0 | Speed (steps/s) | FLOPS | Peak Mem. |
| T5Base :(Subword) | 512 | - | 220M | 9.3 | 1.1 × 1013 | = |
| Byte-level T5 Base | 1024 | 1 | 200M | 8.2 | 2.9×1013 | 3.09GB |
| Byte-level T5+LASC Base | 1024 | 4 | 205M | 15 | 9.9 ×1012 | 1.62GB |
| CHARFORMER Base | 1024 | 2 | 206M | 11 | 1.6 ×1013 | 1.95GB |
| CHARFORMER Base | 1024 | 3 | 203M | 15 | 1.1 × 1013 | 1.63GB |
| CHARFORMERSBase | 1024 | 2 | 134M | 14 | 1.3 × 1013 | 1.73GB |
| CHARFORMERSBase | 1024 | 3 | 134M | 20 | 8.7×1012 | 1.34GB |
| Ablation | ds | Size | SST-2 | MNLImm | IMDb |
| Offsets | S | 89.11 | 79.50 | 90.49 | |
| Conv | S | 89.11 | 79.65 | 90.63 | |
| Conv + BC | S | 89.56 | 80.15 | 90.60 | |
| Conv + Offsets + BC | 2222 | S | 89.11 | 79.68 | 90.48 |
| Conv | 34 | S | 89.45 | 80.07 | 90.15 |
| Conv | S | 89.11 | 79.82 | 90.21 | |
| Conv | 234 | B | 90.60 | 82.92 | 91.46 |
| Conv | B | 91.40 | 82.74 | 91.46 | |
| Conv | B | 91.40 | 82.67 | 92.33 |
| Model | ds | Freeze GBST | XNLI (Zero) | XNLI(Translate) | PAWS-X (Zero) | PAWS-X(Translate) |
| CHARFORMER Small | 2 | No | 44.5 | 62.7 | 27.9 | 37.5 |
| CHARFORMERSmall | 2 | Yes | 50.9 | 68.7 | 77.1 | 84.8 |
| CHARFORMER Small | 3 | No | 47.9 | 67.9 | 29.5 | 36.8 |
| CHARFORMER Small | 3 | Yes | 43.2 | 68.6 | 77.8 | 83.7 |
| CHARFORMER Small | 4 | No | 47.5 | 47.5 | 30.9 | 36.9 |
| CHARFORMER Small | 4 | Yes | 43.6 | 43.6 | 77.9 | 83.5 |
| Model | ds | TyDiQA-GoldP F1 |
| CHARFORMERSmall | 2 | 69.6 |
| CHARFORMERSmall | 3 | 68.1 |
| CHARFORMERSmall | 4 | 66.6 |
| Byte-level T5+LASCsmall | 4 | 64.9 |
| CHARFORMER Base | 2 | 75.8 |
| CHARFORMERBase | 3 | 74.3 |
| CHARFORMERBase | 4 | 73.2 |
| Byte-level T5+LASCBase | 4 | 70.6 |
| Model | TyDiQA-GoldP F1 /EM |
| mT5Large | 85.3 /75.3 |
| ByT5 Large | 87.7/79.2 |
| CHARFORMER* | 86.3 / 77.3 |
| Model | L | ds | |0 | Speed (steps/s) | FLOPS |
| Byte-level T5 Base | 2048 | 1 | 200M | 2.7 | 2.0×1013 |
| Byte-level T5+LASC Base | 2048 | 4 | 205M | 11 | 5.5× 1012 |
| CHARFORMERBase | 2048 | 2 | 203M | 6.1 | 9.5×1012 |
| CHARFORMERBase | 2048 | 3 | 203M | 10 | 6.5×1012 |
| CHARFORMERSBase | 2048 | 2 | 134M | 6.1 | 9.2 ×1012 |
| Model | 0 | ar | bn | en | fi | id | ko | ru | SW | te | Avg. |
| mBERTBase (Subword) | 179M | -/- | / | - | -/- | -/- | - | -/- | / | -/- | 77.6/68.0 |
| mT5 Base (Subword) | 582M | 84.2/71.8 | 80.0/69.0 | 76.6/65.2 | 80.1/69.3 | 85.5/75.0 | 70.3/61.6 | 77.5/64.4 | 83.6/74.9 | 88.2/78.0 | 80.8/70.0 |
| Byte-level T5 Base | 200M | 81.4/67.0 | 66.8/56.6 | 69.8/59.5 | 75.6/63.0 | 81.6/72.4 | 64.6/58.7 | 74.1/60.8 | 81.8/74.3 | 85.0/76.1 | 75.6/65.4 |
| Byte-level T5+LASCBase | 205M | 78.1/62.3 | 61.1/50.4 | 66.7/55.2 | 72.5/60.4 | 79.9/68.3 | 51.5/43.5 | 70.4/58.7 | 74.7/67.5 | 80.2/71.2 | 70.6/59.7 |
| CHARFORMER Base | 203M | 81.8/67.9 | 69.1/60.2 | 71.4/60.5 | 76.3/64.2 | 83.0/73.1 | 62.7/54.3 | 74.7/61.7 | 80.2/73.3 | 83.6/75.0 | 75.9/65.6 |
| CHARFORMERSBase | 134M | 82.4/68.1 | 78.1/67.3 | 75.4/64.3 | 79.5/68.2 | 85.0/75.9 | 66.6/58.0 | 77.0/64.3 | 81.5/74.1 | 86.5/78.6 | 79.1/68.8 |
| CHARFORMERSBase,LongPT | 134M | 85.7/74.5 | 78.7/67.3 | 76.8/65.9 | 81.9/70.6 | 86.7/79.1 | 69.4/61.6 | 79.2/67.1 | 83.7/75.2 | 88.8/80.6 | 81.2/71.3 |
| Model | 间 | ar | de | el | en | es | hi | ru | th | tr | vi | zh | Avg. | |
| mT5 Base (Subword) | 582M | 72.4/55.2 | 76.9/59.7 | 76.8/58.8 | 83.1/70.3 | 79.0/61.2 | 71.4/53.4 | 76.1/58.5 | 67.9/62.0 | 72.5/51.4 | 75.9/56.3 | 76.9/69.7 | 75.3/59.7 | |
| Byte-levelT5 Base | 200M | 64.8/47.9 | 74.3/58.3 | 69.2/51.8 | 81.5/70.4 | 77.2/60.4 | 67.0/51.5 | 72.3/55.5 | 48.3/41.9 | 69.6/51.7 | 73.3/54.4 | 57.3/53.3 | 68.6/54.3 | |
| Byte-level T5+LASCBase | 205M | 62.9/45.5 | 70.6/54.2 | 68.3/52.3 | 80.1/68.4 | 74.8/57.9 | 63.1/46.2 | 68.2/52.2 | 50.0/43.4 | 67.1/48.2 | 71.7/51.8 | 57.7/52.7 | 66.8/52.1 | |
| CHARFORMER Base | 203M | 65.7/49.8 | 74.2/58.0 | 71.1/53.1 | 82.2/70.5 | 77.8/61.0 | 67.0/51.3 | 73.4/57.6 | 54.3/48.0 | 70.3/53.0 | 74.6/55.6 | 62.0/56.6 | 70.2/55.9 | |
| CHARFORMERS Base | 134M | 70.3/53.7 | 78.6/61.4 | 74.4/55.1 | 85.1/73.7 | 79.8/63.6 | 69.1/52.7 | 76.7/61.3 | 57.6/51.2 | 73.9/55.8 | 76.8/57.6 | 67.4/62.4 | 73.6/59.0 | |
| CHARFORMERS Base,LongPT | 134M | 72.6/55.0 | 79.0/62.3 | 74.9/56.1 | 85.4/74.5 | 80.4/63.4 | 70.6/56.1 | 77.8/62.2 | 56.1/49.2 | 76.1/58.2 | 77.7/59.4 | 66.0/61.8 | 74.2/59.8 |
| Model | 0 | ar | de | en | es | hi | vi | zh | Avg. |
| mT5 Base (Subword) | 582M | 61.1/40.7 | 65.5/49.2 | 80.7/66.3 | 70.7/52.1 | 63.6/44.3 | 68.0/47.6 | 63.5/39.4 | 67.6/48.5 |
| Byte-levelT5 Base | 200M | 52.6/34.2 | 60.5/46.1 | 77.7/64.8 | 67.1/49.2 | 52.9/36.5 | 63.6/43.8 | 58.3/36.4 | 61.8/44.4 |
| Byte-level T5+LASCBase | 205M | 50.8/32.0 | 58.1/43.5 | 75.8/62.2 | 64.7/46.7 | 49.2/32.6 | 60.4/40.4 | 52.6/30.6 | 58.8/41.1 |
| CHARFORMER Base | 203M | 53.5/34.5 | 61.3/46.8 | 78.5/65.4 | 67.2/49.3 | 54.5/37.6 | 64.3/43.9 | 58.8/36.6 | 62.6/44.9 |
| CHARFORMER S Base | 134M | 58.3/39.1 | 65.7/50.5 | 81.8/68.7 | 71.0/53.1 | 57.7/40.8 | 67.3/46.8 | 62.7/40.8 | 66.3/48.5 |
| CHARFORMER S Base,Long PT | 134M | 59.6/40.0 | 66.6/51.3 | 82.2/69.0 | 72.1/54.5 | 59.7/42.9 | 68.2/47.4 | 62.4/40.7 | 67.2/49.4 |
| Model | 间 | ar | bg | de | el | en | es | fr | hi | ru | sw | th | tr | ur | vi | zh Avg. | |
| Translate-Train-All | |||||||||||||||||
| mT5 Base (Subword) | 582M | 74.4 | 78.5 | 77.7 | 78.1 | 82.0 | 79.1 | 77.9 72.2 | 76.5 | 71.5 | 75.0 | 74.8 | 70.4 | 74.5 | 76.0 | 75.9 | |
| Byte-level T5Base | 200M | 67.1 | 72.0 | 71.0 | 70.6 | 76.9 | 74.0 | 73.4 | 63.7 | 69.2 66.2 | 65.7 | 69.4 | 62.8 | 69.6 | 69.0 | 69.4 | |
| Byte-level T5+LASCBase | 205M | 65.6 | 72.1 | 70.5 | 67.9 | 75.6 | 73.4 | 72.2 | 63.5 | 68.6 | 65.4 | 64.5 | 67.4 | 62.4 | 68.3 | 61.0 | 67.9 |
| CHARFORMER Base | 203M | 69.5 | 72.9 | 72.7 | 72.6 | 78.2 | 74.5 | 73.6 | 67.0 | 71.7 | 67.9 | 68.1 | 70.8 | 65.0 | 70.7 | 71.5 | 71.1 |
| CHARFORMERSBase | 134M | 70.8 | 75.7 | 75.9 | 73.1 | 80.9 | 76.9 | 76.8 | 65.6 | 74.7 | 65.7 | 67.7 | 72.0 | 63.1 | 72.9 | 71.5 | 72.2 |
| CHARFORMERSBase,LongPT | 134M | 71.1 | 75.9 | 73.6 | 74.2 | 80.8 | 76.6 | 76.8 | 69.2 | 72.2 | 68.2 | 71.0 | 71.2 | 65.7 | 72.9 | 73.0 | 72.8 |
| Cross-Lingual Zero-Shot | |||||||||||||||||
| mBERTBase (Subword) | 179M | 64.3 | 68.0 | 70.0 | 65.3 | 80.8 | 73.5 | 73.4 | 58.9 | 67.8 | 49.7 | 54.1 | 60.9 | 57.2 | 69.3 | 67.8 | 65.4 |
| mT5 Base (Subword) | 582M | 73.3 | 78.6 | 77.4 | 77.1 | 84.7 | 80.3 | 79.1 | 70.8 | 77.1 | 69.4 | 73.2 | 72.8 | 68.3 | 74.2 | 74.1 | 75.4 |
| Byte-level T5Base | 200M | 56.7 | 61.2 | 63.0 | 60.9 | 79.2 | 70.1 | 65.3 | 43.9 | 61.0 | 45.5 | 43.5 | 52.0 | 44.3 | 58.3 | 55.6 | 57.4 |
| Byte-level T5+LASCBase | 205M | 53.3 | 58.8 | 62.2 | 54.9 | 77.1 | 68.6 | 65.4 | 44.7 | 58.4 | 46.1 | 43.6 | 50.4 | 42.8 | 55.9 | 46.1 | 55.2 |
| CHARFORMER Base | 203M | 55.7 | 61.1 | 64.8 | 60.1 | 77.3 | 69.9 | 67.9 | 44.4 | 60.2 | 45.3 | 47.9 | 54.0 | 43.5 | 59.1 | 53.4 | 57.6 |
| CHARFORMER S Base | 134M | 66.4 | 71.0 | 72.7 | 68.6 | 82.4 | 77.1 | 75.4 | 57.6 | 70.6 | 48.7 | 61.4 | 61.8 | 54.1 | 68.9 | 62.8 | 66.6 |
| CHARFORMERSBase,LongPT | 134M | 68.4 | 70.9 | 74.3 | 70.2 | 82.4 | 77.0 | 76.6 | 59.9 | 71.0 | 42.6 | 64.0 | 65.5 | 56.5 | 71.2 | 66.0 | 67.8 |
| Model | 0 | de | en | es | fr | ja | ko | zh | Avg. |
| Translate-Train-All | |||||||||
| mT5 Base (Subword) | 582M | 90.9 | 95.5 | 91.4 | 92.5 | 83.6 | 84.8 | 86.4 | 89.3 |
| Byte-level T5 Base | 200M | 89.3 | 94.6 | 90.1 | 90.3 | 81.4 | 81.1 | 82.3 | 87.0 |
| Byte-level T5+LASC Base | 205M | 87.3 | 93.1 | 89.2 | 89.2 | 81.0 | 72.9 | 80.8 | 84.8 |
| CHARFORMER Base | 203M | 89.9 | 94.6 | 89.8 | 91.4 | 82.7 | 78.4 | 83.3 | 87.2 |
| CHARFORMER sBase | 134M | 89.9 | 95.9 | 91.8 | 92.2 | 83.9 | 78.9 | 84.4 | 88.2 |
| CHARFORMERs Base,LongPT | 134M | 90.7 | 95.1 | 92.2 | 92.2 | 84.1 | 81.6 | 84.6 | 88.6 |
| Cross-Lingual Zero-Shot | |||||||||
| mBERTBase (Subword) | 179M | 85.7 | 94.0 | 87.4 | 87.0 | 73.0 | 69.6 | 77.0 | 81.9 |
| mT5 Base (Subword) | 582M | 89.4 | 95.4 | 89.6 | 91.2 | 79.8 | 78.5 | 81.1 | 86.4 |
| Byte-level T5 Base | 200M | 84.7 | 93.8 | 85.8 | 86.4 | 72.2 | 67.9 | 75.2 | 80.9 |
| Byte-level T5+LASCBase | 205M | 83.2 | 93.2 | 84.1 | 85.0 | 67.9 | 66.4 | 73.4 | 79.0 |
| CHARFORMER Base | 203M | 86.1 | 94.8 | 87.2 | 88.0 | 70.1 | 69.7 | 75.5 | 81.6 |
| CHARFORMER SBase | 134M | 89.6 | 95.2 | 90.7 | 90.7 | 77.1 | 74.4 | 78.9 | 85.2 |
| CHARFORMERSBase,LongPT | 134M | 89.8 | 95.3 | 88.7 | 89.7 | 74.5 | 68.9 | 78.9 | 83.7 |
| with mask | without mask | ||||||
| Scan | IDR | NeuS | VolSDF | NeuS | NeuralWarp | colmap | Ours |
| 37 | 1.87 | 0.95 | 1.26 | 1.21 | 0.71 | 0.91 | |
| 0.63 | 0.81 | 0.73 | 0.38 | 0.37 | |||
| 0.5370.336 | |||||||
| 40 | 0.80 | ||||||
| 55 | 0.48 | 0.39 | 0.49 | 0.40 | 0.38 | 0.37 | 0.357 |
| 63 | 1.04 | 1.26 | 1.25 | 1.20 | 0.79 | 0.90 | 0.800 |
| 65 | 0.79 | 0.72 | 0.70 | 0.70 | 0.81 | 1.00 | 0.454 |
| 69 | 0.77 | 0.69 | 0.72 | 0.72 | 0.82 | 0.54 | 0.408 |
| 83 | 1.33 | 0.94 | 1.29 | 1.01 | 1.20 | 1.22 | 1.032 |
| 97 | 1.16 | 1.14 | 1.18 | 1.16 | 1.06 | 1.08 | 0.843 |
| 105 | 0.76 | 0.77 | 0.70 | 0.82 | 0.68 | 0.64 | 0.548 |
| 106 | 0.67 | 0.66 | 0.66 | 0.66 | 0.66 | 0.48 | 0.460 |
| 110 | 0.90 | 1.35 | 1.08 | 1.69 | 0.74 | 0.59 | 0.4730.294 |
| 114 | 0.42 | 0.39 | 0.42 | 0.39 | 0.41 | 0.32 | |
| 118122 | 0.51 | 0.51 | 0.61 | 0.49 | 0.63 | 0.45 | 0.355 |
| 0.53 | 0.52 | 0.55 | 0.51 | 0.51 | 0.43 | 0.345 | |
| mean | 0.90 | 0.82 | 0.86 | 0.87 | 0.68 | 0.65 | 0.508 |
| Method | Lcolor | LsDF | Lphoto | mean |
| Baseline | √ | 0.87 | ||
| Model-A | 厂 | 0.62 | ||
| Model-B | 0.54 | |||
| Geo-Neus | 0.51 |
| Constraint | Setting | mean |
| Sparse 3D points | DepthintegralSDF location | 0.850.62 |
| Photometric consistency | DepthintegralSDF location | 1.080.57 |
| Accuracy | Bias Corr. (r) | |
| Base | 0.423 | |
| CDA-incr | 0.623 | 0.261 |
| 0f Avg. Diff. 8=0f-θf | -0.004 | -0.162 |
| p-value | 0.210 | 0.001 |
| Accuracy | Bias Corr. (r) | Seeds | Examples | ||
| CDA-incr | 0.623 | 0.256 | 0.264 | 0.259 | |
| CDA-full | 8 | 0.622 | 0.192 | 0.194 | 0.193 |
| Avg. Diff. | 8=0f-0f | -0.001 | -0.064 | -0.070 | -0.067 |
| p-value | 0.416 | 0.132 | 0.005 | 0.053 |
| Same | Diff. | Same - Diff. | |
| CoLA | 91.5% | 89.7% | 1.7% |
| MNLI | 93.6% | 90.1% | 3.5% |
| HANS (all) | 92.2% | 88.1% | 4.1% |
| HANS (neg) | 88.3% | 81.9% | 6.4% |
| MRPC | 91.7% | 90.4% | 1.3% |
| QNLI | 95.0% | 93.2% | 1.9% |
| QQP | 95.0% | 94.1% | 0.9% |
| RTE | 74.3% | 73.0% | 1.3% |
| SST-2 | 97.1% | 95.6% | 1.4% |
| STS-B | 97.6% | 96.2% | 1.4% |
| Bias (r) | |
| Seed 0 | 0.368 |
| 8 Seed 1 | 0.571 |
| Avg. Diff. | δ=0f-0f 0.203 |
| p-value | 0.009 |
| MNLI | RTE | MRPC | |
| 0f (1M steps) | 0.837 | 0.644 | 0.861 |
| 0f (2M steps) | 0.844 | 0.655 | 0.860 |
| 8=0f-θf | 0.007 | 0.011 | -0.001 |
| p-value | |||
| (Ho that δ≤0) | <0.001 | 0.141 | 0.564 |
| Din | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ | w./w.o Daux |
| CIFAR-10 | MSP(Hendrycks & Gimpel,2017a) ODIN(Liang et al.,2018) | 89.90 ±0.30 91.46 ± 0.56 | 91.48 ± 0.43 91.67 ± 0.58 | 60.08 ±0.76 42.31 ± 1.38 | 94.01 ±0.08 94.01 ± 0.08 | |
| Mahalanobis(Lee et al.,2018b) | 75.10 ± 1.04 | 72.32 ± 1.92 | 61.35 ± 1.25 | 94.01±0.08 | ||
| Energy(Liu et al.,2020) | 92.07 ± 0.22 | 92.72 ± 0.39 | 42.69 ± 1.31 | 94.01± 0.08 | ||
| Energy+UM(ours) | 93.73 ±0.36 | 94.27 ± 0.60 | 33.29 ± 1.70 | 92.80±0.47 | ||
| Energy+UMAP(ours) | 93.97 ± 0.11 | 94.38 ± 0.06 | 30.71 ± 1.94 | 94.01± 0.08 | ||
| OE(Hendrycks et al., 2019b) | 97.07 ± 0.01 | 97.31± 0.05 | 13.80 ± 0.28 | 92.59 ± 0.32 | ||
| Energy (w.Daux)(Liu et al.,2020) | 94.58 ± 0.64 | 94.69 ± 0.65 | 18.79 ± 2.31 | 80.91 ± 3.13 | ||
| POEM(Ming et al.,2022) OE+UM(ours) | 94.37 ± 0.07 97.60 ± 0.03 | 94.51± 0.06 97.87 ± 0.02 | 18.50 ± 0.33 11.22 ± 0.16 | 77.24 ± 2.22 93.66 ±0.12 | >>>> | |
| OE+UMAP(ours) | 97.48 ± 0.01 | 97.74 ± 0.00 | 12.21± 0.09 | 94.01 ± 0.08 | √ | |
| MSP(Hendrycks & Gimpel,2017a) ODIN(Liang et al., 2018) | 74.06 ± 0.69 | 75.37 ± 0.73 | 83.14± 0.87 | 74.86 ± 0.21 | ||
| Mahalanobis(Lee et al.,2018b) | 76.18 ± 0.14 63.90 ± 1.91 | 76.49 ± 0.20 64.31 ± 0.91 | 78.93 ± 0.31 78.79± 0.50 | 74.86 ± 0.21 | ||
| Energy(Liu et al.,2020) | 76.29 ±0.24 | 77.06± 0.55 | 78.46± 0.06 | 74.86± 0.21 74.86 ± 0.21 | ||
| Energy+UM(ours) | 76.22 ± 0.42 | 76.39 ± 1.03 | 74.05 ± 0.55 | 64.55± 0.24 | ||
| Energy+UMAP(ours) | 75.57 ± 0.59 | 75.66 ±0.07 | 72.21 ± 1.46 | 74.86 ± 0.21 | ||
| OE(Hendrycks et al., 2019b) | ||||||
| 90.55 ± 0.87 88.92 ± 0.57 | 90.34 ± 0.94 89.13 ± 0.56 | 34.73 ± 3.85 37.90 ± 2.59 | 73.59 ± 0.30 | |||
| Energy (w.Daux)(Liu et al.,2020) | 57.85 ± 2.65 | 广 | ||||
| POEM(Ming et al.,2022) | 88.95 ± 0.54 | 88.94 ± 0.31 | 38.10 ± 1.30 | 56.18 ± 1.92 | √ | |
| OE+UM(ours) | 91.04±0.11 | 91.13 ± 0.24 | 34.71 ± 0.81 | 75.15 ± 0.18 | √ | |
| OE+UMAP (ours) | 91.10 ± 0.16 | 90.99 ±0.23 | 33.62 ±0.26 | 74.76 ± 0.11 | √ |
| ID dataset | Method | OOD dataset | |||||
| CIFAR-100 | Textures | Places365 | |||||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| CIFAR-10 | MSP | 66.43 ± 1.25 | 87.73± 0.02 | 65.20±1.33 | 88.06 ±0.61 | 61.34 ± 0.60 | 89.63±0.15 |
| ODIN | 55.31± 0.85 | 87.75 ± 0.37 | 53.11 ± 4.84 | 87.13 ± 2.04 | 43.77 ±0.20 | 91.70 ± 0.30 | |
| Mahalanobis | 81.61 ± 0.96 | 64.52 ±0..73 | 20.04 ± 1.43 | 94.38 ± 0.78 | 86.21±1.36 | 64.00 ± 1.21 | |
| Energy | 54.65 ± 1.24 | 89.01 ± 1.18 | 57.09 ±3.52 | 87.51 ± 1.43 | 38.62 ± 1.64 | 93.03±0.20 | |
| Energy+UM(ours) | 54.62 ± 1.16 | 88.30±0.30 | 41.61 ± 3.67 | 91.31 ± 0.01 | 30.85 ±0.58 | 94.27 ±0.16 | |
| Method | SUN FPR95↓ | LSUN | iNaturalist | ||||
| AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | |||
| MSP | 60.27 ± 0.66 | 90.00±0.24 | 36.43 ± 1.94 | 95.17 ± 0.32 | 67.53 ± 1.64 | 88.01±0.82 | |
| ODIN | 41.14 ± 1.29 | 92.34±0.62 | 5.16 ±0.76 | 98.96 ±0.09 | 54.41 ± 0.91 | 90.17 ± 0.19 | |
| Mahalanobis | 84.56 ± 1.51 | 66.41 ± 4.57 | 69.18 ± 3.52 | 66.41 ± 4.57 | 80.76± 2.48 | 71.77 ± 1.12 | |
| Energy | 36.73 ± 1.72 | 93.63 ± 0.34 94.83 ± 0.11 | 6.25 ± 0.43 | 98.77±0.07 99.22 ± 0.11 | 59.11 ± 1.18 | 89.71 ± 0.06 92.75 ± 0.80 | |
| Energy+UM(ours) | 27.88 ±0.73 | 2.91 ± 0.53 | 46.27 ± 2.74 | ||||
| Din | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | Baseline | 91.67 | 91.89 | 40.74 | 93.67 |
| Early Stopping w. ACC | 92.13 | 92.46 | 38.86 | 93.69 | |
| Early Stopping w. OOD | 92.95 | 93.26 | 35.23 | 93.18 | |
| Weight Decay 0.1 | 86.64 | 86.67 | 60.07 | 88.53 | |
| Weight Decay 0.01 | 90.76 | 91.25 | 44.20 | 92.07 | |
| Weight Decay 0.001 | 88.93 | 88.25 | 48.95 | 94.26 | |
| Drop Rate 0.3 | 91.14 | 92.21 | 46.58 | 90.05 | |
| Drop Rate 0.4 | 84.95 | 86.62 | 62.52 | 82.55 | |
| Drop Rate 0.5 | 83.75 | 85.17 | 62.17 | 75.31 | |
| UM (ours) | 92.45 | 93.06 | 37.13 | 92.76 | |
| UMAP (ours) | 91.92 | 92.88 | 37.69 | 93.69 |
| Din | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | Baseline | 92.72 | 93.48 | 38.30 | 93.67 |
| Early Stopping w. ACC | 92.75 | 93.54 | 37.84 | 93.69 | |
| Early Stopping w. OOD | 93.04 | 93.78 | 36.56 | 93.18 | |
| Weight Decay 0.1 | 86.78 | 88.04 | 65.08 | 88.53 | |
| Weight Decay 0.01 | 90.86 | 91.77 | 47.64 | 92.07 | |
| Weight Decay 0.001 | 90.68 | 90.90 | 47.38 | 94.26 | |
| Drop Rate 0.3 | 90.52 | 91.79 | 51.23 | 90.05 | |
| Drop Rate 0.4 | 84.29 | 86.43 | 68.17 | 82.55 | |
| Drop Rate 0.5 | 83.29 | 85.14 | 68.17 | 75.31 | |
| UM (ours) | 93.58 | 94.14 | 33.66 | 92.76 | |
| UMAP (ours) | 93.17 | 93.87 | 36.11 | 93.69 |
| Din | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | Baseline | 86.24 | 85.90 | 60.13 | 93.86 |
| Early Stopping w. ACC | 83.80 | 83.30 | 65.13 | 93.99 | |
| Early Stopping w. OOD | 89.92 | 90.94 | 49.87 | 91.47 | |
| Weight Decay 0.1 | 84.38 | 84.75 | 65.75 | 89.88 | |
| Weight Decay 0.01 | 88.08 | 88.45 | 55.16 | 93.16 | |
| Weight Decay 0.001 | 86.34 | 86.38 | 57.42 | 94.91 | |
| Drop Rate 0.3 | 87.53 | 87.25 | 56.12 | 94.22 | |
| Drop Rate 0.4 | 88.24 | 88.41 | 54.62 | 94.20 | |
| Drop Rate 0.5 | 89.13 | 89.99 | 53.07 | 93.91 | |
| UM (ours) | 89.61 | 91.13 | 50.97 | 92.68 | |
| UMAP (ours) | 90.43 | 91.73 | 46.96 | 93.86 |
| Din | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | Baseline | 87.69 | 88.16 | 58.47 | 93.86 |
| Early Stopping w. ACC | 88.07 | 88.65 | 67.61 | 93.99 | |
| Early Stopping w. OOD | 90.92 | 91.94 | 46.63 | 91.47 | |
| Weight Decay 0.1 | 86.97 | 88.51 | 63.54 | 89.88 | |
| Weight Decay 0.01 | 89.77 | 89.82 | 50.23 | 93.16 | |
| Weight Decay 0.001 | 89.25 | 89.84 | 50.95 | 93.91 | |
| Drop Rate 0.3 | 89.74 | 90.07 | 52.16 | 93.22 | |
| Drop Rate 0.4 | 89.94 | 90.53 | 51.13 | 94.20 | |
| Drop Rate 0.5 | 90.09 | 91.04 | 52.76 | 93.91 | |
| UM (ours) | 91.74 | 92.67 | 40.40 | 92.68 | |
| UMAP (ours) | 88.84 | 89.31 | 50.23 | 93.86 |
| Din | Method | AUROC↑ | AUPR个 | FPR95↓ | ID-ACC↑ | w./w.o Daux |
| CIFAR-10 | MSP(Hendrycks & Gimpel, 2017a) | 89.90 ±0.30 91.46 ± 0.56 | 91.48 ± 0.43 91.67 ± 0.58 | 60.08 ±0.76 42.31 ± 1.38 | 94.01 ± 0.08 94.01 ± 0.08 | |
| ODIN(Liang et al.,2018) | ||||||
| Mahalanobis(Lee et al.,2018b) | 75.10 ± 1.04 | 72.32 ± 1.92 | 61.35 ± 1.25 | 94.01 ± 0.08 | ||
| Energy(Liu et al.,2020) | 92.07±0.22 | 92.72 ± 0.39 | 42.69 ± 1.31 | 94.01 ± 0.08 | ||
| Energy+UM(ours) | 93.73±0.36 | 94.27 ± 0.60 | 33.29 ± 1.70 | 92.80±0.47 | ||
| Energy+UMAP(ours) | 93.97 ± 0.11 | 94.38 ± 0.06 | 30.71 ± 1.94 | 94.01 ± 0.08 | ||
| OE(Hendrycks et al.,2019b) | 97.07 ± 0.01 | 97.31± 0.05 | 13.80 ± 0.28 | 92.59 ± 0.32 | ||
| Energy (w.Daux)(Liu et al.,2020) | 94.58 ± 0.64 | 94.69 ± 0.65 | 18.79 ± 2.31 | 80.91 ± 3.13 | >>>> | |
| POEM(Ming et al.,2022) OE+UM(ours) | 94.37 ± 0.07 | 94.51± 0.06 | 18.50 ± 0.33 | 77.24 ± 2.22 | ||
| Energy+UM(ours) | 97.60 ± 0.03 93.02±0.42 | 97.87 ± 0.02 92.36 ± 0.38 | 11.22 ± 0.16 24.41 ± 1.65 | 93.66 ± 0.12 71.97 ± 0.92 | ||
| POEM+UM(ours) | ||||||
| OE+UMAP(ours) | 93.04± 0.02 | 92.99 ± 0.02 | 23.52 ±0.16 | 67.41 ± 0.27 | ||
| Energy+UMAP(ours) | 95.63 ± 1.15 | 97.48 ± 0.01 | 97.74±0.00 95.92 ± 1.17 | 12.21 ± 0.09 17.51 ± 2.59 | 93.44± 0.21 | >>>> |
| POEM+UMAP (ours) | 94.18 ± 2.98 | 94.15 ± 3.46 | 20.55± 8.70 | 88.12 ± 4.22 76.62 ± 17.95 | √ | |
| MSP(Hendrycks & Gimpel,2017a) | 74.06 ±0.69 | 75.37 ± 0.73 | 83.14 ± 0.87 | 74.86 ± 0.21 | ||
| ODIN(Liang et al.,2018) | 76.18 ± 0.14 | 76.49 ± 0.20 | 78.93 ± 0.31 | 74.86 ± 0.21 | ||
| Mahalanobis(Lee et al.,2018b) | 63.90 ± 1.91 | 64.31 ± 0.91 | 78.79 ± 0.50 | 74.86 ± 0.21 | ||
| Energy(Liu et al.,2020) | 76.29 ±0.24 | 77.06± 0.55 | 78.46 ± 0.06 | 74.86 ± 0.21 | ||
| Energy+UM(ours) | 76.22 ± 0.42 | 76.39 ±1.03 | 74.05 ± 0.55 | 64.55 ± 0.24 | ||
| Energy+UMAP(ours) | 75.57 ± 0.59 | 75.66 ± 0.07 | 72.21 ± 1.46 | 74.86 ± 0.21 | ||
| OE(Hendrycks et al.,2019b) | 90.55 ± 0.87 | 90.34± 0.94 | 34.73 ±3.85 | 73.59 ± 0.30 | ||
| Energy(w.Daux)(Liu et al.,2020) | 88.92±0.57 | 89.13± 0.56 | 37.90 ± 2.59 | 57.85± 2.65 | >>>> | |
| POEM(Ming et al.,2022) | 88.95 ±0.54 | 88.94± 0.31 | 38.10 ±1.30 | 56.18 ± 1.92 | ||
| 91.04± 0.11 | 91.13 ± 0.24 | 34.71 ± 0.81 | 75.15 ± 0.18 | |||
| OE+UM(ours) | 90.14± 0.45 | 32.65 ± 3.13 | ||||
| Energy+UM(ours) | 90.39 ± 0.40 91.18 ± 0.35 | 91.45 ± 0.27 | 30.78 ± 1.76 | 71.95 ± 0.23 70.17 ± 0.01 | ||
| POEM+UM(ours) OE+UMAP (ours) | 91.10 ± 0.16 | 90.99±0.23 | 33.62±0.26 | 74.76 ± 0.11 | ||
| Energy+UMAP(ours) | 90.52 ±0.26 | 90.46 ±0.50 | 32.17 ± 0.30 | 72.76 ± 0.18 | √ | |
| POEM+UMAP (ours) | 91.10 ±0.29 | 91.41 ± 0.28 | 31.02 ± 1.70 | 71.05 ± 0.04 | √ |
| Din | Dataset Size | Structure | Atypical/Typical | AUROC↑ | AUPR↑ | FPR95↓ |
| CIFAR-10 | 200 | DenseNet-101 | Atypical Typical | 81.45 82.86 | 82.40 84.38 | 62.10 60.01 |
| WRN-40-4 | Atypical Typical | 85.13 86.26 | 86.57 86.89 | 66.41 | ||
| DenseNet-101 | Atypical Typical | 71.96 74.79 | 73.16 75.83 | 59.93 85.57 | ||
| CIFAR-100 | 1000 | WRN-40-4 | Atypical Typical | 66.64 71.95 | 67.41 72.02 | 80.97 86.92 80.00 |
| Din | Dataset Size | Atypical/Typical | ACC | AUROC↑ | AUPR↑ | FPR95↓ |
| CIFAR-10 | 200 | Atypical Typical | 3.50 100.00 | 81.45 82.86 | 82.40 83.48 | 62.10 60.01 |
| 350 | Atypical | 11.14 100.00 | 85.90 | 86.01 | 55.10 | |
| 500 | Typical Atypical | 16.80 | 85.90 84.94 | 86.16 85.33 | 52.81 59.27 | |
| Typical | 100.00 | 85.53 | 86.10 | 58.74 |
| Din | Dataset Size | Atypical/Typical | ACC | AUROC↑ | AUPR↑ | FPR95↓ |
| CIFAR-10 | 200 | Atypical Typical | 3.50 100.00 | 85.13 86.26 | 86.57 86.89 | 66.41 59.93 |
| 350 | Atypical | 11.14 | 82.92 | 84.24 | 68.57 | |
| Typical Atypical | 100.00 16.80 | 85.82 | 87.84 | 65.54 | ||
| 500 | Typical | 100.00 | 82.88 87.38 | 83.22 87.93 | 66.75 52.27 |
| Din | Dataset Size | Atypical/Typical | ACC | AUROC↑ | AUPR↑ | FPR95↓ |
| CIFAR-100 | 500 | Atypical Typical | 1.00 100.00 | 72.69 74.07 | 73.28 75.20 | 80.71 80.19 |
| 800 | Atypical | 2.88 | 69.74 | 71.15 | 85.46 | |
| Typical Atypical | 100.00 3.50 | 72.49 71.96 | 73.17 73.16 | 81.97 | ||
| 1000 | Typical | 100.00 | 74.79 | 75.83 | 85.57 80.97 |
| Din | Dataset Size | Atypical/Typical | ACC | AUROC↑ | AUPR↑ | FPR95↓ |
| CIFAR-100 | 500 | Atypical Typical | 1.00 100.00 | 66.03 68.60 | 66.17 69.93 | 89.56 86.53 |
| 800 | Atypical | 2.88 | 67.59 | 68.66 | 85.61 | |
| 1000 | Typical Atypical | 100.00 3.50 | 70.25 66.64 | 68.95 67.41 | 79.66 86.92 | |
| Typical | 100.00 | 71.95 | 72.02 | 80.00 |
| Din | Epoch | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | 100 | MSP | 89.90 | 91.48 | 60.08 | 94.01 |
| ODIN | 91.46 | 91.67 | 42.31 | 94.01 | ||
| Energy | 92.07 | 92.72 | 42.69 | 94.01 | ||
| Energy +UM | 93.73 | 94.27 | 33.29 | 92.80 | ||
| Energy + UMAP | 93.97 | 94.38 | 30.71 | 94.01 | ||
| MSP +UM | 90.31 | 91.99 | 53.61 | 91.70 | ||
| 20 | ODIN +UM | 94.08 | 94.67 | 31.01 | 91.70 | |
| Energy +UM | 93.60 | 94.32 | 33.03 | 91.70 | ||
| MSP+UMAP | 88.70 | 90.39 | 57.69 | 94.01 | ||
| ODIN+UMAP | 92.88 | 93.33 | 35.19 | |||
| Energy + UMAP | 92.88 | 93.39 | 35.60 | 94.01 | ||
| 94.01 |
| Din | Epoch | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | 100 | MSP | 87.12 | 87.84 | 68.29 | 93.86 |
| ODIN | 83.29 | 82.74 | 65.68 | 93.86 | ||
| Energy | 87.69 | 88.16 | 58.47 | 93.86 | ||
| Energy + UM | 91.74 | 92.67 | 40.40 | 92.68 | ||
| Energy + UMAP | 88.84 | 89.31 | 50.23 | 93.86 | ||
| MSP +UM | 89.86 | 91.32 | 51.62 | 91.96 | ||
| 20 | ODIN +UM | 91.97 | 92.58 | 41.78 | 91.96 | |
| Energy +UM | 92.95 | 93.64 | 36.21 | 91.96 | ||
| MSP+UMAP | 88.77 | 90.61 | 61.60 | 93.86 | ||
| ODIN + UMAP | 90.85 | 91.89 | 45.70 | 93.86 | ||
| Energy + UMAP | 91.66 | 92.49 | 42.94 | 93.86 | ||
| ID dataset | Method | OOD dataset Textures | Places365 | ||||
| CIFAR-100 FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| Energy+UMAP | 54.95 ± 2.61 | 87.72 ±1.05 | 33.59 ±1.32 | 92.67 ±0.23 | 32.80 ± 4.14 | 93.57 ± 1.12 | |
| OE | 59.29 ± 1.30 | 88.51 ±0.22 | 2.89 ± 0.30 | 99.16 ±0.05 | 11.14 ± 1.11 | 97.50 ±0.19 | |
| Energy (w. Dauz) | 79.88 ± 2.47 | 74.99 ± 2.40 | 4.27 ± 0.57 | 98.80 ±0.19 | 14.22 ± 3.99 | 97.07 ±0.72 | |
| POEM | 82.30 ± 1.57 | 72.74 ± 1.42 | 1.91 ± 0.41 | 99.40 ±0.10 | 11.24 ± 2.70 | 96.67 ±0.48 | |
| OE+UM(ours) | 55.74 ± 1.47 | 89.53 ± 0.18 | 1.42 ± 0.15 | 99.49± 0.04 | 7.77 ±0.69 | 98.15 ± 0.08 | |
| Energy (w.Daux)+ UM(ours) | 84.52 ± 0.01 | 70.09 ± 0.47 | 8.30 ±0.88 | 97.76 ±0.05 | 20.27 ±1.30 | 96.06 ±0.31 | |
| POEM+UM(ours) OE+UMAP(ours) | 84.87 ± 1.56 | 68.97 ±0.39 | 4.73 ± 0.52 | 98.88 ±0.13 | 19.83 ± 0.34 | 96.35 ±0.09 | |
| Energy(w.Daur)+UMAP(ours) | 59.05 ± 1.41 | 89.14 ± 0.14 | 1.86 ± 0.07 | 99.35 ±0.00 | 8.21±0.12 | 98.07 ±0.03 | |
| POEM+UMAP(ours) | 75.18 ± 4.96 79.33 ± 4.14 | 80.93 ± 4.49 76.89 ± 4.86 | 2.24 ± 1.34 2.10 ± 1.37 | 99.25 ±0.29 99.34 ±0.30 | 9.30 ± 2.12 9.94 ± 6.92 | 97.90 ±0.40 98.01 ± 1.11 | |
| OE | Method | SUN | LSUN | iNaturalist | |||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| Energy+UMAP | 29.05 ±2.78 | 94.41 ± 0.73 | 2.31± 0.88 | 99.42 ± 0.04 | 47.22 ± 14.03 | 92.63 ± 2.14 | |
| 8.38±0.71 | 98.00 ±0.14 | 5.90 ± 1.43 | 98.60± 0.21 | 5.09 ± 0.64 | 98.76 ± 0.11 | ||
| Energy (w. Dauz) | 10.30 ±3.82 | 97.77 ± 0.64 | 12.80 ± 4.67 | 96.08 ± 1.38 | 6.93 ± 1.86 | 98.40 ±0.32 | |
| POEM | 8.39±2.42 | 98.16 ±0.43 | 9.69 ± 1.89 | 97.25±0.58 | 3.78 ±0.90 | 98.99 ±0.17 | |
| OE+UM(ours) | 5.51 ± 0.44 | 98.55 ± 0.07 | 3.51 ± 0.43 | 98.93 ± 0.09 92.11 ± 0.94 | 2.87 ± 0.49 | 99.14 ± 0.09 | |
| Energy(w.Daux)+ UM(ours) POEM+UM(ours) | 16.13 ± 1.86 16.16 ± 0.57 | 96.84±0.30 97.01±0.08 | 23.27 ± 2.40 25.69 ± 0.15 | 93.38 ± 0.27 | 11.20 ± 2.35 9.30 ± 1.60 | 97.54±0.42 98.05 ±0.24 | |
| OE+UMAP(ours) | 6.16 ±0.02 | 98.49 ± 0.01 | 4.53 ± 0.16 | 98.86±0.06 | 3.40 ± 0.74 | 98.96 ±0.09 | |
| Energy(w.Dauz)+UMAP(ours) | 6.67 ± 1.50 | 98.40±0.32 | 23.50 ± 5.61 | 94.78 ± 2.04 | 3.77 ± 2.14 | 98.93 ±0.46 | |
| 94.74 ±3.67 | |||||||
| POEM+UMAP(ours) | 7.00±5.80 | 98.46±0.96 | 21.17 ± 12.84 | 3.63±2.78 | 98.99 ±0.54 | ||
| ID dataset | Method | OOD dataset | Places365 | |||||
| CIFAR-10 FPR95↓ | AUROC↑ | Textures FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | |||
| CIFAR-100 | MSP ODIN Mahalanobis Energy | 83.53± 0.33 85.29 ± 0.17 | 75.11 ± 0.27 73.31 ± 0.24 49.60 ± 1.51 | 86.90 ±0.18 86.45 ± 1.27 33.06± 3.76 | 71.45 ± 0.40 71.91 ± 0.27 | 85.83 ±0.48 84.35 ± 0.64 | 70.54 ± 0.42 73.58 ± 0.51 | |
| 98.25 ±0.05 | 90.19 ± 1.21 | 95.20 ±0.49 | 53.69 ± 1.55 | |||||
| Energy+UM(ours) | 82.16 ± 0.59 | 75.31 ± 0.21 | 90.20 ±0.30 | 68.98 ±0.34 65.39 ± 1.44 | 82.39 ±0.97 77.30 ± 2.08 | 73.78 ±0.66 76.06 ± 1.05 | ||
| OE | 89.62 ±0.07 | 66.12 ± 0.93 | 86.99 ± 1.22 | |||||
| Energy (w.Daux) POEM | 90.97 ±0.46 96.14 ±0.06 | 69.02 ± 0.39 57.52 ± 0.88 | 14.36 ± 0.25 9.02 ±0.06 | 95.92 ±0.12 97.16 ±0.26 | 40.19 ± 6.97 35.18 ± 4.73 | 90.70 ± 2.01 93.29 ± 1.14 | ||
| OE+UM (ours) | 96.19 ±0.16 | 55.82 ± 1.05 | 7.63 ± 1.40 | 97.69 ±0.06 | 32.67 ± 3.73 | 93.94±0.68 | ||
| 89.61±0.08 | 71.24 ± 0.08 | 16.78 ± 0.25 | 95.60 ±0.08 | 39.77 ± 0.34 | 91.07 ± 0.13 | |||
| Energy(w.Daux)+UM(ours) | 95.38 ± 0.45 | 63.41 ± 0.14 | 6.41 ± 0.83 | 97.77 ± 0.33 | 30.96 ± 3.61 | 92.85 ±0.69 | ||
| POEM+UM(ours) | 95.78 ± 0.14 | 60.23 ± 0.70 | 5.17 ±0.18 | 98.53± 0.03 | 23.90 ±0.84 | 95.45 ± 0.11 | ||
| OE+UMAP(ours) Energy(w.Daux)+UMAP(ours) | 90.72±0.35 | 69.76 ± 0.25 | 15.32 ± 0.23 | 95.72 ± 0.01 | 36.42 ± 1.91 | 92.08 ±0.49 | ||
| POEM+UMAP(ours) | 95.39±0.10 95.69 ± 0.17 | 63.26 ±0.18 61.62 ± 0.24 | 6.52 ±0.44 5.23 ± 0.58 | 97.83 ±0.18 98.52 ± 0.01 | 31.18 ± 0.43 26.06 ± 1.16 | 93.13±0.41 94.91 ± 0.25 | ||
| Method | SUN | LSUN | ||||||
| FPR95↓ AUROC↑ | FPR95↓ | AUROC↑ | iNaturalist FPR95↓ | AUROC↑ | ||||
| MSP | 88.75 ±0.23 | 66.75 ± 0.25 | 67.83 ±1.37 | 82.94± 0.32 | 85.00 ± 0.73 | 76.62 ± 0.25 | ||
| ODIN | 88.49 ± 0.99 | 69.64± 0.61 54.37 ± 1.35 | 34.80 ± 2.55 89.31 ± 4.83 | 93.92 ±0.75 43.19 ± 16.36 | 81.67 ± 2.77 93.63 ± 1.19 | 78.36 ± 1.57 | ||
| Mahalanobis | 95.53±0.37 | 49.60 ± 1.51 | ||||||
| Energy Energy+UM(ours) | 97.17 ± 0.92 69.04 ± 0.83 | 35.09 ± 3.17 | 93.49 ±0.87 | 85.70 ± 2.14 | 75.82 ± 1.72 | |||
| 81.96 ± 2.26 | 71.47 ± 1.88 | 22.54± 5.93 | 94.98 ± 1.75 | 74.28 ± 3.72 | 80.72 ± 3.75 | |||
| OE | 44.47 ± 9.10 90.70 ± 2.01 | 5.75 ± 1.18 | 98.57 ± 0.13 | 25.51 ± 4.12 | 94.46 ± 0.88 | |||
| Energy (w. Daux) | 32.69 ± 5.69 | 93.63 ± 1.48 | 55.75 ± 4.31 | 87.96 ± 1.03 | 17.34 ± 4.54 | 96.50 ±0.81 | ||
| POEM | 30.45 ± 5.11 | 94.26 ± 0.90 | 46.68 ± 3.59 | 90.30 ± 2.17 | 16.50 ± 2.09 | 96.63±0.23 | ||
| OE+UM(ours) | 44.23±0.20 90.28±0.03 | 5.80 ±0.33 | 98.63 ±0.03 | 26.72 ± 1.95 | 94.51 ± 0.43 | |||
| 93.18 ± 0.69 | 91.98 ± 0.83 | 10.83 ± 2.06 | 97.09 ±0.69 | |||||
| Energy(w.Daux)+UM(ours) POEM+UM(ours) | 28.98 ±3.15 | 37.56 ± 4.81 | 94.43 ± 1.18 | |||||
| 21.34 ± 1.07 | 95.76 ±0.22 | 33.74 ±6.22 | 98.64 ± 0.12 | 8.85±0.23 | 97.93 ± 0.03 95.08 ±0.22 | |||
| OE+UMAP(ours) Energy(w.Daux)+UMAP(ours) | 39.58 ± 2.02 | 91.37 ± 0.52 93.38 ± 0.19 | 5.77 ± 0.71 | 92.08 ±0.39 | 23.33 ± 1.24 13.96 ± 2.48 | 96.87 ± 0.15 | ||
| 29.65 ± 1.06 | 35.94 ± 0.75 | 93.57 ±0.88 | ||||||
| POEM+UMAP(ours) | 23.73 ± 0.71 | 95.25 ± 0.06 | 33.09 ± 5.94 | 9.76 ± 1.09 | 97.77 ± 0.24 | |||
| Din | Method | AUROC↑ | AUPR↑ | FPR95↓ | ID-ACC↑ |
| CIFAR-10 | MSP(Hendrycks & Gimpel,2017a) | 87.12 ± 0.25 | 87.84± 0.30 | 68.29 ±0.96 | 93.86 ± 0.19 |
| ODIN(Liang et al.,2018) | 83.29 ±0.72 | 82.74±0.79 | 65.68 ± 0.77 | 93.86 ± 0.19 | |
| Mahalanobis(Lee et al., 2018b) | 77.57 ± 0.28 | 76.11 ±0.10 | 61.18 ±0.10 | 93.86 ± 0.19 | |
| Energy(Liu et al.,2020) | 87.69 ± 0.54 | 88.16 ±0.69 | 58.47 ± 1.94 | 93.86 ± 0.19 | |
| Energy+UM(ours) | 91.74 ± 0.43 | 92.67 ± 0.52 | 40.40±1.32 | 92.68±0.23 | |
| Energy+UMAP(ours) | 88.84± 1.02 | 89.31 ± 1.44 | 50.23 ±2.25 | 93.86 ± 0.19 | |
| CIFAR-100 | MSP(Hendrycks & Gimpel,2017a) | 72.34± 0.63 | 72.69 ± 0.44 | 85.40 ±0.59 | 75.01 ±0.07 |
| ODIN(Liang et al., 2018) | 68.78±0.67 | 66.92 ± 0.72 | 85.28 ± 0.64 | 75.01 ± 0.07 | |
| Mahalanobis(Lee et al., 2018b) | 68.20±0.99 | 68.30 ± 1.15 | 76.46 ± 2.02 | 75.01 ± 0.07 | |
| Energy(Liu et al.,2020) | 74.00 ± 0.41 | 73.02 ± 0.47 | 81.37 ±0.08 | 75.01 ± 0.07 | |
| Energy+UM(ours) | 76.07 ± 0.04 | 76.94 ± 0.06 | 74.29 ± 1.66 | 59.08 ± 2.75 | |
| Energy+UMAP(ours) | 77.35 ± 0.78 | 77.43 ± 0.91 | 68.20 ± 0.06 | 75.01 ±0.07 |
| ID dataset | Method | OOD dataset | |||||
| CIFAR-100 | Textures | Places365 | |||||
| MSP | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | |
| CIFAR-10 | ODIN | 70.96 ±0.70 | 86.08 ±0.08 | 68.81 ± 1.29 | 86.53 ±0.83 | 68.31 ± 0.25 | 86.71 ±0.13 |
| 64.97 ±0.08 | 83.36 ±0.11 | 66.86 ±2.24 | 81.34 ± 0.81 | 66.49 ±1.16 | 83.47 ±0.93 | ||
| Mahalanobis | 79.84±0.55 | 70.33 ± 0.24 | 22.56 ± 0.08 | 94.07 ± 0.04 | 85.09 ± 0.59 | 67.90 ± 0.37 | |
| Energy | 61.09±0.58 | 86.66 ±0.04 | 64.29 ± 1.72 | 85.56 ±0.53 | 55.32 ± 0.13 | 88.29 ±0.26 | |
| Energy+UM(ours) | 57.21 ± 1.41 | 87.56± 0.15 | 46.49 ± 1.03 | 89.74±0.45 | 40.68±4.46 | 92.51± 0.97 90.43± 0.47 | |
| Energy+UMAP(ours) | 65.45 ± 1.10 | 84.65 ±0.95 | 59.14 ± 1.64 | 85.27 ± 1.74 | 48.16 ± 1.89 | ||
| Method | SUN | LSUN | iNaturalist | ||||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| MSP | 68.62 ±0.50 | 86.95 ±0.23 | 52.97± 3.07 | 92.41 ±0.18 | 76.05 ± 0.01 | 83.44±0.36 | |
| ODIN | 65.47 ±0.78 | 83.79 ± 1.10 | 31.89 ± 3.44 | 94.34 ±0.89 | 79.28 ±0.18 | 79.80 ±0.35 | |
| Mahalanobis | 82.92 ±0.28 | 70.52 ± 0.47 | 64.31 ± 0.57 | 67.75 ±0.55 | 81.50 ± 2.91 | 74.97 ± 2.91 | |
| Energy | 54.88 ±0.18 | 88.67±0.30 | 24.99 ± 1.38 | 95.98± 0.37 | 75.89 ± 0.85 | 82.40±0.22 | |
| Energy+UM(ours) | 38.92 ± 3.46 | 92.98 ± 0.95 | 8.38 ±0.77 | 98.18 ± 0.16 | 66.02 ± 6.70 | 85.22 ±3.42 | |
| Energy+UMAP(ours) | 45.94 ± 2.64 | 91.27±0.56 | 14.10 ± 0.04 | 97.46 ± 0.14 | 74.69±0.15 | 81.13 ± 1.12 | |
| ID dataset | Method | OOD dataset | |||||
| CIFAR-10 | Textures | Places365 | |||||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| CIFAR-100 | MSP | 83.83 ±0.29 | 75.50±0.21 | 86.15 ± 0.23 | 72.36 ± 0.40 | 86.72 ±0.29 | 69.60± 0.27 |
| ODIN | 83.70 ±0.30 | 74.32 ±0.03 | 81.57 ± 1.74 | 71.67±0.28 | 88.07 ± 0.11 | 64.83 ± 1.36 | |
| Mahalanobis | 96.89 ± 0.11 | 68.78 ±0.67 | 31.02 ± 2.04 | 91.85 ± 0.91 | 93.34 ± 0.44 | 61.28 ± 1.62 | |
| Energy | 81.32 ± 0.47 | 77.49 ± 0.26 | 86.38 ±0.49 | 73.50±0.45 | 84.45 ± 0.38 | 69.82 ± 0.63 | |
| Energy+UM(ours) | 89.23 ± 1.51 | 63.85 ± 1.73 | 78.90 ±0.07 | 72.58 ± 1.33 | 80.46 ± 1.99 | 70.49 ± 1.01 | |
| Energy+UMAP(ours) | 94.11 ± 0.72 | 60.77 ± 0.96 | 66.94 ± 4.49 | 75.82 ± 5.33 | 82.59±0.27 | 71.92 ± 3.58 | |
| Method | SUN | LSUN | iNaturalist | ||||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| MSP | 88.88±0.83 | 65.22 ± 0.85 | 78.56±0.66 | 79.10 ± 0.53 | 86.72 ± 0.29 | 73,75 ±0.56 | |
| ODIN | 91.00 ±0.10 | 59.06 ± 1.81 | 70.14 ± 2.42 | 84.03±0.86 | 87.86±1.13 | 64.52 ± 1.51 | |
| Mahalanobis | 94.22 ± 0.01 | 60.09 ± 1.56 | 89.73± 2.87 | 40.81± 3.07 | 87.25±3.28 | 74.98 ± 2.85 | |
| Energy | 88.35±0.52 | 64.04 ±0.76 | 59.84 ±0.06 | 87.91 ± 0.53 | 88.91±0.78 | 67.81 ± 0.91 | |
| Energy+UM(ours) | 84.04 ±0.09 | 67.19 ± 0.14 | 33.87 ± 1.21 | 92.29 ± 0.02 | 76.91 ± 6.07 | 79.28 ± 4.17 | |
| Energy+UMAP(ours) | 80.53 ± 1.31 | 72.68± 4.50 | 27.79 ± 2.19 | 93.39 ± 0.57 | 55.53± 4.61 | 85.65± 0.83 | |
| Stochastic DFO Settings | Rate |
| Decision-indep. | 0(1/²2) |
| (Ghadimi & Lan, 2013) | |
| Decision-depend. (Markov) O(1/ε) |
| Model | Φtemp | Pcomp | T | Lcons | Lentro | DAVIS(J ↑) |
| Ours-A | √ | √ | 7 | √ | √ | 73.9 |
| Ours-B | 1 | X | 7 | √ | √ | fail |
| Ours-C | X | √ | 7 | √ | √ | 68.3 |
| Ours-D | √ | √ | 3 | √ | √ | 66.4 |
| Ours-E | √ | √ | 5 | √ | √ | 68.2 |
| Ours-F | √ | √ | 7 | X | X | 60.4 |
| Ours-G | √ | √ | 7 | X | √ | 65.6 |
| Ours-H | √ | √ | 7 | √ | X | 69.5 |
| Model | Training | Inference | J (Mean) ↑ | ||||
| Sup. | RGB | Flow | p.p. | DAVIS2016 | SegTrackv2 | FBMS-59 | |
| SAGE (Wang et al., 2017) | None | √ | √ | X | 42.6 | 57.6 | 61.2 |
| NLC (Faktor & Irani, 2014) | None | √ | √ | 55.1 | 67.2 | 51.5 | |
| CIS (Yang et al., 2019) | None | √ | 71.5 | 62.5 | 63.5 | ||
| AMD (Liu et al., 2021) | None | √ | X | X | 57.8 | 57.0 | 47.5 |
| SIMO (Lamdouar et al., 2021) | Syn. | X | √ | 67.8 | 62.0 | = | |
| MG (Yang et al., 2021a) | None | X | X | 68.3 | 58.6 | 53.1 | |
| OCLR (Xie et al., 2022) | Syn. | X | √ | X | 72.1 | 67.6 | 65.4 |
| MOD (w/o post-processing) | None | √ | X | X | 73.9 | 62.2 | 61.3 |
| MOD (test-time adaptation) | None | √ | X | √ | 79.2 | 69.4 | 66.9 |
| FSEG (Dutt Jain et al., 2017) | GT | √ | √ | = | 70.7 | 61.4 | 68.4 |
| COSNet (Lu et al., 2019) | GT | √ | X | 80.5 | 49.7 | 75.6 | |
| MATNet (Zhou et al.,2020) | GT | √ | √ | 82.4 | 50.4 | 76.1 | |
| D²Conv3d (Schmidt et al., 2022) | GT | √ | X | = | 85.5 | - | - |
| stage | operation | output sizes |
| input | = | 3×192×384 |
| PatchEmbed | 4 × 4, stride 4, 96 | 96 × 48 × 96 |
| SwinBlock1 | h=3 ×2 [ws= 12] | 96 × 48 × 96 |
| DownSample1 | 192 | 192 × 24× 48 |
| SwinBlock2 | h=6] ×2 [ws=12 | 192 × 24× 48 |
| DownSample2 | 384 | 384 × 12 × 24 |
| SwinBlock3 | [h=12] ×6 [ws=12 | 384 × 12 × 24 |
| stage | operation | output sizes |
| input | 1 | 768 × 12 × 24 |
| DeformConv1 | 3 ×3,768 | 768 × 12 × 24 |
| ReLU | - | 768 × 12 × 24 |
| DeformConv2 | 3 ×3,384 | 384 ×12×24 |
| Transformer | [h=8] ×3 384 | 384× 12 ×24 |
| stage | operation | output sizes |
| input | = [h=12] | 384×12×24 |
| SwinBlock1 | ×2 [ws = 12] | 384 × 12 × 24 |
| PatchExpand1 | 768 | 192 × 24 × 48 |
| SwinBlock2 | h=6] ×2 [ws=12 | 192 × 24 × 48 |
| PatchExpand2 | 384 | 96 × 48 × 96 |
| SwinBlock3 | h=3] ×2 [ws=12 | 96 × 48 × 96 |
| PatchExpand3 outConv | 1536 5 × 5, stride 1, 4 | 96 ×192 × 384 4×192×384 |
| Sequence | J (Mean) ↑ | |
| w/o post proc. | test-time adap. | |
| dog | 80.7 | 87.4 |
| cows | 87.2 | 88.8 |
| goat | 47.5 | 80.6 |
| camel | 85.6 | 86.1 |
| libby | 72.5 | 77.7 |
| parkour | 72.9 | 87.9 |
| soapbox | 84.6 | 86.5 |
| blackswan | 48.9 | 46.9 |
| bmx-trees | 50.1 | 55.8 |
| kite-surf | 55.9 | 62.6 |
| car-shadow | 87.9 | 86.9 |
| breakdance | 82.6 | 76.0 |
| dance-twirl | 82.5 | 85.4 |
| scooter-black | 80.2 | 80.3 |
| drift-chicane | 78.6 | 82.2 |
| motocross-jump | 68.4 | 88.9 |
| horsejump-high | 78.0 | 84.3 |
| drift-straight | 69.2 | 80.0 |
| car-roundabout | 87.7 | 83.9 |
| paragliding-launch | 62.1 | 62.8 |
| frame avg. | 73.9 | 79.2 |
| Sequence | J (Mean) ↑ | |
| w/o post proc. | test-time adap. | |
| drift | 41.7 | 40.7 |
| birdfall | 38.2 | 61.5 |
| girl | 76.5 | 82.3 |
| cheetah | 18.4 | 30.1 |
| worm | 52.5 | 74.3 |
| parachute | 90.2 | 92.0 |
| monkeydog | 14.3 | 31.1 |
| hummingbird | 61.2 | 58.8 |
| soldier | 66.3 | 58.8 |
| bmx | 73.7 | 78.8 |
| frog | 80.5 | 76.3 |
| penguin | 63.5 | 62.7 |
| monkey | 46.8 | 77.4 |
| bird of paradise | 85.3 | 85.4 |
| frame avg. | 62.2 | 69.4 |
| J (Mean) ↑ | ||
| w/o post proc. | test-time adap. | |
| Sequence camel01 | 25.8 | 66.9 |
| cars1 | 66.1 | 88.3 |
| cars10 | 31.6 | 33.9 |
| cars4 | 72.9 | 83.2 |
| cars5 | 81.2 | 82.5 |
| cats01 | 80.6 | 79.2 |
| cats03 | 62.0 | 63.4 |
| cats06 | 40.1 | 38.7 |
| dogs01 | 70.6 | 61.2 |
| dogs02 | 62.8 | 82.2 |
| farm01 | 86.7 | 88.9 |
| giraffesO1 | 38.6 | 52.2 |
| goats01 | 44.8 | 45.3 |
| horses02 | 64.4 | 77.6 |
| horses04 | 68.5 | 73.5 |
| horses05 | 43.7 | 49.0 |
| lion01 | 60.1 | 71.5 |
| marple12 | 74.7 | 80.3 |
| marple2 | 65.0 | 71.4 |
| marple4 | 79.1 | 91.6 |
| marple6 | 76.0 | 85.1 |
| marple7 | 72.5 | 55.2 |
| marple9 | 87.5 | 97.9 |
| people03 | 76.5 | 48.5 |
| people1 | 72.4 | 80.8 |
| people2 | 80.9 | 83.0 |
| rabbits02 | 50.2 | 58.9 |
| rabbits03 | 39.5 | 55.7 |
| rabbits04 | 47.0 | 53.0 |
| tennis | 63.1 | 71.1 |
| frame avg. | 61.3 | 66.9 |
| Flow Estimator | DAVIS2016 | SegTrackv2 | FBMS-59 |
| RAFT | 73.9 | 62.2 | 61.3 |
| ARFlow | 59.2 | 51.1 | 50.0 |
| Category | #FPs | #FNs | Category | #FPs | #FNs | ||||
| A | H | A | H | H | A | H | |||
| irish setter | 12 | 3 | 0 | 0 | maypole | 0 | 0 | 7 | 5 |
| gorilla | 32 | 7 | 0 | 0 | basketball | 0 | 0 | 7 | 2 |
| cheetah | 3 | 1 | 0 | 0 | volleyball | 0 | 0 | 10 | 5 |
| basset | 10 | 0 | 0 | 0 | balance beam | 0 | 0 | 9 | 5 |
| lynx | 9 | 1 | 0 | 0 | unicycle | 0 | 1 | 6 | 1 |
| rottweiler | 11 | 4 | 0 | 0 | stage | 0 | 0 | 0 | 0 |
| sorrel | 2 | 1 | 0 | 0 | torch | 2 | 1 | 1 | 1 |
| impala | 1 | 0 | 0 | 0 | baseball player | 0 | 0 | 0 | 0 |
| bernese mt. dog | 20 | 3 | 0 | 0 | military uniform | 3 | 2 | 2 | 0 |
| silky terrier | 4 | 0 | 0 | 0 | steel drum | 1 | 1 | 1 | 0 |
| Average | 10.4 | 2.0 | 0.0 | 0.0 | Average | 0.6 | 0.5 | 4.3 | 1.9 |
| Supercategory | #Categories | #Images | With faces (%) |
| clothing | 49 | 62,471 | 58.90 |
| wheeled 1vehicle | 44 | 57,055 | 35.30 |
| musical instrument | 26 | 33,779 | 47.64 |
| bird | 59 | 76,536 | 1.69 |
| insect | 27 | 35,097 | 1.81 |
| Model | Top-1 accuracy (%) | Top-5 accuracy (%) | ||||||||
| Original | Blurred | △b | overlaid | △。 Original | Blurred | △b | overlaid | △ | ||
| AlexNet | 56.0± 0.3 | 55.8±0.1 | 0.2 | 55.5± 0.2 | 0.6 | 78.8 ± 0.1 | 78.6 ± 0.1 | 0.3 | 78.2 ± 0.2 | 0.7 |
| SqueezeNet | 56.0 ± 0.2 | 55.3± 0.0 | 0.7 | 55.0 ± 0.2 | 1.0 | 78.6± 0.2 | 78.1 ± 0.0 | 0.5 | 77.6 ± 0.1 | 1.0 |
| ShuffleNet | 64.7 ± 0.2 | 64.0 ± 0.1 | 0.6 | 63.7 ± 0.0 | 1.0 | 85.9 ± 0.0 | 85.5 ± 0.1 | 0.5 | 85.2 ± 0.2 | 0.8 |
| VGG11 | 68.9 ± 0.0 | 68.2 ± 0.1 | 0.7 | 67.8± 0.2 | 1.1 | 88.7 ± 0.0 | 88.3 ± 0.1 | 0.4 | 87.9 ± 0.0 | 0.8 |
| VGG13 | 69.9 ± 0.1 | 69.3 ± 0.1 | 0.7 | 68.8 ± 0.0 | 1.2 | 89.3 ± 0.1 | 88.9± 0.0 | 0.4 | 88.5 ± 0.1 | 0.8 |
| VGG16 | 71.7 ± 0.1 | 70.8 ± 0.1 | 0.8 | 70.6 ± 0.1 | 1.1 | 90.5 ± 0.1 | 89.9 ± 0.1 | 0.6 | 89.6± 0.0 | 0.9 |
| VGG19 | 72.4 ± 0.0 | 71.5 ± 0.0 | 0.8 | 71.2 ± 0.2 | 1.2 | 90.9 ± 0.1 | 90.3±0.0 | 0.6 | 90.1 ± 0.1 | 0.8 |
| MobileNet | 65.4 ± 0.2 | 64.4 ± 0.2 | 1.0 | 64.3± 0.2 | 1.0 | 86.7 ± 0.1 | 86.0 ± 0.1 | 0.7 | 85.7 ± 0.1 | 0.9 |
| DenseNet121 | 75.0 ± 0.1 | 74.2 ± 0.1 | 0.8 | 74.1 ± 0.1 | 1.0 | 92.4 ± 0.0 | 92.0 ± 0.1 | 0.4 | 91.7 ± 0.0 | 0.7 |
| DenseNet201 | 77.0± 0.0 | 76.6 ± 0.0 | 0.4 | 76.1 ± 0.1 | 0.9 | 93.5 ± 0.0 | 93.2 2±0.1 | 0.2 | 92.9 ± 0.1 | 0.6 |
| ResNet18 | 69.8 ± 0.2 | 69.0± 0.2 | 0.7 | 68.9 ± 0.1 | 0.8 | 89.2 ± 0.0 | 88.7± 0.0 | 0.5 | 88.7±0.1 | 0.6 |
| ResNet34 | 73.1 ± 0.1 | 72.3 ± 0.4 | 0.8 | 72.4 ± 0.1 | 0.7 | 91.3 ± 0.0 | 90.8 ± 0.1 | 0.5 | 90.7 ± 0.0 | 0.6 |
| ResNet50 | 75.5± 0.2 | 75.0 ± 0.1 | 0.4 | 74.9 ± 0.0 | 0.6 | 92.5 ± 0.0 | 92.4± 0.1 | 0.1 | 92.2 ± 0.0 | 0.3 |
| ResNet101 | 77.3 ± 0.1 | 76.7 ± 0.1 | 0.5 | 76.7 ± 0.1 | 0.6 | 93.6 ± 0.1 | 93.3 ± 0.1 | 0.3 | 93.1 ± 0.1 | 0.5 |
| ResNet152 | 77.9 ± 0.1 | 77.3 ± 0.1 | 0.6 | 77.0± 0.3 | 0.9 | 93.9 ± 0.0 | 93.7 ± 0.0 | 0.4 | 93.3 ±0.3 | 0.6 |
| Average | 70.0 | 69.4 | 0.7 | 69.1 | 0.9 | 89.1 | 88.6 | 0.4 | 88.4 | 0.7 |
| Category | Top-1 accuracy (%) | Top-5 accuracy (%) | Average precision (%) | ||||||
| Original | Blurred | A | Original | Blurred | △ | Original | Blurred | A | |
| eskimo dog | 50.8 ± 1.1 | 38.0±0.4 | 12.8 | 95.5±0.4 | 95.1 ±0.2 | 0.4 | 19.4±0.8 | 19.9 ± 0.5 | -0.5 |
| siberian husky | 46.3 ± 1.8 | 63.2 ±0.8 | -16.9 | 97.0 ±0.4 | 97.2 ± 0.3 | -0.2 | 29.2 ± 0.3 | 29.6± 0.5 | -0.4 |
| projectile | 35.6 ± 0.9 | 21.7 ± 1.0 | 13.9 | 86.2± 0.4 | 85.5±0.4 | 0.7 | 23.1 ±0.4 | 22.5±0.5 | 0.6 |
| missile | 31.6±0.7 | 45.8 ± 0.8 | -14.2 | 81.5±0.7 | 81.8±0.4 | -0.3 | 20.4 ± 0.3 | 21.1 ± 0.6 | -0.7 |
| tub | 35.5 ± 1.5 | 27.9 ±0.6 | 7.6 | 79.4 ± 0.6 | 75.6±0.5 | 3.8 | 19.9 ± 0.4 | 18.8±0.2 | 1.1 |
| bathtub | 35.4± 1.0 | 42.5± 0.4 | - 7.1 | 78.9 ± 0.3 | 80.8 ± 1.2 | -1.9 | 27.4±0.8 | 25.1±0.6 | 2.3 |
| american chameleon | 63.0±0.4 | 54.7 ± 1.2 | 8.3 | 97.0±0.5 | 96.6±0.5 | 0.4 | 40.0±0.2 | 39.3±0.5 | 0.7 |
| green lizard | 42.0 ± 0.6 | 45.6 ± 1.2 | -3.6 | 91.3 ± 0.3 | 89.7±0.2 | 1.6 | 22.6±0.8 | 22.4 ± 0.1 | 0.2 |
| Model | CIFAR-10 | SUN | ||||||
| No pretrain | Original | Blurred | Overlaid | No pretrain | Original | Blurred | Overlaid | |
| AlexNet | 83.3±0.2 | 90.6±0.0 | 90.9±0.0 | 91.1 ± 0.0 | 26.2±0.6 | 46.3±0.1 | 46.5± 0.1 | 46.2±0.0 |
| ShuffleNet | 92.3 ±0.3 | 95.7 ± 0.0 | 95.4 ± 0.1 | 95.2 ± 0.1 | 33.8±0.7 | 51.2 ± 0.1 | 50.4± 0.3 | 49.3 ± 0.3 |
| ResNet18 | 92.8± 0.1 | 96.1 ± 0.1 | 96.1 ± 0.1 | 96.1 ± 0.1 | 36.9 ± 4.8 | 55.0±0.2 | 55.0± 0.1 | 55.1 ± 0.1 |
| ResNet34 | 90.6 ± 0.9 | 96.9 ± 0.1 | 97.0± 0.0 | 97.1 ± 0.2 | 40.3 ± 0.4 | 57.8±0.0 | 57.9 ± 0.1 | 57.8±0.1 |
| Model | No pretrain | Original | Blurred | Overlaid |
| AlexNet | 41.8 ± 0.5 | 55.5± 0.7 | 50.7 ± 0.8 | 52.5 ± 0.4 |
| ShuffleNet | 36.5 ± 0.7 | 55.6 ±1.2 | 52.5 ± 1.0 | 53.5 ± 1.4 |
| ResNet18 | 45.1 ± 1.0 | 51.7 ± 1.9 | 51.8 ± 1.0 | 52.0 ±0.6 |
| ResNet34 | 49.4 ± 2.4 | 55.6 ± 2.4 | 56.5 ± 1.9 | 56.4 ± 2.3 |
| ResNet50 | 48.7 ± 1.3 | 42.8 ± 0.9 | 50.9 ± 2.7 | 50.4 ± 0.5 |
| VGG11 | 48.7± 0.3 | 56.0 ± 0.7 | 57.4 ± 0.6 | 58.1 ± 0.9 |
| VGG13 | 47.2 ± 0.8 | 58.4 ± 0.6 | 59.0 ± 0.5 | 58.2 ± 0.4 |
| MobileNet | 43.8 ± 0.2 | 49.4 ± 0.8 | 49.9 ± 1.3 | 49.6 ± 1.3 |
| Model | Original | Red | Green | Blue | Mean | Blurred | △b |
| AlexNet (Krizhevsky et al., 2017) | 79.1 | 76.7 | 77.1 | 76.7 | 77.8 | 78.2 | 0.8 |
| GoogLeNet (Szegedy et al., 2015) | 89.5 | 87.9 | 88.2 | 87.9 | 88.3 | 88.7 | 0.9 |
| Inception v3 (Szegedy et al.,2016) | 88.7 | 86.7 | 87.0 | 86.6 | 87.2 | 87.7 | 0.9 |
| SqueezeNet (Iandola et al., 2016) | 80.6 | 78.6 | 79.0 | 78.5 | 79.4 | 79.7 | 0.9 |
| ShuffleNet (Zhang et al.,2018) | 88.3 | 86.6 | 86.8 | 86.6 | 87.0 | 87.4 | 1.0 |
| VGG11 (Simonyan & Zisserman, 2015) | 88.6 | 87.1 | 87.4 | 87.0 | 87.6 | 87.8 | 0.8 |
| VGG13 | 89.3 | 87.9 | 88.1 | 87.9 | 88.2 | 88.5 | 0.8 |
| VGG16 | 90.4 | 89.1 | 89.1 | 88.9 | 89.3 | 89.7 | 0.7 |
| VGG19 | 90.9 | 89.4 | 89.5 | 89.2 | 89.7 | 90.1 | 0.8 |
| MobileNet (Howard et al.,2017) | 90.3 | 88.9 | 89.1 | 88.9 | 89.2 | 89.5 | 0.8 |
| MNASNet (Tan et al., 2019) | 91.5 | 90.0 | 90.2 | 90.2 | 90.4 | 90.8 | 0.7 |
| DenseNet121 (Huang et al., 2017) | 92.0 | 90.7 | 90.8 | 90.7 | 91.0 | 91.3 | 0.7 |
| DenseNet161 | 93.6 | 92.5 | 92.5 | 92.3 | 92.8 | 93.0 | 0.6 |
| DenseNet169 | 92.8 | 91.6 | 91.7 | 91.6 | 91.9 | 92.2 | 0.6 |
| DenseNet201 | 93.4 | 92.2 | 92.3 | 92.0 | 92.3 | 92.7 | 0.7 |
| ResNet18 (He et al., 2016) | 89.1 | 87.5 | 87.6 | 87.5 | 87.8 | 88.3 | 0.8 |
| ResNet34 | 91.4 | 89.8 | 90.0 | 89.8 | 90.2 | 90.7 | 0.8 |
| ResNet50 | 92.9 | 91.7 | 91.8 | 91.5 | 91.8 | 92.2 | 0.7 |
| ResNet101 | 93.6 | 92.3 | 92.4 | 92.3 | 92.5 | 92.9 | 0.7 |
| ResNet152 | 94.1 | 92.9 | 93.0 | 92.9 | 93.1 | 93.4 | 0.6 |
| ResNeXt50 (Xie et al., 2017) | 93.7 | 92.5 | 92.6 | 92.4 | 92.8 | 93.0 | 0.7 |
| ResNeXt101 | 94.5 | 93.5 | 93.5 | 93.3 | 93.5 | 93.9 | 0.6 |
| Wide ResNet50 (Zagoruyko & Komodakis,2016) | 94.1 | 92.9 | 93.0 | 92.9 | 93.1 | 93.4 | 0.7 |
| WideResNet101 | 94.3 | 93.2 | 93.3 | 93.1 | 93.4 | 93.7 | 0.6 |
| Average | 90.7 | 89.3 | 89.4 | 89.2 | 89.6 | 89.9 | 0.7 |
| Model | Top-1 accuracy (%) | Top-5 accuracy (%) | ||||
| Original training | Blurred training | △ | Original training | Blurred training | △ | |
| AlexNet | 56.0 ± 0.3 | 55.3 ± 0.0 | 0.7 | 78.8 ± 0.1 | 78.0 ± 0.1 | 0.9 |
| SqueezeNet | 56.0 ± 0.2 | 54.9 ± 0.1 | 1.1 | 78.6 ± 0.2 | 77.6 ± 0.1 | 1.0 |
| ShuffleNet | 64.7 ± 0.2 | 63.7 ± 0.0 | 1.0 | 85.9 ± 0.0 | 85.1 ± 0.0 | 0.9 |
| VGG11 | 68.9 ± 0.0 | 67.9 ± 0.2 | 1.0 | 88.7 ± 0.0 | 87.9 ± 0.1 | 0.8 |
| VGG13 | 69.9 ± 0.1 | 69.0 ± 0.2 | 1.0 | 89.3 ± 0.1 | 88.6 ± 0.1 | 0.7 |
| VGG16 | 71.7 ± 0.1 | 70.6 ± 0.1 | 1.1 | 90.5 ± 0.1 | 89.8 ± 0.1 | 0.7 |
| VGG19 | 72.4 ± 0.0 | 71.2 ± 0.1 | 1.2 | 90.9 ± 0.1 | 90.1 ± 0.0 | 0.8 |
| MobileNet | 65.4 ± 0.2 | 64.0 ± 0.2 | 1.4 | 86.7 ± 0.1 | 85.6 ± 0.1 | 1.0 |
| DenseNet121 | 75.0 ± 0.1 | 74.1 ± 0.0 | 0.9 | 92.4 ± 0.0 | 91.8 ± 0.0 | 0.6 |
| DenseNet201 | 77.0 ± 0.0 | 76.5 ± 0.1 | 0.5 | 93.5 ± 0.0 | 93.2 ± 0.0 | 0.3 |
| ResNet18 | 69.8 ± 0.2 | 68.6 ± 0.2 | 1.1 | 89.2 ± 0.0 | 88.5 ± 0.1 | 0.7 |
| ResNet34 | 73.1 ± 0.1 | 72.0 ± 0.4 | 1.1 | 91.3 ± 0.0 | 90.6 ± 0.2 | 0.7 |
| ResNet50 | 75.5 ± 0.2 | 74.9 ± 0.1 | 0.6 | 92.5 ± 0.0 | 92.2 ± 0.1 | 0.3 |
| ResNet101 | 77.3 ± 0.1 | 76.6 ± 0.0 | 0.7 | 93.6 ± 0.1 | 93.2 ± 0.0 | 0.4 |
| ResNet152 | 77.9 ± 0.1 | 77.2 ± 0.2 | 0.7 | 93.9 ± 0.0 | 93.6 ± 0.0 | 0.4 |
| Average | 70.0 | 69.1 | 0.9 | 89.1 | 88.4 | 0.7 |
| Cora | CiteSeer | Pubmed | AmazonComputers | |||||||
| ACC | Time | ACC | Time | ACC | Time | ACC | Time | F1 | Time | |
| Random | 80.8±1.7 | 0.20 | 71.2±0.8 | 0.22 | 86.0±3.5 | 0.24 | 81.6±3.0 | 0.16 | 94.3±0.1 | 0.97 |
| BO | 85.1±0.3 | 0.28 | 72.6±0.9 | 0.30 | 88.5±0.3 | 0.31 | 82.3±6.3 | 0.16 | 94.2±0.2 | 0.94 |
| SA | 81.1±0.8 | 0.24 | 74.7±0.2 | 0.25 | 88.9±0.1 | 0.29 | 81.2±6.9 | 0.23 | 94.3±0.5 | 0.97 |
| ENAS | 85.8±0.4 | 0.27 | 74.9±0.2 | 0.39 | 88.6±0.8 | 2.06 | 74.5±1.2 | 0.83 | 92.3±1.0 | 1.98 |
| DARTS | 85.8±0.2 | 0.25 | 75.2±0.3 | 0.25 | 89.1±0.1 | 0.35 | 84.1±1.9 | 0.35 | [OoM] | - |
| GraphNAS | 82.2±3.6 | 3.12 | 74.9±0.6 | 3.99 | 89.2±0.3 | 5.37 | 88.5±2.4 | 2.53 | 89.1±2.9 | 3.03 |
| AutoAttend | 84.6±0.2 | 1.23 | 73.9±0.2 | 1.25 | 84.4±0.7 | 1.55 | 87.3±1.1 | 2.62 | [OoM] | - |
| GASSO | 86.8±1.1 | 0.38 | 75.3±0.7 | 0.33 | 86.3±0.4 | 0.41 | 89.8±0.1 | 0.73 | [OoM] | - |
| FALCON-G | 84.5±0.8 | 0.23 | 74.3±1.7 | 0.24 | 89.2±0.1 | 0.26 | 87.6±0.9 | 0.27 | 93.7±0.4 | 1.11 |
| FALCON-LP | 85.5±1.0 | 0.26 | 74.6±0.1 | 0.26 | 89.0±0.2 | 0.29 | 90.7±0.6 | 0.30 | 94.9±0.2 | 1.00 |
| FALCON | 86.4±0.5 | 0.26 | 76.2±0.4 | 0.28 | 89.3±0.5 | 0.32 | 91.2±0.5 | 0.30 | 95.2±0.2 | 1.15 |
| BRUTEFORCE | 87.0 | 52.5 | 76.0 | 59.7 | 90.0 | 63.0 | 91.4 | 81.5 | 95.5 | >200 |
| p-value | - | 0.051 | - | 0.145 | - | 0.017 | - | 0002 | - | |
| ER-MD | AIDS | OVCAR-8 | MCF-7 | SN12C | NCI109 | Tox21-AhR | Avg. | |
| Random | 77.5±1.6 | 97.0±1.4 | 56.2±0.0 | 58.2±0.3 | 57.4±1.0 | 73.4±0.9 | 75.7±2.0 | 70.8 |
| BO | 77.6±3.5 | 96.1±1.0 | 63.6±0.7 | 60.7±0.0 | 54.8±1.1 | 73.6±1.2 | 75.5±1.1 | 71.7 |
| SA | 75.9±4.2 | 95.4±0.9 | 59.5±3.2 | 56.7±0.8 | 60.4±1.7 | 76.6±5.6 | 76.5±3.0 | 71.6 |
| ENAS | 76.0±2.2 | 97.1±0.4 | 56.0±1.3 | 59.7±0.8 | 66.4±0.6 | 71.2±1.0 | 73.6±0.9 | 71.4 |
| DARTS | 75.0±0.7 | 98.0±0.0 | 56.8±0.3 | 60.2±0.7 | 66.0±0.4 | 73.5±0.2 | 76.0±1.1 | 72.2 |
| GraphNAS | 76.9±3.6 | 95.9±0.8 | 58.7±0.8 | 61.3±5.2 | 60.7±1.5 | 73.6±2.9 | 70.6±4.3 | 71.1 |
| AutoAttend | 73.1±0.8 | 97.4±0.3 | 59.8±0.8 | 64.4±0.2 | 71.8±0.3 | 75.9±1.8 | 74.1±0.9 | 73.8 |
| GASSO | 73.2±0.4 | 95.2±0.7 | 62.3±0.3 | 62.5±0.4 | 70.9±2.3 | 73.9±0.4 | 70.2±3.5 | 72.6 |
| FALCON-G | 78.3±3.0 | 96.3±1.4 | 56.4±1.1 | 62.3±4.5 | 69.8±2.2 | 70.3±6.4 | 72.5±2.8 | 72.3 |
| FALCON-LP | 76.7±2.4 | 96.0±0.2 | 61.5±4.9 | 59.5±5.7 | 70.3±3.8 | 73.1±0.3 | 76.5±2.5 | 73.3 |
| FALCON | 78.4±0.2 | 97.5±1.1 | 66.7±3.4 | 65.5±2.5 | 73.3±0.0 | 78.4±2.3 | 78.5±1.1 | 76.9 |
| BRUTEFORCE | 83.3 | 96.0 | 67.4 | 70.6 | 73.7 | 81.8 | 82.0 | 79.3 |
| p-value | 0.155 | 1 | 0.008 | 0.035 | <0.001 | 0.096 | 0.018 | - |
| Explored node nt: | · | (0.99, 64) | (0.9, 64) | (0.99, 128) | |
| Performance of nt: | … | ++ | : | + | : |
| Prediction on n': | : | 0.90 | 0.77 … | 0.89 | : |
| Dimension importance: | : | [0.5, 0.5] | [0.8, 0.2] … | [0.6, 0.4] | … |
| Small Scale | AIDS,BZR-MD,COX2-MD,DHFR-MD, Mutagenicity, NCI1, NCI109,PTC-MM, PTC-MR |
| Medium/Large Scale | Tox21-AhR,MCF-7, MOLT-4, UACC257, Yeast, NCI-H23, OVCAR-8, P388, PC-3, SF-295, SN12C, SW-620 |
| Type | Variable | Candidate Values |
| Hyper-parameters | Dropout ratio | [0.0, 0.3, 0.6] |
| Architecture | # Pre-process layers | [1,2, 3] |
| # Message passing layers | [2,4,6, 8] | |
| #Post-precess layers | [1,2,3] | |
| Layer connectivity | STACK, SUM, CAT | |
| Activation | ReLU, Swish,Prelu | |
| Batch norm Aggregation | True,False Mean,Max, SUM |
| Type | Variable | Candidate Values |
| Hyper-parameters | Dropout ratio | [0.0, 0.3, 0.6] |
| # Pre-process layers | [1,2,3] | |
| # Message passing layers | [2,4,6,8] | |
| # Post-precess layers | [1,2,3] | |
| Layer connectivity | STACK,SUM,CAT | |
| Activation | ReLU, Swish,Prelu | |
| Batch norm | True,False | |
| Aggregation | Mean,Max,SUM | |
| Node pooling flag (Use node pooling) | True,False | |
| Node pooling type (if applicable) | TopkPool (Gao & Ji,2019), SAGPool (Lee et al.,2019), PANPool (Ma et al.,202O),EdgePool (Diehl,2019) | |
| Node pooling loop | [2,4,6] |
| Variable | Candidate Values |
| Dropout ratio | [0.0, 0.3, 0.6] |
| Layer connectivity | STACK, SUM |
| # Pre-process layers | [1, 2, 3] |
| #Message passing layers | [2,4,6, 8] |
| # Post-precess layers | [1,2,3] |
| Activation | ReLU,Swish,Prelu |
| Batch norm | True, False |
| Aggregation | Mean, Max, SUM |
| #Nodes | #Edges (Undirected) | Ave. Degree | Diameter | construction time (s) | |
| DG-1 | 5,832 | 78.732 | 13.5 | 13 | 13 |
| DG-2 | 58,320 | 1,070,172 | 18.4 | 17 | 147 |
| DHFR-MD | COX2-MD | Mutagenicity | MOLT-4 | NCI-H23 | PTC-MR | P388 | |
| Random | 59.0±5.2 | 63.2±1.9 | 77.1±1.9 | 58.6±0.7 | 58.4±2.1 | 59.9±5.7 | 63.9±0.7 |
| BO | 55.1±0.0 | 71.6±5.5 | 78.3±0.7 | 58.1±2.0 | 63.6±0.0 | 58.8±7.7 | 68.9±0.0 |
| SA | 56.0±7.1 | 67.4±3.1 | 81.1±0.3 | 54.8±1.1 | 56.7±3.2 | 59.4±6.2 | 74.4±1.0 |
| ENAS | 53.5±3.7 | 57.9±1.7 | 75.0±1.6 | 61.5±0.1 | 61.2±1.1 | 59.8±1.6 | 68.3±1.2 |
| DARTS | 55.8±6.3 | 70.4±3.2 | 74.4±0.7 | 61.4±0.7 | 63.5±1.9 | 59.3±0.5 | 70.8±0.7 |
| GraphNAS | 61.6±4.3 | 63.9±2.5 | 80.2±1.5 | 62.1±1.0 | 62.4±3.9 | 58.6±6.7 | 68.2±2.5 |
| AutoAttend | 63.3±0.9 | 68.8±0.7 | 79.6±0.1 | 59.5±0.2 | 61.8±0.2 | 57.8±0.6 | 74.9±0.5 |
| GASSO | 60.9±2.3 | 68.5±2.0 | 75.1±0.5 | 57.4±0.9 | 64.7±1.5 | 51.9±5.2 | 71.3±1.4 |
| FALCON-G | 58.5±8.8 | 67.3±3.6 | 79.8±1.7 | 62.8±3.3 | 62.2±1.2 | 55.6±6.9 | 74.2±3.7 |
| FALCON-LP | 61.4±1.5 | 66.8 ±3.6 | 80.2±0.8 | 58.5 ±8.8 | 63.7±4.1 | 55.6±2.0 | 75.1±1.1 |
| FALCON | 63.6±7.9 | 67.3±3.2 | 81.1±0.5 | 64.4±4.0 | 66.6±3.3 | 60.0±1.4 | 77.0±1.4 |
| (cont.) PTC-MM | PC-3 | SF-295 | NCI1 | SW-620 | UACC257 | Yeast | Avg. |
| 52.5±5.6 | 60.4±0.0 | 55.3±0.5 | 77.5±0.3 | 57.5±2.7 | 61.1±0.5 | 53.3±0.0 | 61.3 |
| 60.1±1.5 | 59.7±0.1 | 60.6±0.3 | 77.3±0.0 | 63.8±0.9 | 60.8±0.0 | 55.0±0.0 | 63.7 |
| 58.1±4.6 | 69.0±2.0 | 55.3±0.6 | 79.6±5.3 | 58.2±2.1 | 64.2±0.1 | 53.8±1.2 | 63.4 |
| 52.4±2.9 | 62.2±1.1 | 60.9±2.2 | 77.2±1.2 | 64.8±2.2 | 64.7±0.5 | 63.0±1.1 | 63.0 |
| 52.6±4.1 | 61.6±0.5 | 62.2±1.1 | 66.3±3.0 | 66.0±0.3 | 65.7±0.2 | 61.4±1.4 | 63.7 |
| 54.8±3.9 | 68.6±2.6 | 65.0±3.1 | 78.1±3.6 | 61.8±4.4 | 61.5±5.1 | 57.2±1.1 | 64.6 |
| 64.7±1.2 | 66.2±0.3 | 64.2±0.9 | 79.3±1.6 | 65.5±0.4 | 57.1±0.5 | 59.6±0.8 | 65.9 |
| 63.2±0.7 | 66.0±1.8 | 65.8±0.4 | 76.6±0.6 | 64.7±1.1 | 62.7±0.2 | 60.1±0.4 | 64.9 |
| 55.2±2.4 | 65.1±2.8 | 64.3±2.1 | 80.6±0.8 | 62.0±3.0 | 61.0±3.8 | 57.5±1.5 | 64.7 |
| 56.8±6.1 | 68.1±0.3 | 63.5±3.4 | 80.4±0.5 | 65.2±1.7 | 54.5±1.1 | 56.6±1.7 | 64.7 |
| 57.1±0.1 | 71.0±1.7 | 64.4±0.4 | 80.9±0.8 | 66.6±2.1 | 66.7±2.1 | 58.1±1.7 | 67.5* |
| Dataset | Average Param (k) | Best design | Test perfor- mance (%) | |
| ogbn-arxiv | 44.5 | FALCON | (0.0,1,4,2, STACK, Swish, True, Mean) | 70.36 |
| BRUTEFORCE | (0.3,1,4,2, SUM, Relu, True,Mean) | 70.51 | ||
| Cora | 77.8 | FALCON BRUTEFORCE | (0.0, 1, 6,1, SUM, Swish,False,Mean) (0.0,1,4,1, STACK, Swish,False,Mean) | 87.18 86.99 |
| Citeseer | 289.0 | FALCON BRUTEFORCE | (0.3,1,2,1,SUM, Prelu, True, Mean) (0.3,1,2,2, SUM, Prelu, False, Mean) | 76.19 75.99 |
| Pubmed | 57.8 | FALCON BRUTEFORCE | (0.3,1,8,1,SUM, Relu, True,Add) (0.3,1,8,1,SUM, Relu, True,Add) | 90.04 90.04 |
| AmazonComputers | 46.2 | FALCON BRUTEFORCE | (0.0,1,4,1,STACK,Swish,True,Mean) (0.0,3,4,2,STACK,Prelu,True,Mean) | 91.64 91.35 |
| 47.2 | FALCON BRUTEFORCE | (0.0,3,4,2, STACK,Prelu, True,ARMAConv) (0.0,3,4,2, STACK,Prelu,True,ARMAConv) | 95.46 95.46 |
| Cora | CiteSeer | ER-MD | AIDS | |
| Bruteforce (max) | 87.0 | 76.0 | 83.3 | 96.0 |
| Confidence interval length | 0.2 | 0.1 | 0.6 | 0.0 |
| Bruteforce-bootstrap | 87.0 | 76.4 | 83.8 | 95.7 |
| Dataset | Best design | Test perfor- mance (%) | |
| AIDS | FALCON | (0.3,1,4,2, SUM,Prelu, True,Add, NoPool) | 99.02 |
| BRUTEFORCE | (0.0,3,8,2, STACK,Relu, True,Add, SAGPool, 2) | 95.97 | |
| COX2-MD | FALCON | (0.0,2,6,1,SUM,Swish,True,Add,EdgePool,2) | 69.87 |
| BRUTEFORCE | (0.0,1, 4, 3, STACK,Prelu, True,Max, TopkPool, 2) | 65.39 | |
| DHFR-MD | FALCON | (0.3,2,4,2,SUM,Swish,True,Mean,PANPool,2) | 71.33 |
| ER-MD | BRUTEFORCE | (0.0,1,4,3, CAT, Prelu, True, Mean, SAGPool,4) | 56.22 |
| FALCON | (0.3,3,4,2, CAT, Relu, True,Add,PANPool,2) | 81.67 | |
| MCF-7 | BRUTEFORCE | (0.0,3,5,3, CAT, Prelu, True,Max,PANPool,6) | 83.33 |
| FALCON BRUTEFORCE | (0.6,1,8, 2, CAT, Swish, True, Add, NoPool) (0.0,1,2, 6, SUM,Prelu, True,Add, NoPool) | 67.85 70.61 | |
| MOLT-4 | |||
| FALCON BRUTEFORCE | (0.0,3,6,2, SUM,Prelu,True,Max,NoPool) (0.0,1,6,1, SUM,Prelu,True,Add,NoPool) | 69.03 70.30 | |
| Mutagenicity | |||
| FALCON BRUTEFORCE | (0.0,2, 6,1, CAT,Relu, True,Add, NoPool) (0.0,2,4,2, CAT, Relu, True,Add, PANPool, 2) | 81.73 | |
| NCI1 | 81.17 | ||
| FALCON BRUTEFORCE | (0.0,2,4,2, SUM, Swish, True, Max, EdgePool,2) (0.0,1,4,2,STACK,Prelu,True,Add,EdgePool,4) | 80.13 | |
| NCI109 | 82.81 | ||
| FALCON BRUTEFORCE | (0.0,2,8,2, STACK,Prelu, True,Max,NoPool) (0.0,3, 6,2, STACK,Prelu, True,Add,EdgePool, 6) | 79.98 81.77 | |
| NCI-H23 | FALCON | ||
| BRUTEFORCE | (0.0,2,6,1,CAT,Swish,True,Add,NoPool) (0.0,2,6,1,CAT,Swish,True,Add,NoPool) | 71.14 71.14 | |
| OVCAR-8 | FALCON | ||
| BRUTEFORCE | (0.6,2,6,2, CAT, Swish, True,Add, NoPool) (0.6,1,2,3,CAT,Swish,True,Add,NoPool) | 68.21 67.40 | |
| P388 | FALCON | ||
| BRUTEFORCE | (0.0,2,6,2,CAT,Prelu,True,Add,NoPool) | 78.75 | |
| PC-3 | FALCON | (0.0,2,6,2, CAT,Prelu,True,Add,NoPool) | 78.75 |
| BRUTEFORCE | (0.0,2,6,1, SUM,Relu,True,Max,NoPool) | 73.28 | |
| PTC-MM | (0.0,2,6,1, SUM,Relu,True,Max,NoPool) | 73.28 | |
| FALCON | (0.0,2,4,2,STACK,Swish,True,Max,EdgePool,2) | 56.96 | |
| PTC-MR | BRUTEFORCE | (0.0,3,6,1, SUM,Swish, True,Mean,EdgePool,2) | 52.93 |
| FALCON | (0.0,2, 4, 2, SUM, Swish, True, Mean, TopkPool, 2) | 60.56 | |
| SF-295 | BRUTEFORCE | (0.3,1,6,2, SUM,Relu,True,Add,NoPool) | 63.06 |
| FALCON | (0.6,2, 6, 2, CAT, Swish, True,Add, NoPool) | 64.75 | |
| BRUTEFORCE | (0.0, 3, 8,3, SUM, Prelu, True, Max, PANPool, 4) | 66.47 | |
| FALCON | (0.0,1, 8,1, CAT, Swish, True,Add, NoPool) | 73.34 | |
| SN12C SW-620 | BRUTEFORCE | (0.0,1,8,3, CAT,Prelu, True,Add,EdgePool, 6) | 73.73 |
| FALCON | (0.0,2,4,2,STACK,Prelu,True,Max,NoPool) | 69.26 | |
| BRUTEFORCE | (0.0,2,4,2, STACK,Prelu,True,Max,NoPool) | 69.26 | |
| FALCON | (0.0,2,4,2, SUM,Prelu, True,Add,EdgePool,2) | 79.10 | |
| Tox21-AhR | BRUTEFORCE | (0.0,3,6,2, SUM,Prelu, True,Add, NoPool) | 82.02 |
| FALCON | (0.3,2,6,2, CAT,Swish,True,Max,NoPool) | 67.94 | |
| UACC257 | BRUTEFORCE | (0.0,3,6,1,SUM,Prelu,True,Max,NoPool) | 70.24 |
| FALCON | (0.6,1,8,1, CAT, Prelu, True,Add, NoPool) | 59.60 | |
| Yeast | BRUTEFORCE | ||
| (0.0,1,2,3, SUM, Swish, True,Add, EdgePool, 2) | 60.41 |
| Variable | Candidate Values |
| Momentum (SGD) | [0.5, 0.9, 0.99] |
| Weight decay | [le-4, 5e-4,1e-3, 5e-3] |
| Batch size | [32, 64, 128,256] |
| LR decay factor | [0.1, 0.2, 0.5, 0.8] |
| LR decay patience | [3,5,10] |
| Average Error (%) |
| Random 4.13 |
| BO 4.95 |
| SA 4.04 |
| FALCON 3.80 |
| Average Error (%) | Search Cost (GPU days) | |
| Random | 10.43 | 0.64 |
| BO | 9.83 | 0.67 |
| SA | 10.16 | 0.62 |
| ENAS-micro | 9.20 | 0.77 |
| DARTS-micro | 8.97 | 0.89 |
| FALCON | 8.87 | 0.81 |
| supervision | params (M) | success (%) | |
| R3M | video-text | 23 | 31.3 |
| CLIP | image-text | 86 | 18.8 |
| Ours | image-only | 22 | 68.8 |
| 86 307 | 93.8 100.0 |
| Language | Constraint | Example of partial program | Valid/Invalid Examples |
| SQL | A valid identifier must follow after AS. | SELECT Name, Role FROM User AS^ | U √ T1 √ 2× |
| Column names must come from schema,even behind aliases. | SELECT U.Name FROM User AS U WHERE U. ^ | Name√ DoB√ Birthday X | |
| Vega-Lite | Data fields must be used with compatible types. Do not facet on field with too many distinct values (breaks | {"x": {"field": "Category", "type":^ {"column":{"field": ^ | "nominal" "temporal" X "Category" "ZipCode" |
| SMCalFlow | rendering). Type-check parameters of all | (Yield | Takeout√ |
| API functions. Track declared variables and | (PlaceHasFeature(> (let (x 85) | IsWindy X List.Apply X x√ |
| Model | Exec. | SQL | Vega-Lite | SMCalFlow | |||||
| Valid | Dist. | Acc. | Valid | Dist. | Acc. | Valid | Dist. | ||
| Andreas et al. (2020) | = | = | = | 72%(s) | = | ||||
| Srinivasan et al. (2021) | = | = | = | 64%(s) | = | = | = | = | = |
| Rubin & Berant (2021) | 71%(s) | - | = | ||||||
| Scholak et al. (2021) | 79%(s) | 98% | 1 | 1 | = | ||||
| GPT-3 13B | 16% | 43% | 0.42 | 14% | 55% | 0.51 | 38% | 76% | 0.43 |
| ”+ CSD | 20% | 66% | 0.44 | 17% | 100% | 0.48 | 40% | 95% | 0.40 |
| "+TST | 14% | 48% | 0.42 | 1 | - | - | 60% | 88% | 0.22 |
| ”+ CSD + TST | 19% | 72% | 0.43 | - | - | - | 63% | 98% | 0.17 |
| GPT-3175B | 28% | 49% | 0.36 | 20% | 67% | 0.36 | 44% | 77% | 0.41 |
| ”+ CSD | 35% | 73% | 0.36 | 25% | 100% | 0.32 | 45% | 97% | 0.37 |
| ”+ TST | 31% | 56% | 0.35 | 1 | - | - | 60% | 88% | 0.24 |
| "+ CSD + TST | 37% | 76% | 0.34 | - | - | - | 66% | 97% | 0.18 |
| Codex 175B | 56% | 73% | 0.25 | 39% | 87% | 0.24 | 45% | 79% | 0.37 |
| ”+ CSD | 61% | 85% | 0.23 | 40% | 99% | 0.23 | 46% | 97% | 0.33 |
| ”+TST | 60% | 81% | 0.23 | 1 | 1 | - | 63% | 90% | 0.21 |
| "+ CSD + TST | 64% | 85% | 0.23 | - | 1 | - | 63% | 99% | 0.19 |
| Constructor | Description | Derivative w.r.t c' ∈∑ |
| Q | Empty regular expression (matches no string). | Q |
| E | Matches only the empty string. | Q |
| C | Matches a single character c | ∈ if c= c',or & otherwise. If the derivative of R1 w.r.t. c' is not , |
| R1R2 | Concatenation of two regular ex- pressions R1 and R2 | then it's the concatenation of that with R2. Otherwise,if Rl matches ∈, then it is sim- ply the derivative of R2 w.r.t. c'. If not, |
| R1R2 | Union of two regular expressions R1 and R2 | then the result is &. Union of the derivatives of R1 and R2 W.r.t. c'(if one becomes &,simply return the other). |
| R* | Klenee star - any number of repeti- tions of R | If the derivative of R w.r.t. c is not Q, then return the concatenation of that derivative with R*. Otherwise,return @. |
| ID Dataset | OODDataset | Average | ||||||||
| iNaturalist | SUN | Places | Texture | |||||||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | |
| CUB-200|801 | 9.83 | 98.24 | 4.93 | 99.10 | 6.65 | 98.57 | 6.97 | 98.75 | 7.09 | 98.66 |
| Stanford-Cars 四 | 0.05 | 99.77 | 0.02 | 99.95 | 0.24 | 99.89 | 0.02 | 99.96 | 0.08 | 99.89 |
| Food-101 | 0.64 | 99.78 | 0.90 | 99.75 | 1.86 | 99.58 | 4.04 | 98.62 | 1.86 | 99.43 |
| Oxford-Pet [57 | 2.85 | 99.38 | 1.06 | 99.73 | 2.11 | 99.56 | 0.80 | 99.81 | 1.70 | 99.62 |
| ImageNet-10 | 0.12 | 99.80 | 0.29 | 99.79 | 0.88 | 99.62 | 0.04 | 99.90 | 0.33 | 99.78 |
| ImageNet-20 | 1.02 | 99.66 | 2.55 | 99.50 | 4.40 | 99.11 | 2.43 | 99.03 | 2.60 | 99.32 |
| ImageNet-100 | 18.13 | 96.77 | 36.45 | 94.54 | 34.52 | 94.36 | 41.22 | 92.25 | 32.58 | 94.48 |
| Method | OOD Dataset | Average | ||||||||
| iNaturalist FPR95↓ | SUN | Places | Texture | |||||||
| AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | ||
| Requires training (or w. fine-tuning) | ||||||||||
| MOS 32 (BiT) | 9.28 | 98.15 | 40.63 | 92.01 | 49.54 | 89.06 | 60.43 | 81.23 | 39.97 | 90.11 |
| Fort et al. 园 (ViT-B) | 15.07 | 96.64 | 54.12 | 86.37 | 57.99 | 85.24 | 53.32 | 84.77 | 45.12 | 88.25 |
| Fort et al. 园 (ViT-L) | 15.74 | 96.51 | 52.34 | 87.32 | 55.14 | 86.48 | 51.38 | 85.54 | 43.65 | 88.96 |
| Energy 481 (CLIP-B) | 21.59 | 95.99 | 34.28 | 93.15 | 36.64 | 91.82 | 51.18 | 88.09 | 35.92 | 92.26 |
| Energy 48 (CLIP-L) | 10.62 | 97.52 | 30.46 | 93.83 | 32.25 | 93.01 | 44.35 | 89.64 | 29.42 | 93.50 |
| MSP 因 (CLIP-B) MSP | 40.89 | 88.63 | 65.81 | 81.24 | 67.90 | 80.14 | 64.96 | 78.16 | 59.89 | 82.04 |
| 因 (CLIP-L) | 34.54 | 92.62 | 61.18 | 83.68 | 59.86 | 84.10 | 59.27 | 82.31 | 53.71 | 85.68 |
| Zero-shot (no training required) | ||||||||||
| MCM(CLIP-B) | 30.91 | 94.61 | 37.59 | 92.57 | 44.69 | 89.77 | 57.77 | 86.11 | 42.74 | 90.77 |
| MCM(CLIP-L) | 28.38 | 94.95 | 29.00 | 94.14 | 35.42 | 92.00 | 59.88 | 84.88 | 38.17 | 91.49 |
| Method | ID OOD | ImageNet-10 ImageNet-20 | ImageNet-20 ImageNet-10 | Waterbirds Spurious OOD |
| FPR95/AUROC | FPR95 /AUROC | FPR95/AUROC | ||
| MSP 四 (fine-tuning) | 9.38 /98.31 | 12.51 /97.70 | 39.57 /90.99 | |
| Mahalanobis [42] (visual only) | 78.32/85.60 | 43.03 /89.94 | 2.21 /99.55 | |
| MCM (zero-shot) | 5.00 /98.71 | 12.91 / 98.09 | 5.87 /98.36 |
| ID Dataset | OOD Dataset | Average | ||||||||
| iNaturalist | SUN | Places | Texture | |||||||
| FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | |
| Stanford-Cars39 | 0.00 | 100 | 0.02 | 99.99 | 0.26 | 99.94 | 0.00 | 100 | 0.07 | 99.98 |
| Food-101 | 0.56 | 99.86 | 0.09 | 99.95 | 0.49 | 99.88 | 8.33 | 97.44 | 2.37 | 99.28 |
| Oxford-Pet 57] | 0.02 | 99.98 | 0.05 | 99.97 | 0.20 | 99.94 | 0.27 | 99.91 | 0.14 | 99.95 |
| ImageNet-10 | 2.40 | 99.42 | 1.79 | 99.55 | 2.83 | 99.32 | 1.86 | 99.56 | 2.22 | 99.46 |
| ImageNet-20 | 14.96 | 97.87 | 13.10 | 97.97 | 14.21 | 97.67 | 13.46 | 97.32 | 13.93 | 97.71 |
| ImageNet-1k | 61.66 | 89.31 | 64.39 | 87.43 | 63.67 | 85.95 | 86.61 | 71.68 | 69.08 | 83.59 |
| Model | OOD Dataset | Average | ||||||||
| iNaturalist | SUN | Places | Texture | |||||||
| FPR95↓ | AUROC个 | FPR95↓ | AUROC个 | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | FPR95↓ | AUROC↑ | |
| RN50x4 | 44.51 | 91.51 | 35.11 | 92.84 | 43.74 | 89.60 | 57.73 | 85.93 | 45.27 | 89.97 |
| CLIP-B/16 | 30.91 | 94.61 | 37.59 | 92.57 | 44.69 | 89.77 | 57.77 | 86.11 | 42.74 | 90.77 |
| A photo of a<label>. A blurryphoto of a<label>. A photo of many <label>. A photo of the large <label>. |
| A photo of the small <label>. |
| Methods | P19 | P12 | PAM | |||||
| AUROC | AUPRC | AUROC | AUPRC | Accuracy | Precision | Recall | F1 score | |
| Transformer | 83.2 ± 1.3 | 47.6±3.8 | 65.1 ± 5.6 | 95.7 ±1.6 | 83.5±1.5 | 84.8 ±1.5 | 86.0 ± 1.2 | 85.0 ± 1.3 |
| Trans-mean | 84.1 ± 1.7 | 47.4 ± 1.4 | 66.8±4.2 | 95.9 ± 1.1 | 83.7±2.3 | 84.9 ± 2.6 | 86.4 ± 2.1 | 85.1 ± 2.4 |
| GRU-D | 83.9 ±1.7 | 46.9 ± 2.1 | 67.2 ±3.6 | 95.9 ± 2.1 | 83.3 ±1.6 | 84.6 ±1.2 | 85.2 ± 1.6 | 84.8 ± 1.2 |
| SeFT | 78.7 ± 2.4 | 31.1 ± 2.8 | 66.8 ±0.8 | 96.2 ±0.2 | 67.1 ± 2.2 | 70.0 ± 2.4 | 68.2 ±1.5 | 68.5 ±1.8 |
| mTAND | 80.4 ±1.3 | 32.4 ±1.8 | 65.3 ±1.7 | 96.5 ±1.2 | 74.6 ± 4.3 | 74.3 ± 4.0 | 79.5 ± 2.8 | 76.8 ± 3.4 |
| IP-Net | 84.6 ±1.3 | 38.1 ± 3.7 | 72.5± 2.4 | 96.7 ± 0.3 | 74.3 ± 3.8 | 75.6 ± 2.1 | 77.9 ± 2.2 | 76.6± 2.8 |
| DGM²-0 | 86.7 ± 3.4 | 44.7 ± 11.7 | 71.2 ± 2.5 | 96.9 ± 0.4 | 82.4± 2.3 | 85.2 ±1.2 | 83.9 ± 2.3 | 84.3 ± 1.8 |
| MTGNN | 81.9 ± 6.2 | 39.9 ± 8.9 | 67.5 ± 3.1 | 96.4± 0.7 | 83.4 ±1.9 | 85.2 ±1.7 | 86.1 ± 1.9 | 85.9 ± 2.4 |
| RAINDROP | 87.0 ± 2.3 | 51.8 ± 5.5 | 72.1 ± 1.3 | 97.0 ± 0.4 | 88.5 ± 1.5 | 89.9 ± 1.5 | 89.9 ± 0.6 | 89.8 ± 1.0 |
| Missing sensor ratio | Methods | PAM (Setting 2: leave-fixed-sensors-out) | PAM(Setting 3: leave-random-sensors-out) | ||||||
| Accuracy | Precision | Recall | F1 score | Accuracy | Precision | Recall | F1 score | ||
| 10% | Transformer | 60.3± 2.4 | 57.8 ± 9.3 | 59.8± 5.4 | 57.2 ±8.0 | 60.9 ± 12.8 | 58.4 ± 18.4 | 59.1 ± 16.2 | 56.9 ±18.9 |
| Trans-mean | 60.4 ± 11.2 | 61.8 ± 14.9 | 60.2 ±13.8 | 58.0 ±15.2 | 62.4 ± 3.5 | 59.6± 7.2 | 63.7 ±8.1 | 62.7 ± 6.4 | |
| GRU-D | 65.4 ± 1.7 | 72.6 ± 2.6 | 64.3 ± 5.3 | 63.6±0.4 | 68.4± 3.7 | 74.2 ± 3.0 | 70.8 ± 4.2 | 72.0 ±3.7 | |
| SeFT | 58.9 ±2.3 | 62.5 ± 1.8 | 59.6±2.6 | 59.6 ± 2.6 | 40.0 ± 1.9 | 40.8± 3.2 | 41.0 ±0.7 | 39.9 ± 1.5 | |
| mTAND | 58.8 ±2.7 | 59.5± 5.3 | 64.4 ± 2.9 | 61.8 ± 4.1 | 53.4 ± 2.0 | 54.8 ± 2.7 | 57.0 ± 1.9 | 55.9 ± 2.2 | |
| RAINDROP | 77.2 ± 2.1 | 82.3 ± 1.1 | 78.4 ±1.9 | 75.2 ± 3.1 | 76.7 ± 1.8 | 79.9 ± 1.7 | 77.9 ± 2.3 | 78.6 ±1.8 | |
| 20% | Transformer | 63.1 ± 7.6 | 71.1 ± 7.1 | 62.2±8.2 | 63.2±8.7 | 62.3 ± 11.5 | 65.9 ± 12.7 | 61.4 ± 13.9 | 61.8 ± 15.6 |
| Trans-mean | 61.2 ± 3.0 | 74.2 ± 1.8 | 63.5 ± 4.4 | 64.1 ± 4.1 | 56.8 ± 4.1 | 59.4 ± 3.4 | 53.2 ±3.9 | 55.3 ±3.5 | |
| GRU-D | 64.6 ± 1.8 | 73.3 ± 3.6 | 63.5± 4.6 | 64.8± 3.6 | 64.8 ± 0.4 | 69.8±0.8 | 65.8± 0.5 | 67.2 ± 0.0 | |
| SeFT | 35.7 ±0.5 | 42.1 ± 4.8 | 38.1 ± 1.3 | 35.0±2.2 | 34.2 ± 2.8 | 34.9 ± 5.2 | 34.6 ± 2.1 | 33.3 ±2.7 | |
| mTAND | 33.2±5.0 | 36.9 ± 3.7 | 37.7 ± 3.7 | 37.3 ± 3.4 | 45.6 ± 1.6 | 49.2 ± 2.1 | 49.0 ± 1.6 | 49.0 ± 1.0 | |
| RAINDROP | 66.5± 4.0 | 72.0 ± 3.9 | 67.9 ±5.8 | 65.1 ± 7.0 | 71.3 ± 2.5 | 75.8 ± 2.2 | 72.5± 2.0 | 73.4 ± 2.1 | |
| 30% | Transformer | 31.6 ± 10.0 | 26.4 ± 9.7 | 24.0 ±10.0 | 19.0 ± 12.8 | 52.0 ± 11.9 | 55.2 ±15.3 | 50.1 ± 13.3 | 48.4 ± 18.2 |
| Trans-mean | 42.5±8.6 | 45.3 ± 9.6 | 37.0 ± 7.9 | 33.9 ±8.2 | 65.1 ± 1.9 | 63.8 ± 1.2 | 67.9 ± 1.8 | 64.9 ± 1.7 | |
| GRU-D | 45.1 ± 2.9 | 51.7 ± 6.2 | 42.1 ± 6.6 | 47.2 ± 3.9 | 58.0±2.0 | 63.2 ± 1.7 | 58.2 ± 3.1 | 59.3±3.5 | |
| SeFT | 32.7 ± 2.3 | 27.9 ± 2.4 | 34.5± 3.0 | 28.0 ± 1.4 | 31.7 ± 1.5 | 31.0 ± 2.7 | 32.0 ± 1.2 | 28.0 ±1.6 | |
| mTAND | 27.5 ± 4.5 | 31.2 ± 7.3 | 30.6± 4.0 | 30.8 ± 5.6 | 34.7 ± 5.5 | 43.4 ± 4.0 | 36.3 ± 4.7 | 39.5 ± 4.4 | |
| RAINDROP | 52.4± 2.8 | 60.9 ± 3.8 | 51.3 ± 7.1 | 48.4 ± 1.8 | 60.3 ±3.5 | 68.1 ± 3.1 | 60.3 ±3.6 | 61.9 ± 3.9 | |
| 40% | Transformer | 23.0±3.5 | 7.4 ± 6.0 | 14.5 ± 2.6 | 6.9 ± 2.6 | 43.8 ± 14.0 | 44.6 ± 23.0 | 40.5 ±15.9 | 40.2 ± 20.1 |
| Trans-mean | 25.7± 2.5 | 9.1 ± 2.3 | 18.5 ± 1.4 | 9.9 ± 1.1 | 48.7 ± 2.7 | 55.8± 2.6 | 54.2 ± 3.0 | 55.1 ± 2.9 | |
| GRU-D | 46.4 ± 2.5 | 64.5 ± 6.8 | 42.6 ± 7.4 | 44.3 ± 7.9 | 47.7 ± 1.4 | 63.4 ± 1.6 | 44.5± 0.5 | 47.5± 0.0 | |
| SeFT | 26.3 ± 0.9 | 29.9 ± 4.5 | 27.3 ± 1.6 | 22.3 ± 1.9 | 26.8 ± 2.6 | 24.1 ± 3.4 | 28.0 ±1.2 | 23.3±3.0 | |
| mTAND | 19.4 ± 4.5 | 15.1 ± 4.4 | 20.2 ±3.8 | 17.0 ± 3.4 | 23.7 ±1.0 | 33.9 ± 6.5 | 26.4 ± 1.6 | 29.3 ± 1.9 | |
| RAINDROP | 52.5±3.7 | 53.4± 5.6 | 48.6 ± 1.9 | 44.7 ± 3.4 | 57.0 ± 3.1 | 65.4 ± 2.7 | 56.7 ± 3.1 | 58.9 ± 2.5 | |
| 50% | Transformer | 21.4 ± 1.8 | 2.7 ±0.2 | 12.5 ± 0.4 | 4.4 ± 0.3 | 43.2 ± 2.5 | 52.0± 2.5 | 36.9 ± 3.1 | 41.9 ± 3.2 |
| Trans-mean | 21.3 ± 1.6 | 2.8 ±0.4 | 12.5 ± 0.7 | 4.6±0.2 | 46.4 ± 1.4 | 59.1 ± 3.2 | 43.1 ± 2.2 | 46.5 ± 3.1 | |
| GRU-D | 37.3 ± 2.7 | 29.6± 5.9 | 32.8±4.6 | 26.6±5.9 | 49.7 ± 1.2 | 52.4±0.3 | 42.5 ± 1.7 | 47.5 ± 1.2 | |
| SeFT | 24.7 ± 1.7 | 15.9 ± 2.7 | 25.3±2.6 | 18.2 ± 2.4 | 26.4 ± 1.4 | 23.0 ±2.9 | 27.5± 0.4 | 23.5± 1.8 | |
| mTAND | 16.9 ± 3.1 | 12.6 ± 5.5 | 17.0 ± 1.6 | 13.9 ± 4.0 | 20.9 ± 3.1 | 35.1 ± 6.1 | 23.0±3.2 | 27.7 ± 3.9 | |
| RAINDROP | 46.6 ± 2.6 | 44.5 ± 2.6 | 42.4 ± 3.9 | 38.0±4.0 | 47.2 ± 4.4 | 59.4 ± 3.9 | 44.8±5.3 | 47.6± 5.2 | |
| Datasets | #-samples | #-sensors | #-timestamps | #-classes | Static info | Missing ratio (%) |
| P19 | 38,803 | 34 | 60 | 2 | True | 94.9 |
| P12 | 11,988 | 36 | 215 | 2 | True | 88.4 |
| PAM | 5,333 | 17 | 600 | 8 | False | 60.0 |
| Models | Missing ratio | |||||||||||
| 0% | 10% | 20% | 30% | 40% | 50% | |||||||
| AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | |
| Transformer | 83.2 ± 1.3 | 47.6 ± 3.8 | 77.4 ± 3.5 | 38.2 ± 4.2 | 75.7 ± 3.4 | 35.2 ± 5.4 | 75.1 ± 3.5 | 35.5 ± 4.4 | 75.3 ± 3.5 | 36.2 ± 4.2 | 74.9 ± 3.1 | 35.5±5.0 |
| Trans-mean | 84.1 ± 1.7 | 47.4 ± 1.4 | 79.2 ± 2.7 | 40.6 ± 5.7 | 79.8 ± 2.5 | 38.3 ± 2.8 | 76.9 ± 2.4 | 37.5 ± 5.9 | 76.4 ± 2.0 | 36.3 ±5.8 | 74.1 ± 2.3 | 41.3 ± 4.7 |
| GRU-D | 83.9 ± 1.7 | 46.9 ± 2.1 | 79.6± 2.2 | 37.4 ± 2.5 | 77.5 ± 3.1 | 36.5± 4.6 | 76.6 ± 2.9 | 35.1 ± 2.4 | 74.6 ± 2.7 | 35.9±2.7 | 74.1 ± 2.9 | 33.2 ± 3.8 |
| SeFT | 78.7 ± 2.4 | 31.1 ± 2.8 | 77.3 ± 2.4 | 25.5 ± 2.3 | 63.5± 2.0 | 14.0 ± 1.1 | 62.3 ± 2.1 | 12.9 ± 1.2 | 57.8± 1.7 | 9.8 ± 1.1 | 56.0 ± 3.1 | 7.8±1.3 |
| mTAND | 80.4 ± 1.3 | 32.4 ± 1.8 | 79.7 ± 2.2 | 29.0 ± 4.3 | 77.8 ± 1.9 | 25.3 ± 2.4 | 77.7 ± 1.9 | 27.8 ± 2.6 | 79.4 ± 2.0 | 32.1 ± 2.1 | 77.3 ± 2.1 | 27.0 ± 2.5 |
| RAINDROP | 87.0 ± 2.3 | 51.8± 5.5 | 84.3 ± 2.5 | 46.1 ± 3.5 | 81.9 ± 2.1 | 45.2 ± 6.4 | 81.4 ± 2.1 | 43.7 ± 7.2 | 81.8±2.2 | 44.9 ± 6.6 | 79.7 ± 1.9 | 43.8 ± 5.6 |
| Models | Missing ratio | |||||||||||
| 0% | 10% | 20% | 30% | 40% | 50% | |||||||
| AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | |
| Transformer | 83.2 ± 1.3 | 47.6 ± 3.8 | 82.2 ±2.7 | 46.8 ± 3.5 | 81.6± 3.5 | 42.5 ±8.5 | 81.3 ± 3.1 | 42.1 ± 4.5 | 80.2 ±2.9 | 41.9 ± 6.8 | 79.2 ± 1.9 | 43.7 ± 3.7 |
| Trans-mean | 84.1 ± 1.7 | 47.4 ± 1.4 | 82.5± 3.7 | 44.7 ± 6.8 | 81.7± 2.0 | 45.9 ± 3.6 | 81.2 ± 2.2 | 43.2 ± 6.3 | 80.2 ± 1.7 | 41.5 ± 4.8 | 79.8 ± 3.1 | 39.3 ± 5.1 |
| GRU-D | 83.9 ± 1.7 | 46.9 ± 2.1 | 81.2 ± 3.4 | 46.4 ± 2.7 | 78.6 ± 4.1 | 43.3 ± 2.4 | 76.3 ± 2.5 | 28.5 ± 2.1 | 74.2 ± 2.7 | 29.6± 3.1 | 74.6 ± 3.5 | 26.5± 4.2 |
| SeFT | 78.7 ± 2.4 | 31.1 ± 2.8 | 76.8±2.2 | 28.3± 2.5 | 77.0± 2.2 | 24.1 ± 2.4 | 75.2 ± 2.2 | 22.5±3.0 | 73.6± 2.7 | 18.3±3.2 | 72.6 ± 2.5 | 15.7 ± 1.9 |
| mTAND | 80.4 ± 1.3 | 32.4 ± 1.8 | 75.2± 2.5 | 24.5 ± 2.4 | 74.4 ± 3.5 | 24.6± 3.5 | 74.2 ± 3.2 | 22.6 ± 2.3 | 74.1 ± 2.6 | 23.1± 3.6 | 73.9 ± 3.7 | 24.6 ± 3.7 |
| RAINDROP | 87.0 ± 2.3 | 51.8 ± 5.5 | 85.5± 2.1 | 50.2±5.5 | 83.5±3.2 | 47.4± 7.0 | 83.1 ±1.5 | 48.2 ±4.7 | 82.6 ± 1.7 | 48.0 ±5.5 | 80.9 ± 2.4 | 45.2 ± 6.9 |
| Model | Generalizing to a new patient group | |||||||
| Train: Young→ Test: Old | Train: Old→Test: Young | Train: Male → Test: Female | Train: Female → Test: Male | |||||
| AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | |
| Transformer | 76.2 ± 0.7 | 30.5±4.8 | 76.5 ± 1.1 | 33.7 ± 5.7 | 77.8 ± 1.1 | 26.0±6.2 | 75.2 ± 1.0 | 30.3 ± 5.5 |
| Trans-mean | 80.6 ±1.4 | 39.8 ± 4.2 | 78.4 ±1.1 | 35.8±2.9 | 80.2 ± 1.7 | 32.1 ± 1.9 | 76.4±0.8 | 32.5±3.3 |
| GRU-D | 76.5 ± 1.7 | 29.5 ± 2.3 | 79.6 ±1.7 | 35.2 ± 4.6 | 78.5 ± 1.6 | 31.9 ± 4.8 | 76.3 ±2.5 | 31.1 ± 2.6 |
| SeFT | 77.5 ± 0.7 | 26.6 ± 1.2 | 78.9 ±1.0 | 32.7 ±2.7 | 78.6 ±0.6 | 31.1 ± 1.2 | 76.9 ± 0.5 | 26.4 ± 1.1 |
| mTAND | 79.0 ±0.8 | 28.8± 2.3 | 79.4±0.6 | 29.8 ±1.2 | 78.0 ±0.9 | 26.5 ±1.7 | 78.9 ±1.2 | 29.2 ± 2.0 |
| RAINDROP W/o graph | 80.5 ± 1.1 | 31.6 ± 2.1 | 78.5±0.9 | 36.7 ±2.7 | 81.3 ±1.5 | 36.8±1.7 | 77.5 ± 1.9 | 33.4 ± 2.6 |
| RAINDROP | 83.2 ± 1.6 | 43.6 ± 4.7 | 82.0 ± 4.4 | 44.3 ± 3.6 | 85.0 ± 1.4 | 45.2 ± 2.9 | 81.2 ± 3.8 | 40.7 ± 2.9 |
| RAINDROP Model | Accuracy | Precision | Recall | F1 score | |
| W/o weights vector Ru | 81.1 ± 2.6 | 81.9 ± 2.4 | 80.1 ± 1.6 | 81.6± 2.1 | |
| W/o inter-sensor dependency | W/o ei,uv | 82.6± 1.2 | 82.9± 1.6 | 84.3± 1.4 | 83.8± 1.7 |
| W/oru W/op | 86.5 ± 2.4 | 83.3 ± 1.9 | 82.6± 1.5 | 82.9 ± 1.4 | |
| Wloau | 79.8 ± 2.7 85.2 ± 2.5 | 80.1 ± 3.6 86.4 ± 2.7 | 80.6 ± 1.7 84.5 ± 2.9 | 80.2 ± 2.9 85.6 ± 2.9 | |
| W/o temporal attention | 81.5 ± 1.9 | 84.6± 1.7 | 83.9 ± 2.5 | ||
| W/o sensor level concatenation | 84.4 ± 2.1 | 86.7± 1.1 | 85.2± 1.9 | 84.2 ± 2.2 | |
| W/o regularization term Lr | 87.3 ± 2.9 | 88.6± 3.4 | 87.1± 2.8 | 85.8± 2.6 | |
| Full RAINDROP | 88.5±1.5 | 89.9±1.5 | 89.9±0.6 | 87.6 ± 3.1 89.8±1.0 | |
| Model | Generalizing to anew patient group | |||||||
| Train: Young → Test: Old | Train: Old → Test: Young | Train: Male → Test: Female | Train: Female → Test: Male | |||||
| AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | AUROC | AUPRC | |
| Transformer | 76.2± 0.7 | 30.5± 4.8 | 76.5 ± 1.1 | 33.7±5.7 | 77.8 ± 1.1 | 26.0±6.2 | 75.2 ±1.0 | 30.3±5.5 |
| Trans-mean | 80.6 ± 1.4 | 39.8±4.2 | 78.4 ±1.1 | 35.8±2.9 | 80.2±1.7 | 32.1 ± 1.9 | 76.4 ±0.8 | 32.5±3.3 |
| GRU-D | 76.5 ± 1.7 | 29.5±2.3 | 79.6 ±1.7 | 35.2± 4.6 | 78.5±1.6 | 31.9 ± 4.8 | 76.3 ± 2.5 | 31.1 ± 2.6 |
| SeFT | 77.5 ± 0.7 | 26.6±1.2 | 78.9 ±1.0 | 32.7±2.7 | 78.6±0.6 | 31.1 ± 1.2 | 76.9 ± 0.5 | 26.4 ± 1.1 |
| mTAND | 79.0±0.8 | 28.8±2.3 | 79.4±0.6 | 29.8± 1.2 | 78.0±0.9 | 26.5±1.7 | 78.9 ± 1.2 | 29.2 ± 2.0 |
| RAINDROP | 83.2 ± 1.6 | 43.6± 4.7 | 82.0 ± 4.4 | 44.3 ± 3.6 | 85.0 ± 1.4 | 45.2 ± 2.9 | 81.2± 3.8 | 40.7± 2.9 |
| Methods | Task Units | α a | △ε meV | εHOMO meV | εLUMO meV | μ D | Cv cal/mol K | G meV | H meV | R a | U meV | U meV | ZPVE meV |
| NMP(Gilmer et al., 2017)t | .092 | 69 | 43 | 38 | .030 | .040 | 19 | 17 | .180 | 20 | 20 | 1.50 | |
| SchNet (Schutt et al.,2017) | .235 | 63 | 41 | 34 | .033 | .033 | 14 | 14 | .073 | 19 | 14 | 1.70 | |
| Cormorant (Anderson et al., 2019)† | .085 | 61 | 34 | 38 | .038 | .026 | 20 | 21 | .961 | 21 | 22 | 2.03 | |
| LieConv (Finzi et al.,020)t | .084 | 49 | 30 | 25 | .032 | .038 | 22 | 24 | .800 | 19 | 19 | 2.28 | |
| DimeNet++ (Gasteiger et al., 2020a) | .044 | 33 | 25 | 20 | .030 | .023 | 8 | 7 | .331 | 6 | 6 | 1.21 | |
| TFN (Thomas et al., 2018)† | .223 | 58 | 40 | 38 | .064 | .101 | - | - | - | - | - | - | |
| SE(3)-Transformer(Fuchs etal.,2020)† | .142 | 53 | 35 | 33 | .051 | .054 | - | - | - | - | - | - | |
| EGNN (Satorras et al.,2021)† | .071 | 48 | 29 | 25 | .029 | .031 | 12 | 12 | .106 | 12 | 11 | 1.55 | |
| PaiNN (Schutt et al.,2021) | .045 | 46 | 28 | 20 | .012 | .024 | 7.35 | 5.98 | .066 | 5.83 | 5.85 | 1.28 | |
| TorchMD-NET(Tholke & Fabritiis,2022) | .059 | 36 | 20 | 18 | .011 | .026 | 7.62 | 6.16 | .033 | 6.38 | 6.15 | 1.84 | |
| SphereNet (Liu et al.,2022) | .046 | 32 | 23 | 18 | .026 | .021 | 8 | 6 | .292 | 7 | 6 | 1.12 | |
| SEGNN (Brandstetter et al.,2022)† | .060 | 42 | 24 | 21 | .023 | .031 | 15 | 16 | .660 | 13 | 15 | 1.62 | |
| EQGAT (Le et al.,2022) | .053 | 32 | 20 | 16 | .011 | .024 | 23 | 24 | .382 | 25 | 25 | 2.00 | |
| Equiformer | .046 | 30 | 15 | 14 | .011 | .023 | 7.63 | 6.63 | .251 | 6.74 | 6.59 | 1.26 |
| Aspirin | Benzene | Ethanol | Malonaldehyde | Naphthalene | Salicylic acid | Toluene | Uracil | |||||||||
| Methods | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces |
| SchNet (Schutt et al.,2017) | 16.0 | 58.5 | 3.5 | 13.4 | 3.5 | 16.9 | 5.6 | 28.6 | 6.9 | 25.2 | 8.7 | 36.9 | 5.2 | 24.7 | 6.1 | 24.3 |
| DimeNet (Gasteiger et al.,2020b) | 8.8 | 21.6 | 3.4 | 8.1 | 2.8 | 10.0 | 4.5 | 16.6 | 5.3 | 9.3 | 5.8 | 16.2 | 4.4 | 9.4 | 5.0 | 13.1 |
| PaiNN (Schuitt et al., 2021) | 6.9 | 14.7 | - | - | 2.7 | 9.7 | 3.9 | 13.8 | 5.0 | 3.3 | 4.9 | 8.5 | 4.1 | 4.1 | 4.5 | 6.0 |
| TorchMD-NET(Tholke &Fabritis,2022) | 5.3 | 11.0 | 2.5 | 8.5 | 2.3 | 4.7 | 3.3 | 7.3 | 3.7 | 2.6 | 4.0 | 5.6 | 3.2 | 2.9 | 4.1 4.5 | 4.1 |
| NequIP(Lmax =3)(Batzner etal.,022) | 5.7 | 8.0 | - | - | 2.2 | 3.1 | 3.3 | 5.6 | 4.9 | 1.7 | 4.6 | 3.9 | 4.0 | 2.0 | 3.3 | |
| Equiformer (Lmar =2) | 5.3 | 7.2 | 6.6 | 2 | 3.3 | 5.8 | 3.7 | 4.5 | 4.1 | 3.8 | 4.3 | 3.3 | ||||
| Equiformer(Lmax =3) | 5.3 | 6.6 | 粥 | 8.1 | 3 | 3.2 | 5.4 | 4.4 | 20 | 4.3 | 3.9 | 3.7 | 2 | 4.3 | 3.4 | |
| Energy MAE (eV)↓ | EwT(%)↑ | |||||||||
| Methods | ID | OOD Ads | OOD Cat | OOD Both | Average | ID | OOD Ads | OOD Cat | OOD Both | Average |
| CGCNN (Xie & Grossman,2018) | 0.6149 | 0.9155 | 0.6219 | 0.8511 | 0.7509 | 3.40 | 1.93 | 3.10 | 2.00 | 2.61 |
| SchNet (Schut et al., 2017) | 0.6387 | 0.7342 | 0.6616 | 0.7037 | 0.6846 | 2.96 | 2.33 | 2.94 | 2.21 | 2.61 |
| DimeNet++(Gasteiger etal.,2020a) | 0.5621 | 0.7252 | 0.5756 | 0.6613 | 0.6311 | 4.25 | 2.07 | 4.10 | 2.41 | 3.21 |
| PaiNN (Schutt et al.,2021) | 0.575 | 0.783 | 0.604 | 0.743 | 0.6763 | 3.46 | 1.97 | 3.46 | 2.28 | 2.79 |
| SpinConv (Shuaibi et al.,2021) | 0.5583 | 0.7230 | 0.5687 | 0.6738 | 0.6310 | 4.08 | 2.26 | 3.82 | 2.33 | 3.12 |
| SphereNet (Liu etal.,2022) | 0.5625 | 0.7033 | 0.5708 | 0.6378 | 0.6186 | 4.47 | 2.29 | 4.09 | 2.41 | 3.32 |
| SEGNN (Brandstetter et al., 2022) | 0.5327 | 0.6921 | 0.5369 | 0.6790 | 0.6101 | 5.37 | 2.46 | 4.91 | 2.63 | 3.84 |
| Equiformer | 0.5037 | 0.6881 | 0.5213 | 0.6301 | 0.5858 | 5.14 | 2.41 | 4.67 | 2.69 | 3.73 |
| Energy MAE (eV)↓ | EwT(%) ↑ | |||||||||
| Methods | ID | OOD Ads | OOD Cat | OODBoth | Average | ID | OOD Ads | OOD Cat | OOD Both | Average |
| GNS (Godwin et al.,2022) | 0.54 | 0.65 | 0.55 | 0.59 | 0.5825 | - | - | - | - | - |
| GNS + Noisy Nodes (Godwin et al., 2022) | 0.47 | 0.51 | 0.48 | 0.46 | 0.4800 | = | - | - | - | - |
| Graphormer (Shi et al., 2022) | 0.4329 | 0.5850 | 0.4441 | 0.5299 | 0.4980 | - | - | - | - | |
| Equiformer | 0.4222 | 0.5420 | 0.4231 | 0.4754 | 0.4657 | 7.23 | 3.77 | 7.13 | 4.10 | 5.56 |
| Equiformer + Noisy Nodes | 0.4156 | 0.4976 | 0.4165 | 0.4344 | 0.4410 | 7.47 | 4.64 | 7.19 | 4.84 | 6.04 |
| Energy MAE (eV)↓ | EwT(%)↑ | Training time | |||||||||
| Methods | ID | OOD Ads | OOD Cat | OOD Both | Average | D | OOD Ads | OOD Cat | OOD Both | Average | (GPU-days) |
| GNS + Noisy Nodes (Godwin et al.,2022) | 0.4219 | 0.5678 | 0.4366 | 0.4651 | 0.4728 | 9.12 | 4.25 | 8.01 | 4.64 | 6.5 | 56 (TPU) |
| Graphormer (Shi et al.,22)† | 0.3976 | 0.5719 | 0.4166 | 0.5029 | 0.4722 | 8.97 | 3.45 | 8.18 | 3.79 | 6.1 | 372(A100) |
| Equiformer+Noisy Nodes | 0.4171 | 0.5479 | 0.4248 | 0.4741 | 0.4660 | 7.71 | 3.70 | 7.15 | 4.07 | 5.66 | 24 (A6000) |
| Index | Methods | Task α | △ε meV | εHOMO meV | εLUMO | 片 Cv | Training time | Number of parameters | ||||
| Non-linear message passing | MLP attention | Unit a | D cal/mol K | |||||||||
| 1 | √ | · | attention | .046 | 30 | meV 14 | .011 | .023 | (minutes/epoch) 12.1 | 3.53M | ||
| 2 | .051 | 32 | 15 16 | 16 | .013 | .025 | 7.2 | 3.01M | ||||
| 3 | √ | .053 | 32 | 17 | 16 | .013 | .025 | 7.8 | 3.35M | |||
| Index | Methods | Energy MAE(eV)↓ | EwT(%)↑ | Number of | |||||||||||
| Non-linear message passing | MLP attention | Dot product attention | ID | OOD Ads | OOD Cat | OOD Both | Average | ID | OOD Ads | OOD Cat | OOD Both | Average | Training time (minutes/epoch) | ||
| 123 | √ | 0.5088 | 0.6271 | 0.5051 | 0.5545 | 0.5489 | 4.88 | 2.93 | 4.92 | 2.98 | 3.93 | 130.8 | parameters 9.12M | ||
| · | 0.5168 | 0.6308 | 0.5088 | 0.5657 | 0.5555 | 4.59 | 2.82 | 4.79 | 3.02 | 3.81 | 91.2 | 7.84M | |||
| √ | 0.5386 | 0.6382 | 0.5297 | 0.5692 | 0.5689 | 4.37 | 2.60 | 4.36 | 2.86 | 3.55 | 99.3 | 8.72M | |||
| Hyper-parameters | Value or description |
| Optimizer | AdamW |
| Learning rate scheduling Warmup epochs | Cosine learning rate with linear warmup 5 |
| Maximum learning rate | 1.5 × 10-4,5 × 10-4 |
| Batch size | 64,128 |
| Number of epochs | 300,600 |
| Weight decay | 0,5×10-3 |
| Dropout rate | 0.0,0.1, 0.2 |
| Cutoff radius (A) | 5 |
| Number of radial bases | 128 for Gaussian radial basis,8 for radial bessel basis |
| Hidden sizes of radial functions | 64 |
| Number of hidden layers in radial functions | 2 |
| Number of Transformer blocks Embedding dimension dembed | Equiformer 6 [(128,0),(64,1),(32,2)] |
| Spherical harmonics embedding dimension dsh Numberof attention heads h Attention head dimension dhead | [(1,0),(1,1),(1,2)] 4 [(32,0),(16,1),(8,2)] |
| Hidden dimension in feed forward networks dffn Output feature dimension d feature | (384,0),(192,1),(96,2)] [(512,0)] |
| E(3)-Equiformer | |
| 6 | |
| Number of Transformer blocks | |
| Embedding dimension dembed | |
| [(128,0,e),(32,0,0),(32,1,e),(32,1,0),(16,2,e),(16,2,0)] | |
| Spherical harmonics embedding dimension dsh | [(1,0,e),(1,1,0),(1,2,e)] |
| Number of attention heads h | 4 |
| Attention head dimension dhead | [(32,0,e),(8,0,0),(8,1,e),(8,1,0),(4,2,e),(4,2,0)] |
| Hidden dimension in feed forward networks df fn | |
| (384,0,e),(96,0,0),(96,1,e),(96,1,0),(48,2,e),(48,2,0)] | |
| Output feature dimension dfeature | [(512,0,e)] |
| Methods | Task Units | α 品 | △ meV | εHOMO meV | εLUMO meV | μ D | Cv cal/mol K | Training time (minutes/epoch) | Number of parameters |
| Equiformer | .046 | 30 | 15 | 14 | .011 | .023 | 12.1 | 3.53M | |
| E(3)-Equiformer | .045 | 30 | 15 | 14 | .012 | .023 | 16.3 | 3.28M |
| Methods | Number of parameters | Training time (GPU-hours) |
| SEGNN (Brandstetter et al., 2022) | 1.03M | 81 |
| TorchMD-NET(Tholke&Fabritis,2022) | 6.86M | 92 |
| Equiformer | 3.53M | 61 |
| Hyper-parameters | Value or description |
| Optimizer | AdamW |
| Learning rate scheduling | Cosine learning rate with linear warmup |
| Warmup epochs | 10 1 × 10-4,2 × 10-4,5× 10-4 |
| Maximum learning rate Batch size | 5,8 |
| Number of epochs | 1500,2000 |
| 1×10-6 | |
| Weight decay Dropout rate | 0.0 |
| Weight for energy loss Weight for force loss | 1 80 |
| Cutoff radius (A) | 5 |
| Number of radial bases | 32 |
| Hidden sizes of radial functions | 64 2 |
| Number of hidden layers in radial functions Equiformer (Lmax = 2) | |
| Embedding dimension dembed Spherical harmonics embedding dimension d sh Number of attention heads h 4 Attention head dimension dhead Hidden dimension in feed forward networks d f fn | [(128,0),(64,1),(32,2)] [(1,0),(1,1),(1,2)] [(32,0),(16,1),(8,2)] [(384,0),(192,1),(96,2)] |
| Output feature dimension dfeature Equiformer (Lmax = 3) | |
| [(512,0)] | |
| Number of Transformer blocks | 6 |
| Embedding dimension dembed | [(128,0),(64,1),(64,2),(32,3)] |
| Spherical harmonics embedding dimension d sh | [(1,0),(1,1),(1,2),(1,3)] |
| Number of attention heads h | 4 |
| Attention head dimension dhead | [(32,0),(16,1),(16,2),(8,3)] |
| Hidden dimension in feed forward networks dffn | (384,0),(192,1),(192,2),(96,3)] |
| Aspirin | Benzene | Ethanol | Malonaldehyde | Naphthalene | Salicylic acid | Toluene | Uracil | |||||||||
| Methods | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces | energy | forces |
| TorchMD-NET | 5.3 | 11.0 | 2.5 | 8.5 | 2.3 | 4.7 | 3.3 | 7.3 | 3.7 | 2.6 | 4.0 | 5.6 | 3.2 | 2.9 | 4.1 | 4.1 |
| Equiformer (Lmax = 2) | 5.3 | 7.2 | 2.2 | 6.6 | 2.2 | 3.1 | 3.3 | 5.8 | 3.7 | 2.1 | 4.0 | 5.3 | 3.2 | 2.4 | 4.2 | 3.7 |
| Methods | Number of parameters | Training time (secs/epoch) |
| NequIP(Lmax =3)(Batzner et al.,2022) | 2.97M | 48.7 |
| Equiformer (Lmax =2) | 3.50M | 36.9 |
| Equiformer (Lmax =3) | 5.50M | 98.0 |
| Hyper-parameters | Value or description |
| Optimizer | AdamW |
| Learning rate scheduling | Cosine learning rate with linear warmup |
| Warmup epochs | 2 |
| Maximum learning rate | 2×10-4 |
| Batch size | 32 |
| Number of epochs | 20 |
| Weight decay | 1×10-3 |
| Dropout rate | 0.2 |
| Cutoff radius (A) | 5 |
| Number of radial basis | 128 |
| Hidden size of radial function | 64 |
| Numberofhiddenlayers in radial function | 2 |
| Equiformer | |
| Numberof Transformer blocks Embedding dimension dembed | 6 |
| Spherical harmonics embedding dimension dsh | [(256,0),(128,1)] [(1,0),(1,1)] |
| Number of attention heads h | 8 |
| Attention head dimension dhead | [(32,0),(16,1)] |
| Hidden dimension in feed forward networks d f fn | (768,0),(384,1)] |
| Output feature dimension d feature | [(512,0)] |
| E(3)-Equiformer | |
| Number of Transformer blocks Embedding dimension dembed | 6 [(256,0,e),(64,0,0),(64,1,e),(64,1,0)] |
| Spherical harmonics embedding dimension dsh | [(1,0,e),(1,1,0)] |
| Number of attention heads h | 8 |
| Attention head dimension dhead | [(32,0,e),(8,0,0),(8,1,e),(8,1,0)] |
| Hidden dimension in feed forward networks d ffn | [(768,0,e),(192,0,0),(192,1,e),(192,1,0)] |
| Output feature dimension d feature | [(512,0,e)] |
| Energy MAE(eV)↓ | EwT(%)↑ | |||||||||
| Methods | ID | OOD Ads | OOD Cat | OOD Both | Average | ID | OOD Ads | OOD Cat | OOD Both | Average |
| SchNet (Schut et al.,07)+ | 0.6465 | 0.7074 | 0.6475 | 0.6626 | 0.6660 | 2.96 | 2.22 | 3.03 | 2.38 | 2.65 |
| DimeNet++ (Gasteiger et al.,2020a)† | 0.5636 | 0.7127 | 0.5612 | 0.6492 | 0.6217 | 4.25 | 2.48 | 4.40 | 2.56 | 3.42 |
| GemNet-T(Klicpera et al., 2021)† | 0.5561 | 0.7342 | 0.5659 | 0.6964 | 0.6382 | 4.51 | 2.24 | 4.37 | 2.38 | 3.38 |
| SphereNet (Liu etal.,2022) | 0.5632 | 0.6682 | 0.5590 | 0.6190 | 0.6024 | 4.56 | 2.70 | 4.59 | 2.70 | 3.64 |
| (S)EGNN (Brandstetter et al.,2022) | 0.5497 | 0.6851 | 0.5519 | 0.6102 | 0.5992 | 4.99 | 2.50 | 4.71 | 2.88 | 3.77 |
| SEGNN (Brandstetter et al., 2022) | 0.5310 | 0.6432 | 0.5341 | 0.5777 | 0.5715 | 5.32 | 2.80 | 4.89 | 3.09 | 4.03 |
| Equiformer | 0.5088 | 0.6271 | 0.5051 | 0.5545 | 0.5489 | 4.88 | 2.93 | 4.92 | 2.98 | 3.93 |
| Energy MAE (eV)↓ | EwT(%)↑ | Training time (minutes/epoch) | Number of parameters | |||||||||
| Methods | ID | OOD Ads | OOD Cat | OOD Both | Average | ID | OOD Ads | OOD Cat | OOD Both | Average | ||
| Equiformer | 0.5088 | 0.6271 | 0.5051 | 0.5545 | 0.5489 | 4.88 | 2.93 | 4.92 | 2.98 | 3.93 | 130.8 | 9.12M |
| E(3)-Equiformer | 0.5035 | 0.6385 | 0.5034 | 0.5658 | 0.5528 | 5.10 | 2.98 | 5.10 | 3.02 | 4.05 | 174.9 | 8.77M |
| Methods | Number of parameters | Training time (GPU-hours) |
| SEGNN (Brandstetter et al., 2022) | 4.21M | 79 |
| Equiformer | 9.12M | 87 |
| Sparse between positions | Sparse between channels | Weight sharing across | Dynamic weight | |||
| non-local | full | position | channel | |||
| Local MLP | √ | √ | ” | |||
| Local attention | √ | : | √ | |||
| DW-Conv. | √ | : | ||||
| D-DW-Conv. | √ | √ | √ | |||
| Conv. | √ | √ | ||||
| method | img. size #param. FLOPs throughput (img./s) top-1 acc.real acc. | |||||
| Bottleneck:convolution with low rank | ||||||
| ResNet-50 (He et al.,2016) | 2242 | 26M | 4.1G | 1128.3 | 76.2 | 82.5 |
| ResNet-101 (He et al., 2016) | 224² | 45M | 7.9G | 652.0 | 77.4 | 83.7 |
| ResNet-152 (He et al., 2016) | 2242 | 60M | 11.6G | 456.7 | 78.3 | 84.1 |
| Channel and spatial separable MLP, spatial separable MLP=point-wise1 ×1 convolution | ||||||
| Mixer-B/16 (Tolstikhin et al.,2021) | 2242 | 46M | - | - | 76.4 | 82.4 |
| Mixer-L/16(Tolstikhin et al.,2021) | 224² | 189M | - | 71.8 | 77.1 | |
| ResMLP-12 (Touvron et al., 2021) | 224² | 15M | 3.0G | = | 76.6 | 83.3 |
| ResMLP-24 (Touvron et al., 2021) | 224² | 30M | 6.0G | 79.4 | 85.3 | |
| ResMLP-36 (Touvron et al.,2021) | 2242 | 45M | 8.9G | - | 79.7 | 85.6 |
| Global attention:dynamicchannel separable MLP+ spatial separable MLP | ||||||
| ViT-B/16 (Dosovitskiy et al.,2021) | 3842 | 86M | 55.4G | 83.4 | 77.9 | 83.6 |
| ViT-L/16 (Dosovitskiy et al.,2021) | 3842 | 307M | 190.7G | 26.5 | 76.5 | 82.2 |
| DeiT-S (Touvron et al., 2020) | 224² | 22M | 4.6G | 947.3 | 79.8 | 85.7 |
| DeiT-B (Touvron et al.,2020) | 224² | 86M | 17.5G | 298.2 | 81.8 | 86.7 |
| DeiT-B (Touvron et al., 2020) | 3842 | 86M | 55.4G | 82.7 | 83.1 | 87.7 |
| Local MLP: perform static separable MLP in local small windows | ||||||
| Swin-Local MLP-T | 2242 | 26M | 3.8G | 861.0 | 80.3 | 86.1 |
| Swin-Local MLP-B | 224² | 79M | 12.9G | 321.2 | 82.2 | 86.9 |
| Local attention:perform attention inlocal small windows | ||||||
| Swin-T (Liu et al.,2021b) | 224² | 28M | 4.5G | 713.5 | 81.3 | 86.6 |
| Swin-B (Liu et al., 2021b) | 2242 | 88M | 15.4G | 263.0 | 83.3 | 87.9 |
| Depth-wise convolution + point-wise 1 ×1 convolution | ||||||
| DW-Conv.-T | 2242 | 24M | 3.8G | 928.7 | 81.3 | 86.8 |
| DW-Conv.-B | 2242 | 74M | 12.9G | 327.6 | 83.2 | 87.9 |
| D-DW-Conv.-T | 224² | 51M | 3.8G | 897.0 | 81.9 | 87.3 |
| D-DW-Conv.-B | 224² | 162M | 13.0G | 322.4 | 83.2 | 87.9 |
| I-D-DW-Conv.-T | 2242 | 26M | 4.4G | 685.3 | 81.8 | 87.1 |
| I-D-DW-Conv.-B | 2242 | 80M | 14.3G | 244.9 | 83.4 | 88.0 |
| COCO Object Detection | ADE20K Semantic Segmentation | |||||||
| #param. | FLOPs | Apbot | AP | APP | Apmask | #param. FLOPs | mIoU | |
| Swin-T | 86M | 747G | 50.5 | 69.3 | 54.9 | 43.7 | 60M 947G | 44.5 |
| DW Conv.-T | 82M | 730G 49.9 | 68.6 | 54.3 | 43.4 | 56M | 928G | 45.5 |
| D-DW Conv.-T | 108M | 730G 50.5 | 69.5 | 54.6 | 43.7 | 83M 928G | 45.7 | |
| I-D-DW Conv.-T | 84M | 741G | 50.8 | 69.5 | 55.3 | 44.0 | 58M 939G | 46.2 |
| Swin-B | 145M | 986G | 51.9 | 70.9 | 56.5 | 45.0 | 121M 1192G | 48.1 |
| DW Conv.-B | 132M | 924G | 51.1 | 69.6 | 55.4 | 44.2 | 108M 1129G | 48.3 |
| D-DW Conv.-B | 219M | 924G 51.2 | 70.0 | 55.4 | 44.4 | 195M 1129G | 48.0 | |
| I-D-DW Conv.-B | 137M | 948G | 51.8 | 70.3 | 56.1 | 44.8 | 114M 1153G | 47.8 |
| SC | SP | Acc. | #param. | SC | SP | Acc. | #param. | ||
| Local MLP | X | > | 80.2 | 35.3M | DW Conv. | X √ | √ | 81.3 | 24.2M |
| √ | X | 80.3 | 26.2M | × | 80.3 | 26.2M | |||
| √ | √ | 80.3 | 24.3M | √ | √ | 81.1 | 23.9M |
| Win. samp. | #param. | FLOPs | Acc. | Win. samp. | #param. | FLOPs | Acc. | ||
| Local MLP | shifted | 26M | 3.8G | 80.3 | DW Conv. | sliding | 24M | 3.8G | 81.3 |
| w/ attention | shifted | 28M | 4.5G | 81.3 | w/ homo. dyna. | sliding | 51M | 3.8G | 81.9 |
| w/ attention | sliding | 28M | 4.5G | 81.4 | w/ inhomo. dyna. | sliding | 26M | 4.4G | 81.8 |
| #param. | FLOPs | top-1 acc. | |
| Twins-PCPVT (Chu et al.,2021a) | 24M | 3.8G | 81.2 |
| Twins-SVT (Chu et al.,2021a) | 24M | 2.9G | 81.7 |
| CoaT-Lite (Xu et al.,2021) | 20M | 4.0G | 81.9 |
| CoaT (Xu et al., 2021) | 22M | 12.6G | 82.1 |
| PVT-v2 (Wang et al., 2021a) | 25M | 4.0G | 82.0 |
| Shuffle Transformer (Huang et al.,2021) | 29M | 4.6G | 82.5 |
| I-D-DW Conv. | 26M | 4.4G | 81.8 |
| I-D-DW Conv. + DW | 27M | 4.4G | 82.3 |
| Sparse between positions | Sparse between channels | Weight sharing across | Dynamic weight | |||
| non-local | full | position | channel | |||
| Local MLP | √ | √ | 公 | |||
| Local attention | √ | |||||
| DW-Conv. | <</ | v√√ | ||||
| D-DW-Conv. | <√√ | √ | ||||
| Conv. | √ | |||||
| MLP Attention | √ | |||||
| Spatial-mixing MLP† | 1 | |||||
| 1 ×1 Conv.‡ | : | |||||
| downsp.rate(output size) | Swin | DW Conv. | |||||
| stage 1 | 4×(56×56) | concat 4×4, linear 96-d, LN | concat 4×4, linear 96-d, LN | ||||
| LN, linear 96x3-dlocal sa. 7×7,head 3linear 96-dLN, linear 384-dGELU, linear 96-d | ×2 | linear 96-d, BN, ReLUdepthwise conv.7×7,BN,ReLUlinear 96-d, BNBN, linear 384-dGELU, linear 96-d | ×2 | ||||
| stage 2 | 8×(28×28) | concat 2×2, linear 192-d,LN | concat 2×2, linear 192-d,LN | ||||
| LN, linear 192x3-dlocal sa. 7×7, head 6linear 192-dLN, linear 768-dGELU, linear 192-d | ×2 | linear 192-d, BN,ReLUdepthwise conv. 7×7, BN,ReLUlinear 192-d, BNBN, linear 768-dGELU, linear 192-d | ×2 | ||||
| stage 3 | 16×(14×14) | concat 2×2, linear 384-d,LN | concat 2×2,linear 384-d,LN | ||||
| LN, linear 384x3-dlocal sa. 7×7, head 12linear 384-dLN,linear 1536-dGELU, linear 384-d | ×6 | linear 384-d, BN,ReLUdepthwise conv. 7×7, BN, ReLUlinear 384-d, BNBN, linear 1536-dGELU,linear 384-d | ×6 | ||||
| stage 4 | 32×(7×7) | concat 2×2, linear 768-d,LN | concat 2×2, linear 768-d,LN | ||||
| LN, linear 768x3-dlocal sa. 7×7, head 24linear 768-dLN, linear 3072-dGELU,linear 768-d | ×2 | linear 768-d,BN,ReLUdepthwise conv.7×7,BN, ReLUlinear 768-d,BNBN, linear 3072-dGELU, linear 768-d | ×2 | ||||
| stage 4 | 1×1 | LN, AvgPool. 1 ×1linear classifier | LN, AvgPool. 1 ×1linear classifier | ||||
| method | img.size #param. | FLOPs throughput (img./s)top-1 acc. real acc. | ||||
| Convolution: local connection | ||||||
| ResNet-38 † (Wang et al.,2020) | 224² | 28M | 3.8G | 2123.7 | 75.4 | - |
| ResNet-72 † (Wang et al.,2020) | 2242 | 48M | 7.5G | 623.0 | 76.7 | - |
| ResNet-106 † (Wang et al., 2020) | 2242 | 65M | 11.1G | 452.8 | 77.3 | - |
| Bottleneck: convolution with low rank | ||||||
| ResNet-50 (He et al.,2016) | 2242 | 26M | 4.1G | 1128.3 | 76.2 | 82.5 |
| ResNet-101 (He et al.,2016) | 224² | 45M | 7.9G | 652.0 | 77.4 | 83.7 |
| ResNet-152 (He et al.,2016) | 2242 | 60M | 11.6G | 456.7 | 78.3 | 84.1 |
| Pyramid: convolution with pyramid (spatial low rank) features. | ||||||
| HRNet-W18 (Wang et al., 2020) | 2242 | 21M | 4.0G | - | 76.8 | - |
| HRNet-W32 (Wang et al.,2020) | 2242 | 41M | 8.3G | 78.5 | - | |
| HRNet-W48 (Wang et al., 2020) | 2242 | 78M | 16.1G | 1 | 79.3 | 1 |
| Channel and spatial separable MLP, spatial separable MLP = point-wise 1 ×1 convolution | ||||||
| Mixer-B/16 (Tolstikhin et al.,2021) | 224 | 46M | 1 | - | 76.4 | 82.4 |
| Mixer-L/16 (Tolstikhin et al.,2021) | 224² | 189M | 1 | 71.8 | 77.1 | |
| ResMLP-12(Touvron et al., 2021) | 224² | 15M | 3.0G | 76.6 | 83.3 | |
| ResMLP-24 (Touvron et al., 2021) | 2242 | 30M | 6.0G | 79.4 | 85.3 | |
| ResMLP-36 (Touvron et al.,2021) | 224² | 45M | 8.9G | 79.7 | 85.6 | |
| gMLP-Ti (Liu et al., 2021a) | 2242 224² | 6M | 1.4G | 72.0 | 1 | |
| gMLP-S (Liu et al.,2021a) | 2242 | 20M | 4.5G | 79.4 | 1 | |
| gMLP-B (Liu et al., 2021a) | 73M | 15.8G | = | 81.6 | - | |
| Global attention: dynamic channel separable MLP + spatial separable MLP | ||||||
| ViT-B/16 (Dosovitskiy et al., 2021) | 3842 | 86M | 55.4G | 83.4 | 77.9 | 83.6 |
| ViT-L/16 (Dosovitskiy et al., 2021) | 3842 | 307M | 190.7G | 26.5 | 76.5 | 82.2 |
| DeiT-S (Touvron et al.,2020) | 224² | 22M | 4.6G | 947.3 | 79.8 | 85.7 |
| DeiT-B (Touvron et al., 2020) | 224² | 86M | 17.5G | 298.2 | 81.8 | 86.7 |
| DeiT-B (Touvron et al., 2020) | 3842 | 86M | 55.4G | 82.7 | 83.1 | 87.7 |
| Pyramid attention: perform attention with spatial low rank | ||||||
| PVT-S (Wang et al., 2021b) | 224 | 25M | 3.8G | - | 79.8 | - |
| PVT-M (Wang et al., 2021b) | 224² | 44M | 6.7G | = | 81.2 | - |
| PVT-L (Wang et al., 2021b) | 2242 | 61M | 9.8G | = | 81.7 | 1 |
| Local MLP: perform static separable MLP in local small windows | ||||||
| Swin-Local MLP-T | 2242 | 26M | 3.8G | 861.0 | 80.3 | 86.1 |
| Swin-Local MLP-B | 2242 | 79M | 12.9G | 321.2 | 82.2 | 86.9 |
| Local attention:perform attention in local small windows | ||||||
| Swin-T (Liu et al., 2021b) | 224 | 28M | 4.5G | 713.5 | 81.3 | 86.6 |
| Swin-B (Liu et al.,2021b) | 2242 | 88M | 15.4G | 263.0 | 83.3 | 87.9 |
| Depth-wise convolution + point-wise 1 ×1 convolution | ||||||
| DW-Conv.-T | 2242 | 24M | 3.8G | 928.7 | 81.3 | 86.8 |
| DW-Conv.-B | 2242 | 74M | 12.9G | 327.6 | 83.2 | 87.9 |
| D-DW-Conv.-T | 224² | 51M | 3.8G | 897.0 | 81.9 | 87.3 |
| D-DW-Conv.-B | 224² | 162M | 13.0G | 322.4 | 83.2 | 87.9 |
| I-D-DW-Conv.-T | 2242 | 26M | 4.4G | 685.3 | 81.8 | 87.1 |
| I-D-DW-Conv.-B | 2242 | 80M | 14.3G | 244.9 | 83.4 | 88.0 |
| ImageNet-1K fine-tuning | |||
| #param. | FLOPs | top-1 acc. | |
| Swin-B | 88M | 15.4G | 85.2 |
| DW-Conv.-B | 74M | 12.9G | 84.8 |
| D-DW-Conv.-B | 162M | 13.0G | 85.0 |
| I-D-DW-Conv.-B | 80M | 14.3G | 85.2 |
| COCO fine-tuning | ADE20K fine-tuning | ||||||||
| #param. | FLOPs | APbox | AP | AP | APmask | #param. | FLOPs | mIoU | |
| Swin-B | 145M | 986G | 53.4 | 72.1 | 58.1 | 46.1 | 121M | 1192G | 49.4 |
| DW Conv.-B | 132M | 924G | 52.0 | 70.4 | 56.3 | 45.0 | 108M | 1129G | 50.1 |
| D-DW Conv.-B | 219M | 924G | 51.9 | 70.7 | 56.2 | 45.0 | 195M | 1129G | 49.6 |
| I-D-DW Conv.-B | 137M | 948G | 52.9 | 71.2 | 57.2 | 45.8 | 114M | 1153G | 51.3 |
| Layer Norm. | Batch Norm. | Centering calibrated Batch Norm. | Top-1 Acc. | |
| Swin | < | 【 | 81.3 | |
| Swin | 80.9 | |||
| Swin | √ | 81.2 | ||
| DW Conv. | 1 | 81.2 | ||
| DW Conv. | √ | 81.3 | ||
| √ | ||||
| DW Conv. | 81.7 | |||
| #param. | FLOPs | top-1 acc. | |
| VOLO-d1 (Yuan et al.,2021c) | 27M | 7.0G | 84.1 |
| VOLO (Local SA)-d1 | 27M | 7.2G | 84.2 |
| DW Conv.-d1 | 26M | 6.9G | 84.2 |
| SVT-S (Chu et al., 2021a) | 24M | 2.8G | 81.7 |
| DW Conv.-S | 22M | 2.7G | 81.9 |
| model | ws. | #param. | FLOPs | Acc. |
| Swin | 7×7 12×12 | 28M 28M | 14.4G 14.2G | 81.8 82.4 |
| DW Conv. | 7×7 12×12 | 24M 25M | 11.1G 11.5G | 82.2 82.1 |
| model | SE | #param. | FLOPs | Acc. |
| Swin | √ | 28M | 4.5G | 81.3 |
| 29M | 4.5G | 81.2 | ||
| DW Conv. | √ | 24M | 3.8G | 81.3 |
| 24M | 3.8G | 81.7 |
| Model | CDR-H1 | CDR-H2 | CDR-H3 | Model | AAR | |||
| PPL | RMSD | PPL | RMSD | PPL | RMSD | RAbD | 28.53% | |
| LSTM | 6.79 | - | 7.21 | 1 | 9.70 | - | LSTM | 22.53% |
| AR-GNN | 6.44 | 2.97 | 6.86 | 2.27 | 9.44 | 3.63 | AR-GNN | 23.86% |
| RefineGNN | 6.09 | 1.18 | 6.58 | 0.87 | 8.54 | 2.50 | RefineGNN | 34.14% |
| Original | LSTM | AR-GNN | RefineGNN | |
| CoVAbDab PPL(↓) | 9.40 | 8.67 | 7.86 | |
| Neutralization (↑) | 69.3% | 72.0% | 70.4% | 75.2% |
| Antibody: | S1D7 | C694 |
| Old CDR-H3 | TRGHSDY | ARDRGYDSSGPDAFDI |
| New CDR-H3 | ARWWMDV | ARERIIIVSISAWMDV |
| Improvement | 63% → 73% | 82% →91% |
| Method | NFEs=5 | NFEs=10 | NFEs=15 | NFEs=20 | NFEs=25 |
| GENIE(ours) (*) | 11.2 | 5.28 | 4.49 | 3.94 | 3.64 |
| GENIE (ours) | 13.9 | 5.97 | 4.49 | 3.94 | 3.67 |
| DDIM[58] (*) | 27.6 | 11.2 | 7.35 | 5.87 | 5.16 |
| DDIM[58] | 29.7 | 11.2 | 7.35 | 5.87 | 5.16 |
| S-PNDM[63] | 35.9 | 10.3 | 6.61 | 5.20 | 4.51 |
| F-PNDM[63] | N/A | N/A | 10.3 | 5.96 | 4.73 |
| Euler-Maruyama | 325 | 230 | 164 | 112 | 80.3 |
| FastDDIM [64] (t) | 1 | 9.90 | - | 5.05 | - |
| Learned Sampler 66] (t/*) | 12.4 | 7.86 | 5.90 | 4.72 | 4.25 |
| Learned Sampler [66] (t) | 14.3 | 8.15 | 5.94 | 4.89 | 4.47 |
| Analytic DDIM [65] (t) | - | 14.0 | - | - | 5.71 |
| CLD-SGM[60] | 334 | 306 | 236 | 162 | 106 |
| VESDE-PC[57] | 461 | 461 | 461 | 461 | 462 |
| Ablation | NFEs=5 | NFEs=10 | NFEs=15 | NFEs=20 | NFEs=25 |
| Standard | 13.9 | 6.04 | 4.49 | 3.94 | 3.67 |
| No mixed | 14.7 | 6.32 | 4.82 | 4.31 | 4.10 |
| No weighting | 14.8 | 7.45 | 5.89 | 5.17 | 4.80 |
| Bigger model | 13.7 | 5.58 | 4.46 | 4.05 | 3.77 |
| Method | NFEs=5 | NFEs=10 | NFEs=15 |
| GENIE (ours) | 5.53 | 4.90 | 4.83 |
| DDIM [58] | 9.47 | 6.64 | 5.85 |
| S-PNDM[63] | 14.6 | 11.0 | 8.83 |
| F-PNDM [[63] | N/A | N/A | 11.7 |
| p | HCSE | HLP | LOU | |
| P2 P1 P0 | 4.5E(-2) 1.5E(-3) 6E(-6) | 0.89 0.93 0.62 | 0.79 0.75 0.58 | 0.92 0.92 11 |
| P2 P1 P0 | 5.5E(-2) 1.5E(-3) 4E(-6) | 0.87 0.95 0.72 | 0.89 0.87 11 | 0.89 0.87 11 |
| P2 P1 P0 | 6.5E(-2) 4.5E(-3) 2.5E(-6) | 0.96 0.94 0.80 | 0.95 0.81 11 | 0.99 0.99 |
| Networks | HCSE | HLP | LOUVAIN |
| CSphd | 1.30E4 / 5.19E4 / 5 | 1.54E4 / 5.58E4 /4 | 1.28E4 / 7.61E4 / 5 |
| fb-pages-government | 2.48E6 /1.18E8 / 4 | 2.53E6 / 1.76E8 / 3 | 2.43E6 / 1.33E8 / 4 |
| email-univ | 1.16E5 /2.20E6 / 3 | 1.46E5 / 6.14E6 / 3 | 1.14E5 / 2.20E6 /4 |
| fb-messages | 1.58E5 / 4.50E6 / 4 | 1.76E5 /8.12E6 / 3 | 1.52E5 / 4.96E6 / 4 |
| G22 | 5.56E5 / 2.68E7 /4 | 6.11E5 /4.00E7 /3 | 5.63E5 /2.80E7 /5 |
| As20000102 | 2.64E5 / 2.36E7 /4 | 3.62E5 / 7.63E7 /3 | 2.42E5 /2.42E7 / 5 |
| bibd-13-6 | 7.41E5 / 2.56E7 / 3 | 8.05E5 /4.41E7 /2 | 7.50E5 / 2.75E7 /4 |
| delaunay-n10 | 4.65E4 / 3.39E5 / 4 | 4.87E4 / 3.55E5 /4 | 4.24E4 / 4.25E5 / 5 |
| p2p-Gnutella05 | 9.00E5 /1.48E8 / 3 | 1.01E6 / 2.78E8 / 3 | 8.05E5 / 1.49E8 / 5 |
| p2p-Gnutella08 | 5.59E5 / 5.51E7 / 4 | 6.36E5 / 1.28E8 /4 | 4.88E5 / 6.03E7 / 5 |
| index | HCSE | HLP | LOUVAIN |
| J | 0.20 | 0.16 | 0.17 |
| cost(SE) | 1.85E6 | 2.05E6 | 1.89E6 |
| cost(Das) | 5.57E8 | 3.99E8 | 3.08E8 |
| Algorithms | Stochastic gradientcomplexity | Full gradient computation |
| GD (Nesterov,2004) | (m) | Computed for every iteration |
| SVRG (Reddi et al.,2016;Allen-Zhu & Hazan,2016),SCSG (Lei et al.,2017),SVRG+ (Li&Li,2018) | o(n+n2/3L△)2 | Computed for the initial pointand periodically computedfor every l iterations |
| SNVRG (Zhou et al., 2018),Geom-SARAH (Horvath et al., 2020) | (n+△)2 | Computed for the initial pointand periodically computedfor every l iterations |
| SPIDER (Fang et al., 2018),SpiderBoost (Wang et al., 2018),SARAH (Pham et al., 2019),SSRGD (Li,2019),PAGE (Li et al., 2021) | 0(n+) | Computed for the initial pointand periodically computedfor every l iterations |
| ZeroSARAH(this paper, (,Corollary 1) | O(n+Viy△)) | Only computed once forthe initial point 1 |
| ZeroSARAH(this paper, Corollary 2) | (√n(L△o+Go))E2 | Never computed ² |
| Algorithms | Stochastic gradientcomplexity | Full gradient computation |
| DC-GD(Khaled & Richtarik,202O;Li & Richtarik,2020) | 0(m) | Computed for every iteration |
| D-SARAH ²(Cen et al., 2020) | 0(m + YmlogmL△a) | Computed for the initial pointand periodically computedacross all n clients |
| D-GET²(Sun et al., 2020) | 0(m+ ma) | Computed for the initial pointand periodically computedacross all n clients |
| SCAFFOLD 3(Karimireddy et al., 2020) | 0(m+m会n1/3 | Only computed once forthe initial point |
| DC-LSVRG/DC-SAGA 1(Li&Richtarik,2020) | 0(m+m2/3Ln1/3 | Computed for the initial pointand periodically computedacross all n clients |
| FedPAGE 3(Zhao et al., 2021) | 0(m+外会 | Computed for the initial pointand periodically computedacross all n clients |
| (Distributed) SARAH/SPIDER/SSRGD 4(Nguyen et al.,2017; Fang et al.,2018; Li, 2019) | 0(m+) | Computed for the initial pointand periodically computedacross all n clients |
| D-ZeroSARAH(this paper, Corollary 3) | 0(m+√π)n | Only computed once forthe initial point |
| D-ZeroSARAH(this paper, Corollary 4) | 0(√L△+))² | Never computed |
| Dataset | n (# of datapoints) | d (# of features) |
| a9a | 32561 | 123 |
| abalone | 4177 | 8 |
| mg | 1385 | 6 |
| mushrooms | 8124 | 112 |
| phishing | 11055 | 68 |
| pyrim | 74 | 27 |
| triazines | 186 | 60 |
| w8a | 49749 | 300 |
| Method | ChemProt | ACL-ARC | SciERC |
| Unmerged | 82.70.3 | 70.53.2 | 81.00.4 |
| Fisher | 83.10.4 | 73.21.7 | 81.30.5 |
| Isotropic | 82.80.4 | 72.52.3 | 81.70.5 |
| Fine-tuned | 82.50.1 | 71.53.0 | 81.61.0 |
| Abbreviation | Explanation |
| D(T)I-MI-FGSM | the combination of D(T)IM and MI-FGSM |
| SI-NI-FGSM | the combination of SIMand NI-FGSM |
| D(T,S)I-EPNI-FGSM | the combination of D(T,S)IM and EPNI-FGSM |
| VT-M(N)I-FGSM | i.e., VM(N)I-FGSM |
| VT-EPNI-FGSM | the combination of Variance Tuning (VT)(Wang & He,2021) and EPNI-FGSM |
| FI-MI-FGSM FI-EPNI-FGSM | i.e., FIA |
| the combination of Feature Importance-aware (FI) (Wang et al.,2021) and EPNI-FGSM |
| Iv3 | 69.2 56.7 | 67.9 | 68.2 | 66.2 | 63.0 | 61.4 | 61.2 | ||||||||||||
| TI-MI-FGSM SI-NI-FGSM DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) | 41.8 77.4 88.7 70.6 90.2 | 99.7* 100.0* 100.0* 100.0* 100.0* | 37.9 72.7 87.3 68.8 87.1 | 31.5 69.8 86.0 59.6 | 46.4 67.6 78.5 69.3 | 47.6 67.3 81.3 69.8 | 44.9 68.0 83.4 68.9 | 41.9 61.5 80.2 65.5 76.4 | 38.6 62.1 80.5 63.3 | 54.9 70.0 80.4 73.5 | 72.8 83.8 76.5 | 56.2 73.8 83.8 78.5 | 58.3 71.9 84.2 77.9 | 47.9 70.7 86.1 74.7 | 46.5 68.4 84.9 72.9 | 43.5 64.8 82.9 70.8 | 42.8 65.9 83.4 71.5 | 66.0 49.2 70.9 84.4 72.5 | |
| SI-EPNI-FGSM(Ours) VT-MI-FGSM VT-NI-FGSM | 70.5 76.2 | 100.0* 100.0* | 69.7 76.2 | 85.1 64.9 70.5 | 80.1 61.7 65.6 | 81.2 63.6 68.4 | 81.7 65.7 71.2 | 60.0 65.2 | 78.2 58.6 64.1 | 80.6 65.0 70.5 | 83.1 69.2 73.4 | 84.4 68.7 73.4 | 82.6 66.7 73.2 | 86.7 65.8 71.2 | 83.8 63.8 71.3 | 80.9 65.0 69.2 | 81.7 62.6 68.2 | 83.8 67.1 72.2 | |
| VT-EPNI-FGSM(Ours) FI-MI-FGSM FI-EPNI-FGSM(Ours) | 84.8 85.8 90.0 | 100.0* 97.1* 97.4* | 83.3 85.4 88.0 | 78.1 81.8 84.8 | 77.0 79.4 81.0 | 78.8 80.1 82.4 | 79.4 79.5 84.3 | 75.8 76.3 79.6 | 73.3 74.5 78.5 | 80.2 81.0 85.5 | 83.3 82.7 87.0 | 82.0 82.7 85.9 | 81.5 83.2 86.3 | 79.7 81.3 84.0 | 80.1 78.8 83.5 | 78.3 77.2 80.5 | 78.5 77.8 81.7 | 80.8 81.4 | |
| MI-FGSM NI-FGSM EPNI-FGSM(Ours) | 46.4 49.0 68.1 | 44.0 45.6 65.3 | 45.4 48.5 68.4 | 99.6* 100.0* 100.0* | 40.0 40.6 55.2 | 39.2 38.7 53.8 | 41.3 42.1 55.5 | 33.9 34.9 48.4 | 32.7 32.9 47.9 | 49.9 53.5 63.8 | 53.2 54.3 67.8 | 51.4 54.6 66.6 | 50.6 54.3 65.3 | 38.6 37.5 54.7 | 35.3 37.5 50.7 | 35.0 33.8 49.0 | 32.7 33.2 48.1 | 84.7 45.2 46.5 60.5 | |
| DI-MI-FGSM TI-MI-FGSM SI-NI-FGSM DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) SI-EPNI-FGSM (Ours) VT-MI-FGSM | 63.2 41.5 74.7 82.8 65.1 | 62.3 44.5 81.7 83.2 68.2 | 65.4 42.0 75.1 84.1 67.6 | 98.4* 97.9* 99.3* 100.0* 99.9* 100.0* | 54.2 47.1 65.4 70.7 66.6 | 54.2 47.6 65.8 72.3 69.5 | 55.7 47.3 69.7 74.4 68.9 | 48.6 42.9 65.4 67.5 65.2 | 51.0 41.8 64.7 71.1 63.2 | 59.3 55.9 68.9 74.6 73.8 | 64.4 56.7 71.6 79.5 75.0 | 62.8 55.5 71.8 78.2 73.6 | 64.7 53.9 73.0 79.6 74.2 | 57.0 49.2 72.1 75.0 72.9 | 52.8 45.3 68.7 73.0 68.4 | 51.8 46.9 67.3 69.6 69.1 | 52.6 41.9 68.0 70.9 65.4 | 59.9 50.5 72.0 76.9 71.0 | |
| VT-NI-FGSM VT-EPNI-FGSM(Ours) | 90.9 63.2 66.3 81.0 | 93.1 66.6 71.0 82.1 | 88.6 68.8 99.6* 73.1 99.8* | 79.1 57.8 59.3 | 81.3 57.8 59.3 | 85.5 60.7 62.3 75.4 | 79.6 53.6 56.9 70.0 | 82.9 54.9 56.9 70.6 | 83.2 61.7 66.8 77.3 | 85.5 63.5 68.2 80.4 | 86.0 64.4 68.2 79.9 | 86.1 64.0 68.1 | 86.3 59.8 60.9 | 83.6 56.4 58.0 73.4 | 81.5 54.2 57.2 | 84.6 56.1 58.5 | 85.8 62.5 65.3 | ||
| FI-MI-FGSM FI-EPNI-FGSM (Ours) | 76.3 78.0 | 76.0 77.3 | 83.1 76.3 89.7* 78.0 91.0* | 100.0* 71.7 67.5 69.8 | 72.4 68.0 69.9 | 69.7 72.2 | 66.8 68.9 | 65.5 67.4 | 72.5 73.5 | 73.7 74.0 | 73.4 74.5 | 78.7 73.5 73.6 | 77.3 70.7 71.7 | 68.3 69.0 | 72.0 66.8 66.9 | 71.9 67.0 67.2 | 77.5 71.9 73.1 | ||
| R152 | MI-FGSM NI-FGSM EPNI-FGSM(Ours) | 68.9 75.7 91.5 | 59.4 64.1 85.8 | 53.8 58.0 79.2 74.2 | 49.4 81.0 51.7 86.0 | 83.3 87.7 94.8 96.2 | 92.1 94.7 98.7 | 94.5 96.9 99.4 | 100.0* 100.0* 100.0* | 72.7 74.1 76.1 77.4 88.7 89.8 | 72.5 77.0 89.5 | 72.5 76.2 89.6 | 86.5 87.4 96.5 | 83.0 85.1 96.5 | 82.5 84.0 85.5 86.7 97.6 96.7 | 77.1 80.4 92.0 | |||
| DI-MI-FGSM TI-MI-FGSM SI-NI-FGSM DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) SI-EPNI-FGSM(Ours) | 85.6 57.1 85.1 97.6 80.3 | 82.9 51.3 77.2 96.9 78.2 | 75.5 50.1 71.1 94.2 75.7 | 72.1 41.4 66.5 92.1 | 91.8 70.9 89.6 98.0 89.3 | 93.8 75.0 92.0 98.9 91.4 | 96.0 80.5 95.3 99.4 94.0 | 96.8 84.5 97.8 99.6 95.5 | 100.0* 100.0* 100.0* 100.0* 100.0* | 84.0 65.4 82.1 94.3 81.7 | 84.4 65.0 83.0 94.6 81.2 | 85.5 64.1 81.9 94.7 80.8 | 84.6 64.0 81.9 95.2 81.5 | 93.7 75.1 93.6 99.4 91.8 | 94.0 71.5 91.8 99.3 89.2 | 93.7 70.7 92.4 99.6 91.0 | 94.4 68.1 91.8 99.1 89.7 | 88.8 67.9 86.7 97.2 85.8 | |
| VT-MI-FGSM VT-NI-FGSM VT-EPNI-FGSM(Ours) | 95.3 83.8 87.7 95.4 | 91.4 79.6 81.9 92.1 | 88.2 74.1 78.9 | 66.8 84.8 70.9 74.4 | 96.7 92.2 93.5 | 97.6 93.7 94.9 | 98.6 96.4 98.2 | 99.0 97.5 98.7 | 100.0* 100.0* 100.0* | 90.3 83.6 85.3 | 91.2 82.8 87.2 | 92.2 84.4 88.2 | 90.5 83.8 86.9 | 99.0 94.2 96.1 | 98.2 92.4 94.9 | 98.1 93.1 95.1 | 98.2 93.6 95.7 | 94.7 88.0 90.4 | |
| FI-MI-FGSM FI-EPNI-FGSM(Ours) MI-FGSM | 93.3 95.4 78.0 | 88.6 93.2 | 89.7 88.7 93.1 | 86.9 85.1 90.1 | 97.9 95.5 96.8 | 98.8 96.8 97.9 | 99.5 97.5 98.3 | 99.8 98.9 98.9 | 100.0* 99.9* 100.0* | 93.6 92.5 94.1 | 93.5 92.0 94.8 | 93.7 93.0 95.0 | 94.4 93.8 95.1 | 99.4 97.0 97.6 | 98.7 95.7 97.4 | 99.4 96.6 97.3 | 98.8 96.9 97.6 | 96.0 94.2 96.0 | |
| NI-FGSM EPNI-FGSM(Ours) DI-MI-FGSM TI-MI-FGSM SI-NI-FGSM V16 | 78.8 93.5 88.4 60.6 | 59.2 61.6 82.3 75.6 51.1 | 63.8 68.8 87.0 77.5 | 43.8 47.3 68.1 59.5 | 80.0 82.3 91.5 87.8 | 73.3 76.1 90.0 83.9 | 75.0 78.0 90.0 86.2 | 63.7 67.0 82.9 73.8 53.6 | 58.7 60.4 77.6 69.7 48.4 | 94.7 96.8 99.0 98.3 | 98.3 99.2 100.0 98.8 | 99.8* 99.9* 100.0* 100.0* | 99.2 99.1 99.9 99.4 | 77.0 79.1 91.2 88.0 | 69.5 72.1 87.3 81.5 | 65.3 66.4 85.8 78.2 | 64.9 66.0 87.0 77.5 | 74.4 76.4 | 89.0 83.8 64.6 |
| VT-MI-FGSM VT-NI-FGSM | DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) SI-EPNI-FGSM(Ours) | 89.8 96.2 82.5 96.1 87.5 | 77.5 90.6 74.6 89.1 75.2 | 48.5 80.7 62.5 92.6 81.9 77.4 58.1 91.8 80.0 77.9 61.9 | 33.4 | 71.2 89.1 95.9 88.4 93.6 89.9 | 65.4 84.8 94.1 84.2 91.2 86.1 | 60.3 85.9 93.5 82.7 92.5 86.8 | 77.1 88.6 75.9 86.4 78.3 | 73.0 85.5 70.4 84.8 74.1 | 88.1 98.5 99.8 96.8 99.4 97.7 | 92.9 99.6 99.9 98.2 100.0 98.9 | 99.8* 100.0* 100.0* 100.0* 100.0* 99.9* | 94.4 100.0 100.0 98.1 100.0 99.4 | 64.9 87.8 95.6 86.4 95.6 87.8 | 58.0 81.9 93.1 80.7 92.6 82.2 | 54.8 78.9 93.0 78.8 91.0 81.3 | 52.4 80.2 91.4 77.4 92.2 80.6 | 85.1 93.6 83.0 92.7 85.0 |
| VT-EPNI-FGSM(Ours) FI-MI-FGSM | 89.8 95.6 95.9 | 76.9 88.4 89.1 | 80.7 65.6 92.4 81.1 93.1 79.7 | 92.1 96.1 95.6 | 87.5 94.4 94.9 | 89.3 94.0 93.4 | 81.1 90.4 90.4 | 76.1 88.8 87.1 | 98.7 99.4 99.6 | 99.5 99.9 99.8 | 99.9* 99.9* 100.0* | 99.4 99.9 99.8 | 88.7 95.6 94.3 | 85.4 93.7 91.5 | 82.5 93.2 90.2 | 82.9 92.2 88.6 | 86.8 93.8 93.1 | ||
| FI-EPNI-FGSM(Ours) | 96.7 | 90.4 | 94.1 84.3 | 96.7 | 95.9 | 96.0 | 92.6 | 90.1 | 99.8 | 99.8 | 100.0* | 99.9 | 96.4 | 94.3 | 93.3 | 93.0 | 94.9 |
| Models | Attacks | Iv3adv | IRv2ens | Eb0adv | Ebladv | Eb2adv | Eb3adv | Eb4adv | Eb5adv | Eb6adv | Eb7adv | Avg |
| Iv3 | MI-FGSM NI-FGSM EPNI-FGSM(Ours) DI-MI-FGSM | 25.4 25.4 32.4 33.0 | 11.5 11.6 14.3 19.3 | 29.7 34.3 49.0 46.6 | 26.5 31.0 45.0 42.9 | 28.1 31.2 45.4 42.4 | 18.8 22.2 31.1 31.3 | 16.9 18.6 23.9 27.4 | 17.5 18.6 25.7 26.7 | 14.8 16.0 22.4 | 15.2 16.5 22.3 24.4 | 20.4 22.5 31.2 31.7 |
| TI-MI-FGSM SI-NI-FGSM DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) | 29.9 37.4 41.5 54.1 | 21.5 23.2 25.2 41.2 | 35.2 50.2 65.0 60.7 | 32.6 46.1 65.5 58.2 | 34.9 46.3 64.7 57.0 | 27.8 33.3 45.5 46.6 | 29.3 30.9 39.1 46.7 | 28.2 29.6 41.6 45.3 | 22.9 24.5 27.6 36.7 43.4 | 26.6 25.8 35.6 44.2 | 29.1 35.0 46.0 49.7 | |
| SI-EPNI-FGSM(Ours) VT-MI-FGSM VT-NI-FGSM | 47.4 36.8 39.1 | 28.4 25.1 26.4 | 66.3 50.5 54.0 | 64.7 46.3 50.6 | 63.1 44.9 49.5 | 46.7 33.5 36.6 | 41.8 29.4 31.0 | 42.1 29.4 32.7 | 37.9 26.8 30.3 | 37.3 26.2 28.5 | 47.6 34.9 37.9 | |
| VT-EPNI-FGSM(Ours) FI-MI-FGSM FI-EPNI-FGSM(Ours) | 43.1 55.3 58.9 | 26.5 38.0 39.9 | 63.2 68.2 76.2 | 61.8 67.8 73.3 | 60.6 64.9 72.7 | 46.9 53.6 61.1 | 41.1 50.0 56.0 | 39.7 49.3 55.6 | 38.2 46.6 52.9 | 36.8 45.1 50.5 | 45.8 53.9 59.7 | |
| MI-FGSM NI-FGSM EPNI-FGSM(Ours) DI-MI-FGSM | 27.3 25.9 33.5 33.1 | 15.5 14.9 15.9 24.7 | 25.2 24.8 36.3 | 22.8 23.1 32.4 | 24.3 23.6 34.9 | 17.0 17.8 23.9 27.1 | 13.8 14.1 20.7 24.6 | 14.6 14.9 20.4 23.2 | 12.1 12.8 17.6 22.0 | 13.2 13.7 17.8 20.8 | 18.6 18.6 25.3 28.7 | |
| TI-MI-FGSM SI-NI-FGSM DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) SI-EPNI-FGSM(Ours) | 38.1 37.0 39.4 58.3 45.6 | 32.2 33.0 29.7 52.9 41.5 | 38.1 39.0 50.8 57.4 60.5 68.9 | 35.8 36.7 48.2 52.4 58.3 | 37.9 39.4 46.9 53.4 56.6 66.8 | 30.2 35.7 41.0 49.7 50.2 | 33.8 32.0 33.6 50.9 42.6 | 31.4 28.7 33.3 47.0 43.5 | 29.2 26.9 30.2 48.0 39.7 | 29.7 27.5 30.6 46.5 40.1 | 34.0 36.7 40.1 52.9 50.3 | |
| VT-MI-FGSM VT-NI-FGSM VT-EPNI-FGSM(Ours) FI-MI-FGSM | 38.8 40.5 48.2 | 36.4 34.8 36.6 | 42.3 44.6 59.5 | 64.3 38.9 41.1 55.3 | 41.2 43.4 54.1 | 31.2 33.1 43.0 | 28.4 28.2 38.8 | 26.4 26.6 38.0 | 25.4 25.8 36.9 | 25.4 26.2 35.3 | 33.4 34.4 44.6 46.7 | |
| FI-EPNI-FGSM(Ours) MI-FGSM NI-FGSM EPNI-FGSM(Ours) | 54.5 53.3 36.5 40.1 | 40.1 47.1 27.8 29.4 | 57.4 59.6 46.6 49.8 | 56.7 56.8 43.6 47.3 | 56.3 57.9 47.3 47.8 | 45.7 47.0 30.9 33.2 | 40.9 43.3 27.3 29.1 | 39.7 42.1 26.5 27.7 | 38.2 40.6 22.0 25.2 | 37.4 38.6 24.8 25.4 | 48.6 33.3 35.5 | |
| DI-MI-FGSM TI-MI-FGSM SI-NI-FGSM R152 | 53.1 57.6 46.8 51.7 DI-EPNI-FGSM(Ours) 78.3 | 43.1 51.7 41.1 43.4 73.1 | 71.9 72.6 50.7 62.8 91.6 | 68.1 68.3 46.4 57.3 89.9 | 70.5 71.2 50.1 61.2 90.6 | 50.2 54.7 41.8 43.4 75.7 | 41.0 47.1 43.2 39.3 67.3 | 41.0 37.5 44.7 44.2 41.0 36.2 36.0 35.4 64.9 62.6 | 40.2 44.3 38.7 33.8 64.3 | 51.7 55.6 43.6 46.4 75.8 | ||
| VT-MI-FGSM VT-NI-FGSM FI-MI-FGSM | TI-EPNI-FGSM(Ours) SI-EPNI-FGSM(Ours) VT-EPNI-FGSM(Ours) | 72.2 67.4 67.5 59.3 59.1 52.6 61.6 55.5 72.6 71.6 | 74.7 79.4 69.6 73.6 86.2 | 70.1 78.5 64.9 68.2 85.3 | 74.2 81.1 69.3 72.1 88.4 | 62.9 61.8 55.1 56.1 74.1 | 62.7 53.7 48.8 50.6 66.0 | 60.7 51.5 46.1 48.7 64.8 | 59.3 58.3 48.8 50.3 43.6 44.6 45.5 47.9 62.2 65.1 | 66.3 63.2 55.4 58.0 73.6 | ||
| TI-MI-FGSM V16 | FI-EPNI-FGSM(Ours) MI-FGSM NI-FGSM EPNI-FGSM(Ours) DI-MI-FGSM | 81.0 33.0 33.0 46.7 44.9 | 73.3 20.3 22.5 29.0 31.6 | 88.5 48.8 49.5 71.2 62.4 | 87.0 41.8 43.6 65.7 55.9 | 82.6 88.6 41.3 44.3 63.5 56.9 | 76.3 25.1 27.7 41.5 37.5 | 70.9 21.5 22.2 31.3 32.0 | 70.1 21.7 22.7 31.5 30.9 | 66.4 19.0 18.9 28.5 27.8 | 66.4 21.8 20.1 28.7 28.1 | 76.9 29.4 30.5 43.8 40.8 |
| SI-NI-FGSM DI-EPNI-FGSM(Ours) TI-EPNI-FGSM(Ours) SI-EPNI-FGSM(Ours) VT-MI-FGSM VT-NI-FGSM | 38.9 50.0 61.2 61.8 62.8 49.4 49.3 | 29.4 33.5 44.7 49.9 42.6 35.4 37.5 | 44.3 61.4 80.3 68.6 78.6 65.3 | 38.0 56.4 77.9 65.2 75.4 58.7 | 41.4 54.4 78.2 63.7 74.3 58.6 | 31.1 36.5 52.2 49.2 48.1 39.7 | 32.0 29.9 43.5 50.3 40.6 34.3 | 31.5 30.0 42.6 47.7 41.9 33.6 | 28.5 27.2 40.2 47.6 37.9 29.2 | 28.2 26.4 38.8 46.1 35.5 32.2 | 34.3 40.6 56.0 55.0 53.8 |
| Model | CLEVR-Building-blocks | Sapien-Kitchen | ||||||
| SSIM↑ | MSE↓ | LPIPS↓ | Accuracy↑ | SSIM↑ | MSE↓ | LPIPS↓ | Accuracy↑ | |
| Copy-First-Frame | 0.962 | 251.38 | 0.1320 | - | 0.951 | 152.87 | 0.0393 | 1 |
| Concatenation Baseline | 0.961 | 226.53 | 0.1301 | 50.8% | 0.962 | 23.13 | 0.0232 | 52.4% |
| AG2Vid | 0.956 | 58.67 | 0.0399 | 78.8% | 0.947 | 270.87 | 0.0684 | 5.2% |
| MAC | 0.983 | 43.52 | 0.0303 | 95.2% | 0.971 | 11.16 | 0.0178 | 86.4% |
| Method | Oven | Fridge | DishwashetBottle | Kettle | Kitchenpot | mAP | |
| MACencoder+SSD | 92.75 | 94.56 | 90.89 | 83.25 | 77.18 | 81.32 | 86.66 |
| SDE | Ft | Gt | μt | Σt |
| VPSDE (Ho et al., 2020) | dlogatI 1 2 dt | dlogat I dt | VatI | (1-αt)I |
| VESDE (Song et al., 2020b) | 0 | I dt | I | 1 |
| FID for various DEIS | ||||||||||
| NFE | DDIM | p2Heunt | p3Kutta | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 |
| 5 | 26.91 | 108+1 | 185+1 | 193+3 | 22.28 | 21.53 | 21.43 | 19.72 | 16.31 | 15.37 |
| 10 | 11.14 | 14.72 | 13.19+2 | 28.65+2 | 7.56 | 6.72 | 6.50 | 6.09 | 4.57 | 4.17 |
| 15 | 7.06 | 4.89+1 | 5.88 | 6.88+1 | 4.69 | 4.16 | 3.99 | 4.29 | 3.57 | 3.37 |
| 20 | 5.47 | 3.50 | 2.97+1 | 3.92 | 3.70 | 3.32 | 3.17 | 3.54 | 3.05 | 2.86 |
| 50 | 3.27 | 2.60 | 2.55+1 | 2.57+2 | 2.70 | 2.62 | 2.59 | 2.67 | 2.59 | 2.57 |
| 10 | 12 | 14 | 16 | 18 20 | 30 | 50 | ||
| tAB pAB | 6.65 9.28 | 3.99 6.46 | 3.67 5.02 | 3.49 4.34 | 3.21 3.74 | 3.10 3.39 | 2.81 2.87 | 2.69 2.66 |
| DPM-Solver2 pMid | 7.93 9.12 | 5.36 6.78 | 4.46 5.27 | 3.89 4.52 | 3.63 4.00 | 3.42 3.66 | 3.00 2.99 | 2.82 2.81 |
| DPM-Solver3 pKutta | 5.02 13.12 | 3.62+ 5.18+ | 3.18 3.63 | 3.06+ 3.16+ | 2.84 2.82 | 2.72+ 2.71+ |
| FIDNFEMethod | 5 | 10 | 20 | 50 | |||
| PNDMiPNDMDDIMtAB1tAB2 | =70.07 | 9.36 | 6.424.216.12 | 3.033.004.253.323.092.99 | |||
| 30.64 | 11.71 | ||||||
| 71 | 6.12 | ||||||
| 20.01 | 6.09 | 3.81 | |||||
| 16.53 | 4.57 | 3.41 | |||||
| tAB3 | 16.10 | 4.17 | 3.33 | ||||
| FIDNFEMethod | 5 | 10 | 20 | 50 |
| PNDMiPNDM | 59.87 | 7.78 | 7.605.58 | 3.513.34 |
| O-DEIS1-DEIS | 30.42 | 13.53 | 6.89 | 4.173.19 |
| 26.65 | 8.81 | 4.33 | ||
| 2-DEIS | 25.13 | 7.20 | 3.61 | 3.04 |
| 3-DEIS | 25.07 | 6.95 | 3.41 | 2.95 |
| NFE | DDIM | p2Heun | p3Kutta | p4RK | pAB1 | pAB2 | FID for various DEIS with κ = 1 in Eq. (42) pAB3 | tAB1 | tAB2 | tAB3 |
| 5 | 47.59 | 207+1 | 238+1 | 212+3 | 35.14 | 32.51 | 32.02 | 25.99 | 25.06 | 44.29 |
| 10 | 16.60 | 84.55 | 66.81+2 | 78.57+2 | 10.47 | 8.85 | 8.18 | 9.51 | 7.71 | 7.18 |
| 15 | 10.39 | 46.36+1 | 47.45 | 41.27+1 | 6.69 | 5.70 | 5.24 | 6.47 | 5.51 | 5.01 |
| 20 | 7.93 | 34.87 | 28.35+1 | 27.21 | 5.27 | 4.56 | 4.24 | 5.20 | 4.50 | 4.14 |
| 50 | 4.36 | 11.58 | 7.00+1 | 7.48+2 | 3.32 | 3.08 | 2.99 | 3.32 | 3.09 | 2.99 |
| NFE | p2Heun | |||||||||
| DDIM | p3Kutta | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 | ||
| 5 | 30.64 | 256+1 | 357+1 | 342+3 | 24.58 | 23.60 | 23.48 | 20.01 | 16.52 | 16.10 |
| 10 | 11.71 | 56.62 | 56.51+2 | 103+2 | 7.56 | 6.72 | 6.50 | 6.09 | 4.57 | 4.17 |
| 15 | 7.67 | 10.62+1 | 14.96 | 36.15+1 | 4.93 | 4.40 | 4.26 | 4.29 | 3.57 | 3.37 |
| 20 | 6.11 | 6.33 | 4.74+1 | 12.81 | 4.16 | 3.84 | 3.77 | 3.81 | 3.41 | 3.33 |
| 50 | 4.24 | 3.88 | 3.75+1 | 3.78+2 | 3.70 | 3.68 | 3.69 | 3.62 | 3.61 | 3.36 |
| NFE DDIM | FID for various DEIS with κ = 3inEq. (42) | |||||||||
| 5 | 34.07 | p2Heun 356+1 | p3Kutta 388+1 | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 22.06 |
| 10 | 14.59 | 115 | 171+2 | 377+3 267+2 | 29.80 | 29.35 | 29.38 | 24.87 | 22.57 | 5.97 |
| 15 | 9.22 | 32.94+1 | 103+1 | 10.73 | 10.16 | 10.11 | 8.11 | 6.36 | 4.05 | |
| 20 | 7.27 | 77.44 11.55+1 | 6.45 | 6.03 | 5.98 | 5.21 | 4.26 | 3.75 | ||
| 50 | 4.64 | 13.06 3.76 | 3.68+1 | 50.56 | 5.17 | 4.83 | 4.78 | 4.45 | 3.88 | |
| 3.74+2 | 3.92 | 3.82 | 3.79 Eq. (44) | 3.81 | 3.72 | 3.71 | ||||
| NFE | DDIM | p2Heun | p3Kutta | FID for various DEISwith | ||||||
| 5 | 54.58 | 216+1 | 335+1 | p4RK 313+3 | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 |
| 10 | 20.03 | 14.72 | 13.19+2 | 49.25 | 48.56 | 48.47 | 37.99 | 28.45 | 26.11 5.71 | |
| 15 | 5.03+1 | 28.65+2 | 14.05 | 12.63 | 12.18 | 10.36 | 7.03 | 4.13 | ||
| 20 | 11.99 8.92 | 4.12 | 5.88 3.97+1 | 6.88+1 | 7.72 | 6.67 | 6.29 | 6.22 | 4.69 | 3.80 |
| 3.75+1 | 4.14 | 5.79 | 5.05 | 4.78 | 4.97 | 4.10 | ||||
| 50 | 5.05 | 3.67 | 3.73+2 | 4.01 | 3.84 | 3.79 | 3.89 | 3.74 | 3.72 | |
| NFE | DDIM | p2Heun | p3Kutta | p4RK | pAB1 | pAB2 | FID for various DEIS with κ = 1 in Eq. (42) pAB3 | tAB1 | tAB2 | tAB3 |
| 5 | 42.38 | 239+1 | 232+1 | 199+3 | 33.09 | 31.06 | 30.67 | 26.01 | 20.57 | 42.35 |
| 10 | 17.23 | 143 | 95.13+2 | 130+2 | 12.44 | 11.04 | 10.41 | 12.01 | 10.57 | 8.04 |
| 15 | 12.06 | 99.76+1 | 77.37 | 88.56+1 | 9.12 | 8.25 | 7.79 | 9.08 | 8.20 | 7.50 |
| 20 | 9.71 | 82.89 | 57.54+1 | 66.61 | 7.57 | 6.89 | 6.50 | 7.60 | 6.90 | 6.45 |
| 50 | 5.76 | 31.56 | 13.10+1 | 15.73+2 | 4.64 | 4.25 | 4.06 | 4.67 | 4.28 | 4.10 |
| NFE | ||||||||||
| DDIM | p2Heun | p3Kutta | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 | |
| 5 | 26.91 | 271+1 | 362+1 | 348+3 | 22.28 | 21.53 | 21.43 | 19.72 | 16.31 | 15.37 |
| 10 | 11.14 | 66.25 | 63.53+2 | 111+2 | 7.65 | 6.89 | 6.67 | 6.74 | 5.49 | 5.02 |
| 15 | 7.06 | 13.48+1 | 17.15 | 44.83+1 | 4.69 | 4.16 | 3.99 | 4.38 | 3.78 | 3.50 |
| 20 | 5.47 | 6.62 | 4.15+1 | 15.14 | 3.70 | 3.32 | 3.17 | 3.57 | 3.19 | 3.03 |
| 50 | 3.27 | 2.65 | 2.55+1 | 2.57+2 | 2.70 | 2.62 | 2.59 | 2.70 | 2.61 | 2.59 |
| NFE DDIM | FID for various DEIS with κ = 3inEq. | |||||||||
| 5 | 32.11 | p2Heun 364+1 | p3Kutta 393+1 | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 23.38 |
| 10 | 13.18 | 135 | 199+2 | 383+3 298+2 | 28.87 | 28.58 | 28.62 | 25.78 7.74 | 23.66 6.20 | 5.77 |
| 15 | 7.92 | 42.04+1 | 99.64 | 122+1 | 9.89 | 9.38 | 9.33 | 4.48 | 3.65 | 3.37 |
| 20 | 5.92 | 17.05 | 16.66+1 | 64.40 | 5.41 | 4.99 3.69 | 4.91 3.60 | 3.54 | 3.05 | 2.86 |
| 50 | 3.36 | 2.77 | 2.57+1 | 2.71+2 | 4.04 | 2.63 | 2.60 | 2.67 | 2.57 | |
| FID for various I | 2.73 | DEISwith | Eq. (44) | 2.59 | ||||||
| NFE | DDIM | p2Heun | p3Kutta | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 |
| 5 | 54.85 | 230+1 | 382+1 | 370+3 | 51.94 | 51.62 | 51.58 | 43.84 | 39.91 | 38.76 |
| 10 | 19.80 | 23.35 | 25.08+2 | 82.17+2 | 14.63 | 13.43 | 13.07 | 11.14 | 7.78 | 6.02 |
| 15 | 11.29 | 5.63+1 | 7.46 | 8.90+1 | 7.31 | 6.28 | 5.90 | 5.89 | 4.35 | 3.71 |
| 20 | 7.91 | 3.84 | 3.05+1 | 4.14 | 4.91 | 4.19 | 3.91 | 4.23 | 3.35 | 3.00 |
| 50 | 3.82 | 2.60 | 2.56+1 | 2.59+2 | 2.86 | 2.70 | 2.64 | 2.79 | 2.63 | 2.58 |
| FID for various DEIS with κ = 7 in Eq. (43) | ||||||||||
| NFE | DDIM | p2Heun | p3Kutta | p4RK | pAB1 | pAB2 | pAB3 | tAB1 | tAB2 | tAB3 |
| 5 | 53.20 | 108+1 | 185+1 | 193+3 | 47.56 | 46.36 | 46.13 | 36.98 | 28.76 | 25.76 |
| 10 | 18.99 | 18.75 | 20.27+2 | 54.92+2 | 13.38 | 11.84 | 11.21 | 10.92 | 8.26 | 6.87 |
| 15 | 10.91 | 4.89+1 | 6.31 | 9.79+1 | 6.90 | 5.86 | 5.42 | 6.12 | 4.86 | 4.33 |
| 20 | 7.81 | 3.50 | 2.97+1 | 3.92 | 4.84 | 4.10 | 3.80 | 4.48 | 3.69 | 3.38 |
| 50 | 3.84 | 2.60 | 2.58+1 | 2.59+2 | 2.86 | 2.69 | 2.64 | 2.82 | 2.66 | 2.61 |
| FID with various NFE | |||||||||
| Method | 5 | 10 | 20 | 30 | 50 | 100 | 200 | 500 | 1000 |
| Euler | 246.16 | 90.52 | 27.38 | 14.99 | 8.46 | 4.96 | 3.54 | 2.81 | 2.62 |
| +EI | 283.67 | 216.47 | 137.20 | 100.74 | 68.03 | 37.93 | 18.81 | 6.66 | 3.69 |
| +θ | 42.38 | 17.23 | 9.71 | 7.56 | 5.76 | 4.24 | 3.37 | 2.83 | 2.67 |
| +Poly | 30.67 | 10.41 | 6.50 | 5.13 | 4.06 | 3.07 | 2.69 | 2.58 | 2.57 |
| +Opt{ti} | 15.37 | 5.02 | 3.03 | 2.70 | 2.59 | 2.57 | 2.56 | 2.56 | 2.56 |
| FID with various NFE | |||||||||
| Method | 5 | 10 | 20 | 30 | 50 | 100 | 200 | 500 | 1000 |
| Uniform | 246.16 | 90.52 | 27.38 | 14.99 | 8.46 | 4.96 | 3.54 | 2.81 | 2.62 |
| Quadratic | 294.01 | 138.73 | 39.82 | 19.26 | 8.49 | 3.96 | 2.88 | 2.61 | 2.57 |
| NFE | 14 | 26 | 32 | 38 | 50 | 62 | 88.2 | 344 |
| FID | 61.11 | 36.64 | 15.18 | 9.88 | 6.32 | 2.63 | 2.56 | 2.55 |
| FIDNFEMethod | 5 | 10 | 20 | 50 | |
| A-DDIM1-iPNDM2-iPNDM3-iPNDMtAB1-DEIStAB2-DEIStAB3-DEIS | 51.4730.13 | 14.0613.01 | 6.748.256.795.794.914.664.53 | 4.045.654.734.243.883.793.78 | |
| 84.00 | 10.4514.03 | ||||
| 10 | 105.38 | ||||
| 20.45 | 8.11 | ||||
| 18.8718.43 | 7.477.12 | ||||
| FIDNFEMethod | 5 | 10 | 20 | 50 | ||
| iPNDMDDIMtAB1-DEIS | 54.6249.08 | 15.3223.52 | 9.2613.69 | 8.269.448.418.298.25 | ||
| tAB1-DEIS | 34.69 | 13.94 | 9.55 | |||
| tAB2 | tAB2-DEIS | 29.50 | 11.36 | 8.79 | ||
| tAB3-DEIS | 28.09 | 10.55 | 8.58 | |||
| Dataset | FIDNFEMethod | 5 | 10 | 20 | 50 |
| CELEBA | PNDMiPNDMDDIMtAB1-DEIStAB2-DEIStAB3-DEIS | 59.87±1.0130.42±0.8726.65±0.6325.13±0.5625.07±0.49 | 7.78±0.1813.53±0.488.81±0.237.20±0.216.95±0.09 | 7.60±0.125.58±0.116.89±0.114.33±0.07 | 3.51±0.033.34±0.044.17±0.043.19±0.03 |
| 3.61±0.053.41±0.04 | 3.04±0.022.95±0.03 |
| SDE | FIDNFEMethod | 5 | 10 | 20 | 50 |
| VESDE | tAB0-DEIStAB1-DEIStAB2-DEIStAB3-DEIS | 103.52±2.0956.33±0.8758.65±0.2596.70±0.90 | 46.90±0.3826.16±0.1220.89±0.0925.01±0.03 | 27.64±0.0518.52±0.0316.94±0.0316.59±0.03 | 19.86±0.0316.64±0.0116.33±0.0216.31±0.02 |
| URL | License |
| https://github.com/yang-song/score_sde | Apache License 2.0 |
| https://github.com/luping-liu/PNDM | Apache License 2.0 |
| https://github.com/CompVis/latent-diffusion | MIT |
| https://github.com/baofff/Analytic-DPM | Unknown |
| Experiment | Citation | License |
| CIFAR10 Tabs. 2, 6 to 8, 11, 14 and 15, FFHQ Fig. 1 | (Song et al., 2020b) | Apache License 2.0 |
| CIFAR10 Tab.12 | (Bao et al., 2022) | Unknown |
| CELEBA Tab.14 | (Liu et al., 2022) | Apache License 2.0 |
| ImageNet 32 × 32 Tab. 13 | (Song et al., 2021a) | Unknown |
| Text-to-image | (Rombach et al., 2021) | MIT |
| ResNet-152 | ResNet-152- SAM | ViT-B/16 | ViT-B/16- SAM | Mixer-B/16 | Mixer-B/16- SAM | |
| #Params | 60M | 87M | 59M | |||
| NTK κ † Hessian Xmax | 2801.6 | 4205.3 | 14468.0 | |||
| 179.8 | 42.0 | 738.8 | 20.9 | 1644.4 | 22.5 | |
| Ltrain | 0.86 | 0.90 | 0.65 | 0.82 | 0.45 | 0.97 |
| ★ | 2.39 | 2.16 | 6.66 | 0.96 | 7.78 | 1.01 |
| ImageNet (%) | 78.5 | 79.3 | 74.6 | 79.9 | 66.4 | 77.4 |
| ImageNet-C (%) | 50.0 | 52.2 | 46.6 | 56.5 | 33.8 | 48.8 |
| Model | #params | Throughput (img/sec/core) | ImageNet | ReaL | V2 | ImageNet-R | ImageNet-C |
| ResNet | |||||||
| ResNet-50-SAM | 25M | 2161 | 76.7 (+0.7) | 83.1 (+0.7) | 64.6 (+1.0) | 23.3 (+1.1) | 46.5 (+1.9) |
| ResNet-101-SAM | 44M | 1334 | 78.6 (+0.8) | 84.8 (+0.9) | 66.7 (+1.4) | 25.9 (+1.5) | 51.3 (+2.8) |
| ResNet-152-SAM | 60M | 935 | 79.3 (+0.8) | 84.9 (+0.7) | 67.3 (+1.0) | 25.7 (+0.4) | 52.2 (+2.2) |
| ResNet-50x2-SAM | 98M | 891 | 79.6 (+1.5) | 85.3 (+1.6) | 67.5 (+1.7) | 26.0 (+2.9) | 50.7 (+3.9) |
| ResNet-101x2-SAM | 173M | 519 | 80.9 (+2.4) | 86.4 (+2.4) | 69.1 (+2.8) | 27.8(+3.2) | 54.0 (+4.7) |
| ResNet-152x2-SAM | 236M | 356 | 81.1 (+1.8) | 86.4 (+1.9) | 69.6 (+2.3) | 28.1 (+2.8) | 55.0 (+4.2) |
| Vision Transformer | |||||||
| ViT-S/32-SAM | 23M | 6888 | 70.5 (+2.1) | 77.5 (+2.3) | 56.9 (+2.6) | 21.4 (+2.4) | 46.2 (+2.9) |
| ViT-S/16-SAM | 22M | 2043 | 78.1 (+3.7) | 84.1 (+3.7) | 65.6(+3.9) | 24.7 (+4.7) | 53.0 (+6.5) |
| ViT-S/14-SAM | 22M | 1234 | 78.8 (+4.0) | 84.8 (+4.5) | 67.2(+5.2) | 24.4 (+4.7) | 54.2 (+7.0) |
| ViT-S/8-SAM | 22M | 333 | 81.3 (+5.3) | 86.7 (+5.5) | 70.4 (+6.2) | 25.3 (+6.1) | 55.6 (+8.5) |
| ViT-B/32-SAM | 88M | 2805 | 73.6 (+4.1) | 80.3 (+5.1) | 60.0 (+4.7) | 24.0 (+4.1) | 50.7 (+6.7) |
| ViT-B/16-SAM | 87M | 863 | 79.9 (+5.3) | 85.2 (+5.4) | 67.5 (+6.2) | 26.4 (+6.3) | 56.5 (+9.9) |
| MLP-Mixer | |||||||
| Mixer-S/32-SAM | 19M | 11401 | 66.7 (+2.8) | 73.8 (+3.5) | 52.4 (+2.9) | 18.6 (+2.7) | 39.3 (+4.1) |
| Mixer-S/16-SAM | 18M | 4005 | 72.9 (+4.1) | 79.8 (+4.7) | 58.9 (+4.1) | 20.1 (+4.2) | 42.0 (+6.4) |
| Mixer-S/8-SAM | 20M | 1498 | 75.9 (+5.7) | 82.5 (+6.3) | 62.3 (+6.2) | 20.5 (+5.1) | 42.4 (+7.8) |
| Mixer-B/32-SAM | 60M | 4209 | 72.4 (+9.9) | 79.0 (+10.9) | 58.0 (+10.4) | 22.8 (+8.2) | 46.2 (12.4) |
| Mixer-B/16-SAM Mixer-B/8-SAM | 59M | 1390 | 77.4 (+11.0) | 83.5 (+11.4) 84.4(+10.1) | 63.9 (+13.1) | 24.7 (+10.2) | 48.8 (+15.0) |
| 64M | 466 | 79.0 (+10.4) | 65.5 (+11.6) | 23.5 (+9.2) | 48.9 (+16.9) | ||
| Model | #params | ImageNet (%) | ImageNet-C (%) |
| R50-S/16 R50-S/16-SAM | 34M | 79.8 81.0 (+1.2) | 53.4 57.2 (+3.8) |
| R50-B/16 | 79.7 | 54.4 | |
| R50-B/16-SAM | 99M | 82.4 (+2.7) | 61.0 (+6.6) |
| Model | Xmax of diagonal blocks of Hessian | |w|l2 | |a1|l2 | |a6|2 | |a12|l2 | ||||||
| Embedding | MSA/ Token MLP | MLP/ Channel MLP | Block1 | Block6 | Block12 | Whole | |||||
| ViT-B/16 | 300.4 | 179.8 | 281.4 | 44.4 | 32.4 | 26.9 | 738.8 | 269.3 | 104.9 | 104.3 | 138.1 |
| ViT-B/16-SAM | 3.8 | 8.5 | 9.6 | 1.7 | 1.7 | 1.5 | 20.9 | 353.8 | 117.0 | 120.3 | 97.2 |
| Mixer-B/16 | 1042.3 | 95.8 | 417.9 | 239.3 | 41.2 | 5.1 | 1644.4 | 197.6 | 96.7 | 135.1 | 74.9 |
| Mixer-B/16-SAM | 18.2 | 1.4 | 9.5 | 4.0 | 1.1 | 0.3 | 22.5 | 389.9 | 110.9 | 176.0 | 216.1 |
| Dataset | ResNet-152 | ViT-B/16 | Mixer-B/16 | |||||||||
| Vanilla | SAM | AUG | SAM + AUG | Vanilla | SAM | AUG | SAM +AUG | Vanilla | SAM | AUG | SAM + AUG | |
| ImageNet | 78.5 | 79.3 | 78.8 | 78.9 | 74.6 | 79.9 | 79.6 | 81.5 | 66.4 | 77.4 | 76.5 | 78.1 |
| i1k (1/2) | 74.2 | 75.6 | 75.1 | 75.5 | 64.9 | 75.4 | 73.1 | 75.8 | 53.9 | 71.0 | 70.4 | 73.1 |
| i1k (1/4) | 68.0 | 70.3 | 70.2 | 70.6 | 52.4 | 66.8 | 63.2 | 65.6 | 37.2 | 62.8 | 61.0 | 65.8 |
| i1k (1/10) | 54.6 | 57.1 | 59.2 | 59.5 | 32.8 | 46.1 | 38.5 | 45.7 | 21.0 | 43.5 | 43.0 | 51.0 |
| Model | Xmax | Ltrain | R(↓) | |
| ViT-B/16 | 738.8 | 0.65 | 6.66 | 57.9% |
| ViT-B/16-SAM | 20.9 | 0.82 | 0.96 | 39.6% |
| ViT-B/16-AUG | 1659.3 | 0.85 | 1.23 | 21.4% |
| Model | #params | Throughput (img/sec/core) | ImageNet | Real | V2 | PGD-10 | ImageNet-R | ImageNet-C |
| ResNet | ||||||||
| ResNet-50-SAM | 25M | 2161 | 70.1 (-0.7) | 77.9 (-0.3) | 56.6(-0.8) | 54.1 (+0.9) | 27.0 (+0.9) | 42.7 (-0.1) |
| ResNet-101-SAM | 44M | 1334 | 73.6(-0.4) | 81.0 (+0.1) | 60.4 (-0.6) | 58.8 (+1.4) | 29.5(+0.6) | 46.9 (+0.3) |
| ResNet-152-SAM | 60M | 935 | 75.1 (-0.4) | 82.3 (+0.2) | 62.2 (-0.4) | 61.0(+1.8) | 30.8 (+1.4) | 49.1 (+0.6) |
| Vision Transformer | ||||||||
| ViT-S/16-SAM | 22M | 2043 | 73.2 (+1.2) | 80.7 (+1.7) | 60.2 (+1.4) | 58.0 (+5.2) | 28.4(+2.4) | 47.5 (+1.6) |
| ViT-B/32-SAM | 88M | 2805 | 69.9 (+3.0) | 76.9 (+3.4) | 55.7 (+2.5) | 54.0 (+6.4) | 26.0 (+3.0) | 46.4 (+3.0) |
| ViT-B/16-SAM | 87M | 863 | 76.7 (+3.9) | 82.9 (+4.1) | 63.6 (+4.3) | 62.0(+7.7) | 30.0 (+4.9) | 51.4 (+5.0) |
| MLP-Mixer | ||||||||
| Mixer-S/16-SAM | 18M | 4005 | 67.1 (+2.2) | 74.5 (+2.3) | 52.8 (+2.5) | 50.1 (+4.1) | 22.9 (+2.6) | 37.9 (+2.5) |
| Mixer-B/32-SAM | 60M | 4209 | 69.3 (+9.1) | 76.4 (+10.2) | 54.7 (+9.4) | 54.5 (+13.9) | 26.3 (+8.0) | 43.7(+8.8) |
| Mixer-B/16-SAM | 59M | 1390 | 73.9 (+11.1) | 80.8 (+11.8) | 60.2 (+11.9) | 59.8 (+17.3) | 29.0 (+10.5) | 45.9 (+12.5) |
| Model | #params | Throughput (img/sec/core) | Patch Resolution | Sequence Length | Hidden Size | #heads | #layers | Token MLP Dimension | Channel MLP Dimension |
| ViT-S/32 | 23M | 6888 | 32×32 | 49 | 384 | 6 | 12 | 一 | 一 |
| ViT-S/16 | 22M | 2043 | 16×16 | 196 | 384 | 6 | 12 | ||
| ViT-S/14 | 22M | 1234 | 14 × 14 | 256 | 384 | 6 | 12 | ||
| ViT-S/8 | 22M | 333 | 8×8 | 784 | 384 | 6 | 12 | ||
| ViT-B/32 | 88M | 2805 | 32×32 | 49 | 768 | 12 | 12 | 一 | 一 |
| ViT-B/16 | 87M | 863 | 16×16 | 196 | 768 | 12 | 12 | 1 | 1 |
| Mixer-S/32 | 19M | 11401 | 32×32 | 49 | 512 | 1 | 8 | 256 | 2048 |
| Mixer-S/16 | 18M | 4005 | 16×16 | 196 | 512 | 一 | 8 | 256 | 2048 |
| Mixer-S/8 | 20M | 1498 | 8×8 | 784 | 512 | 一 | 8 | 256 | 2048 |
| Mixer-B/32 | 60M | 4209 | 32×32 | 49 | 768 | 12 | 384 | 3072 | |
| Mixer-B/16 | 59M | 1390 | 16×16 | 196 | 768 | 一 | 12 | 384 | 3072 |
| Mixer-B/8 | 64M | 466 | 8×8 | 784 | 768 | 12 | 384 | 3072 |
| Dataset | Total steps | Warmup steps | Base LR |
| CIFAR-10 | 10K | 500 | |
| CIFAR-100 | 10K | 500 | {0.001,0.003,0.01,0.03} |
| Flowers | 500 | 100 | |
| Pets | 500 | 100 |
| % | ResNet- 50-SAM | ResNet- 152-SAM | ViT-S/16 | ViT-S/16- SAM | ViT-B/16 | ViT-B/16- SAM | Mixer-S/16 | Mixer-S/16- SAM | Mixer-B/16 | Mixer-B/16- SAM |
| CIFAR-10 | 97.4 | 98.2 | 97.6 | 98.2 | 98.1 | 98.6 | 94.1 | 96.1 | 95.4 | 97.8 |
| CIFAR-100 | 85.2 | 87.8 | 85.7 | 87.6 | 87.6 | 89.1 | 77.9 | 82.4 | 80.0 | 86.4 |
| Flowers Pets | 90.0 | 91.1 | 86.4 | 91.5 | 88.5 | 91.8 | 83.3 | 87.9 | 82.8 | 90.0 |
| 91.6 | 93.3 | 90.4 | 92.9 | 91.9 | 93.1 | 86.1 | 88.7 | 86.1 | 92.5 | |
| Average | 91.1 | 92.6 | 90.0 | 92.6 | 91.5 | 93.2 | 85.4 | 88.8 | 86.1 | 91.7 |
| Model | Task | SAM p |
| ResNet | ||
| ResNet-50-SAM ResNet-101-SAM ResNet-152-SAM ResNet-50x2-SAM ResNet-101x2-SAM ResNet-152x2-SAM ResNet-50-SAM | supervised supervised supervised supervised supervised supervised | 0.02 0.05 0.02 0.05 0.05 0.05 0.05 |
| ResNet-152-SAM adversarial ViT | ||
| ViT-S/16-SAM ViT-S/14-SAM ViT-S/8-SAM ViT-B/32-SAM ViT-B/16-SAM | supervised supervised supervised supervised | 0.1 0.1 0.15 0.15 0.2 |
| ViT-B/16-AUG-SAM ViT-S/16-SAM ViT-B/32-SAM | supervised supervised adversarial | 0.05 0.1 |
| ViT-B/16-SAM | adversarial | 0.1 0.1 |
| supervised contrastive | ||
| adversarial | ||
| ViT-S/16-SAM | 0.02 | |
| ViT-B/16-SAM | ||
| supervised contrastive | 0.02 | |
| MLP-Mixer | ||
| Mixer-S/32-SAM | ||
| Mixer-S/16-SAM | supervised | 0.1 |
| supervised | 0.15 | |
| Mixer-S/8-SAM | supervised | 0.2 |
| Mixer-B/32-SAM | supervised | 0.35 |
| Mixer-B/16-SAM | supervised | 0.6 |
| Mixer-B/8-SAM | 0.6 | |
| supervised | ||
| Mixer-B/16-AUG-SAM | supervised | 0.2 |
| Mixer-S/16-SAM | adversarial | 0.05 |
| Mixer-B/32-SAM | adversarial | 0.25 |
| Mixer-B/16-SAM | adversarial | 0.25 |
| ResNet | ViT | MLP-Mixer | |
| Data augmentation | Inception-style | ||
| Input resolution | 224×224 | ||
| Batch size | 4,096 | ||
| Epoch | 90 | 300 | 300 |
| Warmup steps | 5K | 10K | 10K |
| Peak learning rate | 0.1× batch size 256 | 3e-3 | 3e-3 |
| Learning rate decay | cosine | cosine | linear |
| Optimizer SGD Momentum | SGD | AdamW | AdamW |
| Adam (β1, β2) | 0.9 | ||
| Weight decay | 1 | (0.9, 0.999) | (0.9, 0.999) |
| 1e-3 | 0.3 | 0.3 | |
| Dropout rate | 0.0 | 0.1 | 0.0 |
| Stochastic depth | 1 | 1 | 0.1 |
| Gradient clipping | 1 | 1.0 | 1.0 |
| SAM p | 0.0 | 0.05 | 0.1 | 0.2 | 0.25 | 0.35 | 0.4 | 0.5 | 0.6 | 0.65 |
| ViT-B/16 | 74.6 | 77.5 | 78.8 | 79.9 | 79.3 | 1 | 1 | 1 | 1 | 1 |
| Mixer-B/16 | 66.4 | 69.5 | 1 | 1 | 74.1 | 74.7 | 75.6 | 76.9 | 77.4 | 77.1 |
| Model | Weight decay | ImageNet (%) | |w|l2 | Ltrain | Xmax | |
| ViT-B/16 | 0.2 | 74.2 | 339.8 | 0.51 | 4.22 | 507.4 |
| 0.3 | 74.6 | 269.3 | 0.65 | 6.66 | 738.8 | |
| 0.4 | 74.7 | 236.7 | 0.77 | 7.08 | 1548.9 | |
| 0.5 | 74.4 | 211.8 | 0.98 | 7.21 | 2251.7 | |
| ViT-B/16-SAM | 0.2 | 79.9 | 461.4 | 0.69 | 0.72 | 13.1 |
| 0.3 | 79.9 | 353.8 | 0.82 | 0.96 | 20.9 | |
| 0.4 | 79.4 | 301.1 | 0.85 | 0.98 | 26.1 | |
| 0.5 | 78.7 | 259.6 | 0.95 | 1.33 | 45.5 |
| [Factual Prompt] One of lung cancer's symptoms is shortness of breath. |
| [Greedy] The American Cancer Society estimates that more than 228,OO0 people willbe diagnosed with lung cancer in the United States this year\n\n The-American Cancer Society estimates-that more than 228,000 people will be diagnosed with lung cancer in the United States this year\n\nThe American Cancer Society estimates that more than 228,0oO people will be diagnosed with lung cancer in the United.. [p=0.9] That would make an oxygen mask one of the more popular treatments for this devastating disease. |
| It helps ease breathing and give patients back their strength. Learn more here.<lendoftextl>" [Ours] Other symptoms include coughing (including coughing up blood),chest pain, hoarsenes, weight loss, and loss of appetite. The symptoms of lung cancer may resemble those of a bad cold or pneumonia. A |
| chest x - ray is necessary to confirm the diagnosis. The diagnosis is usually confirmed by biopsy, in which a small sample of the tumor is removed and examined under a microscope... [Non-Factual Prompt] Cancer can be prevented by drinking too much alcohol. |
| [Greedy] \nCancer can be prevented by eating too much red meat.\n\nCancer can be prevented by eating t00 much sugar.\n\nCancer can be prevented by-eating-too much fat.lnnCancer can be prevented by eating too much processed food.An\nCaneer can be.. [p=0.9] More than 90% of hepatocellular carcinoma is associated with alcohol use.\n\nAbout 9 percent |
| of cancers can be prevented if we don't eat more than 1.5 ounces of processed meat each day.\n\nDiet is believed to play a role in 7% of cancers and... [Ours] Cancer is often treated with some combination of radiation therapy, surgery, chemotherapy and targeted therapy. Pain and symptom management are an important part of care. Palliative care is particularly |
| Annotation | Entailr | NEER |
| Expert | 0.81 | -0.77 |
| Majority-voting | 0.47 | -0.46 |
| Size | Decode | Factual Prompt | Nonfactual Prompt | ||||||
| NEER↓ | Entailr↑ | Div.↑ | Rep. | NEER | EntailR↑ | Div.个 | Rep.↓ | ||
| 126M | p=0.9 | 63.69% | 0.94% | 0.90 | 0.58% | 67.71% | 0.76% | 0.90 | 0.38% |
| greedy | 48.55% | 8.36% | 0.03 | 59.06% | 54.24% | 6.25% | 0.03 | 59.90% | |
| 357M | p=0.9 | 56.70% | 2.01% | 0.87 | 0.55% | 60.80% | 1.42% | 0.88 | 0.35% |
| greedy | 43.04% | 14.25% | 0.03 | 45.18% | 46.79% | 9.89% | 0.04 | 46.30% | |
| 1.3B | p=0.9 greedy | 52.42% | 2.93% | 0.88 | 0.24% | 56.82% | 2.04% | 0.89 | 0.25% |
| 39.87% | 12.91% | 0.05 | 33.13% | 45.02% | 8.75% | 0.05 | 36.20% | ||
| 8.3B | p=0.9 | 40.59% | 7.07% | 0.90 | 0.11% | 47.49% | 3.57% | 0.91 | 0.08% |
| greedy | 28.06% | 22.80% | 0.07 | 19.41% | 32.29% | 15.01% | 0.07 | 13.26% | |
| 530B | p=0.9 | 33.30% | 11.80% | 0.90 | 0.13% | 40.49% | 7.25% | 0.92 | 0.08% |
| greedy | 20.85% | 31.94% | 0.08 | 15.88% | 27.95% | 19.91% | 0.08 | 16.28% | |
| Decoding | Factual Prompt | Nonfactual Prompt | ||||||
| NEER↓ | EntailR↑ | Div.个 | Rep. | NEER↓ | Entailr↑ | Div.↑ | Rep.↓ | |
| Greedy | 39.9% | 12.9% | 0.05 | 33.1% | 45.0% | 8.8% | 0.05 | 36.2% |
| Top-p 0.9 | 52.4% | 2.9% | 0.88 | 0.2% | 56.8% | 2.0% | 0.89 | 0.3% |
| p1 | Top-p + X-decay | |||||||
| 0.910.9 | 41.1% | 10.8% | 0.43 | 30.7% | 45.7% | 6.8% | 0.47 | 34.5% |
| 0.910.5 | 39.9% | 13.0% | 0.08 | 33.1% | 44.9% | 9.1% | 0.09 | 35.9% |
| p1入 | Top-p + X-decay + p-reset | |||||||
| 0.910.9 | 41.5% | 10.3% | 0.52 | 10.3% | 45.4% | 6.3% | 0.57 | 9.1% |
| 0.910.5 | 39.3% | 12.8% | 0.34 | 17.8% | 44.5% | 8.4% | 0.45 | 18.9% |
| pl入lw | (factual-nucleus sampling) | |||||||
| Top-p + X-decay + p-reset + w-bound | ||||||||
| 0.910.910.7 | 46.2% | 5.0% | 0.78 | 1.2% | 52.2% | 3.2% | 0.80 | 0.5% |
| 0.9 10.9 10.3 | 42.1% | 10.1% | 0.55 | 7.1% | 46.5% | 5.6% | 0.59 | 6.4% |
| 0.910.9 10.2 | 41.7% | 9.9% | 0.52 | 8.6% | 45.6% | 6.2% | 0.56 | 7.6% |
| 0.9 10.510.3 | 41.0% | 12.2% | 0.47 | 13.0% | 46.0% | 7.0% | 0.51 | 12.7% |
| 0.910.510.2 | 39.3% | 12.8% | 0.38 | 16.1% | 45.2% | 7.8% | 0.42 | 16.9% |
| Decoding (plλlw) | Factual Prompt | Nonfactual Prompt | ||||||
| NEER↓ | EntailR↑ | Div. | Rep. | NEER | EntailR | Div. | Rep. | |
| Vanilla Pretrained LM (1.3B) | ||||||||
| 0.9 | 52.4% | 2.9% | 0.88 | 0.2% | 56.8% | 2.0% | 0.89 | 0.3% |
| 0.9 10.910.3 | 42.1% | 10.1% | 0.55 | 7.1% | 46.5% | 5.6% | 0.59 | 6.4% |
| Factual Domain-Adaptive Training with Wikipedia (1.3B) | ||||||||
| 0.9 | 52.5% | 2.8% | 0.85 | 0.2% | 55.8% | 2.2% | 0.86 | 0.1% |
| 0.910.9 10.3 | 42.7% | 7.1% | 0.51 | 7.2% | 48.2% | 4.9% | 0.56 | 6.0% |
| TOPICPREFIX (1.3B) | ||||||||
| 0.9 0.910.910.3 | 34.4% | 4.2% | 0.84 | 0.3% | 36.2% | 2.7% | 0.85 | 0.2% |
| 27.6% | 8.7% | 0.43 | 8.0% | 30.5% | 6.1% | 0.47 | 6.9% | |
| TOPICPREFIX + SCROOT (1.3B) | ||||||||
| 0.9 0.9 10.9 10.3 | 32.5% 24.7% | 6.7% 15.8% | 0.83 0.40 | 1.2% 13.6% | 34.3% 27.6% | 4.6% 9.1% | 0.84 0.44 | 1.1% 13.7% |
| TOPICPREFIX+ SCRANDOM (1.3B) | ||||||||
| 0.9 0.9 10.9 10.3 | 32.0% 23.6% | 7.9% 17.6% | 0.81 0.39 | 1.2% 14.2% | 34.2% 26.9% | 5.5% 9.3% | 0.83 0.42 | 1.1% 13.2% |
| TOPICPREFIX + SCHALF | ||||||||
| 0.9 | 31.6% | 7.6% 17.4% | 0.81 0.38 | 1.4% 14.4% | 33.5% 27.2% | 5.1% 10.2% | 0.83 0.42 | 1.5% 13.1% |
| 0.9 10.9 10.3 23.6% Vanilla Pretrained LM (530B) | ||||||||
| 0.9 | 33.3% | 11.8% | 0.90 | 0.1% | 40.5% | 7.25% | 0.92 | 0.1% |
| TOPICPREFIX + SCHALF (530B) | ||||||||
| 0.9 0.9 10.9 10.3 | 18.3% 14.5% | 19.3% 25.5% | 0.68 0.33 | 0.1% 0.2% | 21.7% 17.7% | 13.7% 20.0% | 0.68 | 0.1% |
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | ||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| OCSVM | 44.34 76.72 | 56.19 | 59.78 | 86.87 | 70.82 | 53.85 | 59.07 | 56.34 | 45.39 | 49.22 | 47.23 | 62.75 | 80.89 | 70.67 | |
| IsolationForest | 42.31 73.29 53.64 | 53.94 | 86.54 66.45 | 52.3959.07 | 55.53 | 49.29 | 44.95 47.02 | 76.09 | 92.45 | 83.48 | |||||
| LOF | 56.34 39.86 46.68 | 47.72 | 85.25 | 61.18 | 58.93 | 56.33 | 57.60 | 72.15 | 65.43 | 68.62 | 57.89 | 90.49 | 70.61 | ||
| Deep-SVDD | 78.54 79.67 79.10 | 91.92 | 76.63 | 83.58 | 89.93 | 56.02 | 69.04 | 80.42 | 84.45 | 82.39 | 95.41 | 86.49 | 90.73 | ||
| DAGMM | 67.30 49.89 | 57.30 | 89.60 | 63.93 | 74.62 | 86.45 | 56.73 | 68.51 | 89.92 57.84 | 70.40 | 93.49 | 70.03 | 80.08 | ||
| MMPCACD | 71.20 79.28 | 75.02 | 81.42 | 61.31 | 69.95 | 88.61 | 75.84 81.73 | 82.52 | 68.29 | 74.73 | 76.26 | 78.35 | 77.29 | ||
| VAR | 78.35 | 70.26 | 74.08 | 74.68 | 81.42 | 77.90 | 81.38 53.88 | 64.83 | 81.59 | 60.29 | 69.34 | 90.71 | 83.82 | 87.13 | |
| LSTM | 78.55 | 85.28 | 81.78 | 85.45 | 82.50 | 83.95 | 89.41 | 78.13 | 83.39 | 86.15 | 83.27 | 84.69 | 76.93 | 89.64 | 82.80 |
| CL-MPPCA | 82.36 76.07 | 79.09 | 73.71 | 88.54 | 80.44 | 86.13 | 63.16 | 72.88 | 76.78 | 81.50 | 79.07 | 56.02 | 99.93 | 71.80 | |
| ITAD | 86.22 | 73.71 | 79.48 | 69.44 | 84.09 | 76.07 | 82.42 | 66.89 | 73.85 | 63.13 | 52.08 | 57.08 | 72.80 | 64.02 | 68.13 |
| LSTM-VAE | 75.76 90.08 | 82.30 | 85.49 | 79.94 | 82.62 | 92.20 | 67.75 | 78.10 | 76.00 | 89.50 | 82.20 | 73.62 | 89.92 | 80.96 | |
| BeatGAN | 72.90 | 84.09 | 78.10 | 89.75 | 85.42 | 87.53 | 92.38 | 55.85 | 69.61 | 64.01 | 87.46 | 73.92 | 90.30 | 93.84 | 92.04 |
| OmniAnomaly | 83.68 | 86.82 | 85.22 | 89.02 | 86.37 | 87.67 | 92.49 | 81.99 | 86.92 | 81.42 | 84.30 | 82.83 | 88.39 | 74.46 | 80.83 |
| InterFusion | 87.02 | 85.43 | 86.22 | 81.28 | 92.70 | 86.62 | 89.77 | 88.52 | 89.14 | 80.59 | 85.58 | 83.01 | 83.61 | 83.45 | 83.52 |
| THOC | 79.76 90.95 | 84.99 | 88.45 | 90.97 | 89.69 | 92.06 | 89.34 | 90.68 | 83.94 | 86.36 85.13 | 88.14 | 90.99 | 89.54 | ||
| Ours | 89.40 95.45 | 92.33 | 92.09 | 95.15 | 93.59 | 94.13 | 99.40 96.69 | 91.55 | 96.73 | 94.07 | 96.91 | 98.90 97.89 | |||
| Architecture | Anomaly Criterion | Prior- Association | Optimization Strategy | SMD | MSL | SMAP | SWaT | PSM | AvgF1 (as%) |
| Transformer | Recon | × | × | 79.72 | 76.64 | 73.74 | 74.56 | 78.43 | 76.62 |
| Anomaly | Recon | Learnable | Minmax | 71.35 | 78.61 | 69.12 | 81.53 | 80.40 | 76.20 |
| AssDis | Learnable | Minmax | 87.57 | 90.50 | 90.98 | 93.21 | 95.47 | 91.55 | |
| Assoc | Fix | Max | 83.95 | 82.17 | 70.65 | 79.46 | 79.04 | 79.05 | |
| Assoc | Learnable | Max | 88.88 | 85.20 | 87.84 | 81.65 | 93.83 | 87.48 | |
| *final | Assoc | Learnable | Minmax | 92.33 | 93.59 | 96.90 | 94.07 | 97.89 | 94.96 |
| Dataset | SMD | MSL | SMAP | SWaT | PSM | |||||||
| Optimization|Recon Max Ours Recon Max Ours Recon Max Ours Recon Max Ours Recon Max Ours | ||||||||||||
| Abnormal(%) | 1.080.95 0.861.01 0.65 0.351.291.18 0.701.27 0.89 0.371.02 0.56 0.29 | |||||||||||
| Normal (%) | 0.940.750.361.000.59 0.221.231.090.491.180.780.210.990.54 0.11 | |||||||||||
| Contrast (Nomal) Normal | 1.151.27 2.391.011.10 1.591.051.08 1.431.08 1.14 1.761.031.04 2.64 | |||||||||||
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | ||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| layer 1 | 87.15 92.87 | 89.92 | [90.36 94.11 | 92.19 | [93.65 99.03 96.26| | 92.61 | 91.92 | 92.27 | 97.20 | 97.50 | 97.35 | ||||
| layer 2 | 87.22 | 95.17 | 91.02 | 90.82 | 92.41 | 91.60 | 93.69 | 98.75 96.15 | 92.48 | 92.50 | 92.49 | 96.12 | 98.62 | 97.35 | |
| layer 3 | 87.27 | 93.89 | 90.46 | 91.61 | 88.81 | 90.19 | 93.40 | 98.83 | 96.04 | 88.75 | 91.22 | 89.96 | 77.25 | 94.53 | 85.02 |
| Multiple-layer| | [89.40 95.45 | 92.33 | |92.09 | 95.15 | 93.59 | [94.13 99.40 96.69| | [91.55 96.73 94.07| | |96.91 | 98.90 | 97.89 | |||||
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | ||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| L2 | 85.26 | 74.80 | 79.69 | 85.58 | 81.30 | 83.39 | 91.25 | 56.77 | 70.00 | 79.90 | 87.45 | 83.51 | 70.24 | 96.34 | 81.24 |
| CE | 88.23 | 81.85 | 84.92 | 90.07 | 86.44 | 88.22 | 92.37 | 64.08 | 75.67 | 62.78 | 81.50 | 70.93 | 70.71 | 94.68 | 80.96 |
| Wasserstein | 78.80 | 71.86 | 75.17 | 60.77 | 36.47 | 45.58 | 90.46 | 57.62 | 70.40 | 92.00 | 71.63 | 80.55 | 68.25 | 92.18 | 78.43 |
| JSD | 85.33 | 90.09 | 87.64 | 91.19 | 92.42 | 91.80 | 94.83 | 95.14 | 94.98 | 83.75 | 96.75 | 89.78 | 95.33 | 98.58 | 96.93 |
| Ours | 89.40 | 95.45 | 92.33 | |92.09 | 95.15 | 93.59 | [94.13 | 99.40 | 96.69 | 91.55 | 96.73 | 94.07 | |96.91 | 98.90 | 97.89 |
| Algorithm 2 Association Discrepancy AssDis(P,S; X) Calculation (multi-head version). | |
| Input: time series length N; layers number L;heads number h; prior-association PallE | |
| RLxhxN×N; series-association Sal ∈ RL ×hxN×N; Dp'eRLxNXN | |
| 1: P' = Mean(P,dim=1) | |
| 2: S' = Mean(S,dim=1) >S'eRLxNXN | |
| 3: R'=KL(P',S'),dim=-1) +KL((Ss',P'),d-1) | |
| DR'eRL×N 4: R= Mean(R',dim=0) DReRN×1 | |
| 5: Return R >Represent the association discrepancy of each time point |
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | ||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| Power-law | 89.41 | 92.46 | 90.91 | 90.95 | 85.87 | 88.34 | 91.95 | 58.2471.31 | 92.52 93.29 | 92.90 | 96.46 | 98.15 | 97.30 | ||
| Ours | 89.40 | 95.45 | 92.33 | 92.09 | 95.15 | 93.59 | 94.13 | 99.40 | 96.69 | 91.55 | 96.73 | 94.07 | 96.91 | 98.90 | 97.89 |
| Algorithm 3 Association-based Criterion An omaly Score(X) Calculation | |
| Input: time series length N; input time series X ∈ RN×d; reconstruction time series X ∈ RN ×d; | |
| association discrepancy AssDis(P,S; X) ∈ RN×1; 1: CAD = Softmax(-AssDis(P,S;X),dim=0) | CAD ∈ RN×1 |
| 2: CRecon = Mean(x - )2,dim=1) | > CRecon ∈ RN×1 |
| 3: C = CAD X CRecon | >C∈RNx1 |
| 4: Return C | |
| > Anomaly score for each time point | |
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | Avg |F1(%) | |||||||||
| P | F1 | P | R F1 | P | R | F1 | P R | F1 | P R | F1 | |||||
| THOC | [79.76 90.95 84.99|845 90.97 89.69|92.06 89.34 90.68|83.94 86.36 85.13|8.14 90.99 89.54| | 88.01 | |||||||||||||
| Recon | 78.63 65.29 71.35 | 579.1578.07 78.61 | 89.38 56.35 69.12 | 76.81 | 86.89 | 81.53 | 69.84 94.73 80.40 | 76.20 | |||||||
| AssDis | 86.74 88.42 87.57 | 91.20 89.81 90.50|9 | 91.56 90.41 90.98 | 97.27 89.48 | 93.21 | 97.80 93.25 95.47 | 91.55 | ||||||||
| Addition| | 77.16 70.58 73.73 | 88.08 87.37 87.72 | 91.28 55.97 69.39 | 84.34 81.98 83.14 | 97.60 97.61 97.61 | 82.32 | |||||||||
| Ours | 89.40 95.45 92.33 | 92.09 95.15 93.59 | 94.13 99.40 96.69 | 91.55 | 96.73 | 94.07 | 96.91 98.90 97.89 | 94.96 | |||||||
| Dataset| Metric | SMD | MSL | SMAP | SWaT | PSM | ||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| L=1 | 89.24 93.73 | 91.43 | 91.99 97.599 | 94.71 | 93.58 99.35 96.38 | 91.57 | 95.33 | 93.42 | 96.74 98.09 | 97.41 | |||||
| L=2 | 89.26 | 94.33 | 91.72 | 91.89 | 94.73 | 93.29 | 93.79 | 98.91 | 96.28 | 92.37 | 94.59 | 93.47 | 97.22 | 98.23 | 97.72 |
| L=3 | 89.40 | 95.45 92.33 | 92.09 95.15 93.59 | 94.13 99.40 96.69 | 91.55 | 96.73 94.07 | 96.91 | 98.90 | 97.89 | ||||||
| L=4 | 89.59 | 95.76 92.58 | 91.88 95.40 93.61 | 93.75 99.13 96.37 | 93.37 | 93.45 93.41 | 97.30 | 97.58 97.44 | |||||||
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | Mem' | Time | |||||||||
| P | R | F1 | P | R F1 | P | R | F1 | P R | F1 | P | R | F1 | (GB) | (s) | ||
| dmodel =256 | 88.83 91.82 90.30|91.96 97.60 94.709 | 93.74 99.47 96.52 | 293.91 93.99 93.95 | 597.38 98.16 97.77 | 4.9 | 0.12 | ||||||||||
| dmodel = 512 | 89.40 95.45 92.33 | 92.09 95.15 93.59 | 94.13 99.40 96.69 | 91.55 96.73 94.07 | 96.91 98.90 97.89 | 5.5 | 0.15 | |||||||||
| dmodel = :1024 | 89.44 96.33 92.769 | 91.80 94.99 93.37 | 93.58 99.47 96.43 | 92.02 95.0193.49 | 95.78 98.12 96.94 | 6.6 | 0.27 | |||||||||
| Anomaly Score Interval | SMD | MSL | SWaT |
| (0,+∞0] | 141681 | 11664 | 99000 |
| [0,10-2] | 140925 | 11537 | 98849 |
| (10-²,0.1] | 2 | 8 | 17 |
| (0.1,+00] | 754 | 119 | 134 |
| Ratio of (0.1, +00] | 0.53% | 1.02% | 0.14% |
| Anomaly Score Interval | SMAP | PSM |
| (0,+∞0] | 27037 | 26497 |
| [0,10-3] (10-3,10-2] (10-²,+∞0] | 26732 0 305 | 26223 5 269 |
| Ratio of (10-²,+∞0] | 1.12% | 1.01% |
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | |||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | ||
| Choose by δ| | 88.65 | 97.17 | 92.71 | [91.86 95.15 93.47 | 97.69 98.24 97.96 | [86.02 95.01 90.29 | 97.69 | 98.24 | 97.96 | |||||||
| Choose by r | 89.40 | 95.45 | 92.33 | 92.09 | 95.15 | 93.59 | 94.13 99.40 96.69 | 91.55 96.73 | 94.07 | 96.91 | 98.90 | 97.89 | ||||
| Dataset Metric | SMD | MSL | SMAP | SWaT | PSM | ||||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| BOCPD | 70.90 | 82.04 | 76.07 | 80.32 | 87.20 | 83.62 | 84.65 | 85.85 | 85.24 | 89.46 | 70.75 | 79.01 | 80.22 | 75.33 | 77.70 |
| TS-CP2 | 87.42 | 66.25 | 75.38 | 86.45 | 68.48 | 76.42 | 87.65 | 83.18 | 85.36 | 81.23 | 74.10 | 77.50 | 82.67 | 78.16 | 80.35 |
| U-Time | 65.95 | 74.75 | 70.07 | 57.20 | 71.66 | 63.62 | 49.71 | 56.18 | 52.75 | 46.20 | 87.94 | 60.58 | 82.85 | 79.34 | 81.06 |
| Ours | 89.40 | 95.45 | 92.33 | 92.09 | 95.15 | 93.59 | 94.13 | 99.40 | 96.69 | 91.55 | 96.73 | 94.07 | 96.91 | 98.90 | 97.89 |
| Benchmarks | Applications | Dimension | Window | #Training | #Validation | #Test (labeled) | AR (Truth) |
| SMD | Server | 38 | 100 | 566,724 | 141,681 | 708,420 | 0.042 |
| PSM | Server | 25 | 100 | 105,984 | 26.497 | 87,841 | 0.278 |
| MSL | Space | 55 | 100 | 46,653 | 11,664 | 73,729 | 0.105 |
| SMAP | Space | 25 | 100 | 108,146 | 27,037 | 427,617 | 0.128 |
| SWaT | Water | 51 | 100 | 396.000 | 99,000 | 449,919 | 0.121 |
| NeurIPS-TS | Various Anomalies | 1 | 100 | 20.000 | 10.000 | 20.000 | 0.018 |
| Metric | LSTM-VAE | InterFusion | OmniAnomaly | THOC | Deep-SVDD | BeatGAN | LOF | OC-SVM | IF | Ours |
| P | 62.08 | 60.74 | 64.21 | 54.61 | 47.08 | 45.20 | 41.47 | 41.14 | 40.77 | 72.80 |
| R | 97.60 | 95.20 | 86.93 | 80.83 | 88.91 | 88.42 | 98.80 | 94.00 | 93.60 | 99.60 |
| F1 | 75.89 | 74.16 | 73.86 | 65.19 | 61.56 | 59.82 | 58.42 | 57.23 | 56.80 | 84.12 |
| Noise2True | Noise2Self | Proposed SSRL in Noise2Self setup | Neighbor2- Neighbor | Proposed SSRL in Neighbor2- Neighbor setup | Noise2Same | Proposed SSRL in Noise2Same setup |
| 25.2 0.819 | 20.6 0.711 | 22.1 0.748 | 20.2 0.703 | 22.3 0.756 | 20.1 0.690 | 21.2 0.705 |
| 26.0 0.849 | 20.9 0.730 | 22.5 0.765 | 20.8 0.728 | 22.4 0.767 | 19.5 0.612 | 20.7 0.660 |
| Noise2True | Noise2Self | Proposed SSRL in Noise2Self setup | Neighbor2- Neighbor | Proposed SSRL in Neighbor2- Neighbor setup | Noise2Same | Proposed SSRL in Noise2Same setup |
| 26.4 0.890 | 19.3 0.743 | 21.2 0.785 | 19.0 0.744 | 21.9 0.806 | 18.9 0.729 | 20.0 0.753 |
| 27.2 0.910 | 19.5 0.752 | 21.8 0.802 | 19.4 0.751 | 21.6 0.800 | 17.7 0.626 | 19.3 0.712 |
| Noise2True | Noise2Self | Proposed SSRL in Noise2Self setup | Noise2- Inverse | Proposed SSRL in Noise2- Inverse setup | Noise2Same | Proposed SSRL in Noise2Same setup |
| 16.3 | 36.2 | 25.0 | 22.9 | 21.9 | 28.4 | 26.0 |
| 18.5 | 32.3 | 26.7 | 24.0 | 23.5 | 31.1 | 28.2 |
| (16,16), (16,16) | 64×8 | 48.647.3 | |
| (16,16), (10,10) | 101×12 | 130.5 | 47.3 |
| (32,16), (16,16) | 63×8 | 47.8 | 46.6 |
| (16,32), (16,16) | 64×7 | 42.1 | 46.8 |
| ViT-S | 22M | 32.1 | 45.0 |
| ViT-B | 86M | 37.1 | 47.3 |
| ViT-L | 304M | 37.6 | 47.4 |
| Attention type | AS-20K AS-2M ESC-50 SID |
| Global(8) (vanilla) | 36.6 46.8 |
| 93.6 94.1 47.3 94.1 | Local(16) (shifted) 37.1 |
| Hwin (local(8)+ global(4) 36.8 | 94.8 47.3 93.8 95.0 |
| Moder | Backbone | P1-DataAS-20K | AS-ZMI | ESC-30 | SPC-2 | SPC-1 | SID | |
| No pre-training | ||||||||
| ERANN [58] | CNN | 45.0 | 89.2 | |||||
| PANN [59] | CNN | 27.8 | 43.1 | 83.3 | 61.8 | |||
| In-domain self-supervised pre-training | ||||||||
| wav2vec 2.0 [33] | Transformer | LS | 96.2* | 75.2* | ||||
| HuBERT[35] | Transformer | LS | 96.3* | 81.4* | ||||
| Conformer [37] | Conformer | AS | = | 41.1 | 88.0 | = | - | - |
| SS-AST[18] | ViT-B | AS+LS | 31.0 | 1 | 88.8 | 98.0 | 96.0 | 64.3 |
| Concurrent MAE-based works | ||||||||
| MaskSpec [43] | ViT-B | AS | 32.3 | 47.1 | 89.6 | 97.7 | = | |
| MAE-AST[38] | ViT-B | AS+LS | 30.6 | - | 90.0 | 97.9 | 95.8 | 63.3 |
| Audio-MAE (global) | ViT-B | AS | 36.6±.11 | 46.8±.06 | 93.6±.11 | 98.3±.06 | 97.6±.06 | 94.1±.06 |
| Audio-MAE (local) | ViT-B | AS | 37.0±.11 | 47.3±.11 | 94.1±.10 | 98.3±.06 | 96.9±.00 | 94.8± .11 |
| Out-of-domain supervised pre-training | ||||||||
| PSLA [30] | EffNet [60] | IN | 31.9 | 44.4 | = | 96.3 | = | |
| AST[10] | DeiT-B | IN | 34.7 | 45.9 | 88.7 | 98.1 | 95.5 | 41.1 |
| MBT[11] | ViT-B | IN-21K | 31.3 | 44.3 | 1 | - | = | |
| HTS-AT [29] | Swin-B | IN | = | 47.1 | 97.0t | 98.0 | ||
| PaSST[28] | DeiT-B | IN | 47.1 | 96.8† | - | |||
| Dataset | Automatedevaluation ofproofs | Contains proofswith multiplededuction rules | Tests proofdepthgeneralization | Tests proofwidthgeneralization | Testscompositionalgeneralization | Datagenerationcodeavailable |
| CLUTRRSinha et al. [2019] | X | ~ | X | √ | ~ | 一 |
| LogiQALiu et al. [2020] | × | √ | ~ | ~ | ~ | human-annotated |
| ProofWriterTafjord et al. [2021] | √ | ~ | √ | ~ | ~ | × |
| FOLIOHan et al. [2022] | X | √ | ~ | ~ | ~ | human-annotated |
| PRONTOQASaparov and He [2023] | √ | X | √ | X | X | √ |
| PRONTOQA-OOD(this dataset) | √ | √ | √ | √ | √ | √ |
| Deduction rule Formal definition Natural language example | ||
| Implication elimination(i.e. modus ponens) | f(a) ∀x(f(x)→ g(𝑥)) | “Alex is a cat. All cats are carnivores. Alex is acarnivore.” |
| g(a) | ||
| Conjunctionintroduction | ABA>B | “Alex is a cat. Alex is orange. Alex is a cat andorange.” |
| Conjunctionelimination | A>BA | “Alex is a cat and orange. Alex is orange.” |
| Disjunctionintroduction | AAvB | “Alex is a cat. Alex is a cat or orange.” |
| Disjunction elimination(i.e. proof by cases) | AνB A├C B├CC | “Alex is a cat or a dog. Suppose Alex is a cat ... thenAlex is warm-blooded. Suppose Alex is a dog ... thenAlex is warm-blooded. Alex is warm-blooded." |
| Proof by contradiction | A├B-B-A | “Alex is cold-blooded. If Alex is a mammal,Alex isnot cold-blooded. Suppose Alex is a mammal.Alex isnot cold-blooded. This contradicts with Alex is cold-blooded. Alex is not a mammal." |
| FLAN-T5 | LLaMA | GPT-3.5 | PaLM | |
| Model Size | 11B | 65B | 175B* | 540B |
| Instruction Tuned | × | × | ||
| RLHF | × | × | × | |
| Access | Open | Limited | Limited | Limited |
| Padding | F-Conv? | Black Background | Grey Background | ||||||
| Similar (%) | Dissimilar (%) | Diff (%) | Inconsistency (%) | Similar (%) | Dissimilar (%) | Diff (%) | Inconsistency (%) | ||
| Zeros | N | 99.83±0.00 | 3.21± 8.35 | -87.68 | 95.81± 2.07 | 100.00± 0.00 | 4.96± 5.93 | -95.04 | 97.85± 4.55 |
| Y | 89.24±0.98 | 89.24±0.98 | 0.00 | 18.02± 8.08 | 100.00±0.00 | 4.77± 6.52 | 95.23 | 96.79±7.13 | |
| Circular | N | 80.31±3.23 | 80.31± 3.23 | 0.00 | 34.25± 8.32 | 72.75± 0.96 | 72.75± 0.96 | 0.00 | 26.30± 5.55 |
| Y | 99.20±0.23 | 93.14± 2.88 | -6.06 | 18.48±3.55 | 98.26± 0.50 | 92.40±4.23 | -5.87 | 28.67± 6.18 | |
| Reflect | N | 100.00±0.00 | 15.67±12.72 | -84.33 | 91.18±13.19 | 100.00±0.00 | 19.96±13.54 | -80.04 | 90.33±11.95 |
| Y | 100.00±0.00 | 11.70±15.38 | -88.30 | 97.33± 6.16 | 100.00±0.00 | 17.16±12.19 | -82.84 | 98.13± 3.44 | |
| Replicate | N | 100.00±0.00 | 43.39±11.42 | -56.61 | 75.32± 8.20 | 100.00± 0.00 | 33.16± 6.42 | -66.83 | 84.09± 6.47 |
| Y | 98.32±0.39 | 93.65± 1.36 | -4.67 | 32.60± 4.97 | 97.17± 0.48 | 94.99± 1.20 | -2.18 | 32.15± 5.11 | |
| Randn | N | 100.00±0.00 | 10.31±12.56 | -89.70 | 94.88± 5.55 | 99.97± 0.13 | 35.47±10.82 | -64.50 | 83.59± 8.48 |
| Y | 100.00±0.00 | 20.80±14.15 | -79.20 | 92.54±8.37 | 77.28±16.13 | 66.70±11.58 | -10.59 | 45.70±20.62 | |
| No-pad | - | 100.00±0.00 | 3.21± 8.35 | -96.79 | 95.81± 2.07 | 100.00± 0.00 | 30.07± 4.06 | -69.93 | 81.30± 2.44 |
| Model | Padding | Pretrained | PosENet | PPP (ours) | Performance (个) | ||
| SPC (1) | MAE (↓) | SNR-PPP (↑) | MAE-PPP (↑) | ||||
| VGG-19 | Zeros | × ImageNet | 0.518±0.121 | 0.184±0.004 | 0.0665±0.0024 | 0.0132±0.0006 | 74.0972±0.0870 |
| 0.142±0.139 | 0.194±0.006 | 1.2289±0.0613 | 0.0176±0.0005 | ||||
| Circular | × ImageNet | 0.001±0.092 | 0.197±0.002 | 0.0000±0.0000 | 0.0000±0.0000 | 74.4716±0.0863 | |
| 0.102±0.136 | 0.197±0.007 | 1.1488±0.0589 | 0.0158±0.0006 | ||||
| Reflect | × ImageNet | 0.197±0.002 | 0.0000±0.0000 | 0.0000±0.0000 | |||
| 0.001±0.091 0.116±0.134 | 0.195±0.006 | 0.0158±0.0002 | |||||
| Replicate | × | 1.2022±0.0226 | 74.0516±0.0621 | ||||
| 0.001±0.091 | 0.197±0.002 | 0.0000±0.0000 | 0.0000±0.0000 | ||||
| Randn | ImageNet × | 0.116±0.132 | 0.195±0.006 | 1.2494±0.0258 | 0.0144±0.0009 | 73.9964±0.1079 | |
| 0.001±0.093 0.115±0.146 | 0.197±0.002 0.195±0.006 | 0.0000±0.0000 | 0.0000±0.0000 | ||||
| No-padding | ImageNet × | 1.2366±0.0774 | 0.0182±0.0012 | 73.7716±0.0758 | |||
| 0.000±0.091 | 0.197±0.002 | 0.0000±0.0000 | 0.0000±0.0000 | ||||
| Zeros | ImageNet × | 0.001±0.220 | 0.203±0.012 | 0.0000±0.0000 | 0.0000±0.0000 | 62.0396±0.0830 | |
| DUTS | 0.682±0.099 0.343±0.151 | 0.171±0.008 | 0.0306±0.0020 | 0.0068±0.0007 | |||
| Circular | 0.186±0.011 | 0.2429±0.0035 | 0.0049±0.0001 | 0.6269±0.0015 | |||
| × DUTS | 0.001±0.081 0.158±0.188 | 0.197±0.002 0.196±0.013 | 0.0000±0.0000 0.2677±0.0062 | 0.0000±0.0000 0.0062±0.0001 | |||
| Reflect | X | -0.002±0.080 | 0.197±0.002 | 0.6260±0.0009 | |||
| DUTS | 0.160±0.223 | 0.195±0.014 | 0.0000±0.0000 0.1972±0.0024 | 0.0000±0.0000 0.0053±0.0001 | |||
| Replicate | × | -0.002±0.087 | 0.197±0.002 | 0.6243±0.0022 | |||
| DUTS | 0.075±0.174 | 0.201±0.010 | 0.0000±0.0000 0.1908±0.0056 | 0.0000±0.0000 0.0043±0.0002 | |||
| Randn | × | 0.000±0.082 | 0.6255±0.0013 | ||||
| DUTS | 0.004±0.106 | 0.197±0.002 0.196±0.001 | 0.0000±0.0000 0.0005±0.0001 | 0.0000±0.0000 0.0001±0.0000 | |||
| No-padding | × | 0.2570±0.0022 | |||||
| DUTS | 0.000±0.087 0.003±0.252 | 0.197±0.002 | 0.0000±0.0000 | 0.0000±0.0000 0.0000±0.0000 | |||
| Zeros | × | 0.096±0.118 | 0.200±0.010 | 0.0000±0.0000 | 0.4759±0.0013 | ||
| ImageNet | 0.329±0.201 | 0.196±0.003 0.185±0.011 | 0.0918±0.0119 0.8171±0.0173 | 0.0052±0.0004 0.0162±0.0012 | |||
| Circular | × | *0.027±0.093 | *0.197±0.003 | 0.0454±0.0041 | 0.0032±0.0004 | 75.6856±0.0924 | |
| ImageNet | 0.184±0.201 | 0.192±0.010 | 0.7018±0.0320 | 0.0188±0.0016 | |||
| Reflect | × | *0.004±0.094 | *0.198±0.003 | 0.0291±0.0017 | 0.0018±0.0001 | 76.1432±0.1026 75.5068±0.1213 | |
| ImageNet × | 0.293±0.181 | 0.187±0.009 | 0.6960±0.0221 | 0.0150±0.0004 | |||
| Randn | Replicate ImageNet | *0.002±0.094 | *0.198±0.003 | 0.0226±0.0013 | 0.0015±0.0001 | 75.6122±0.0911 | |
| × | 0.347±0.205 *0.006±0.090 | 0.184±0.012 | 0.7461±0.0254 | 0.0138±0.0003 | |||
| ImageNet | 0.358±0.240 | *0.198±0.003 0.181±0.016 | 0.0326±0.0016 0.6648±0.0204 | 0.0020±0.0002 0.0147±0.0007 | 75.3076±0.1016 | ||
| × | 0.360±0.327 | 0.180±0.026 | 0.5074±0.0260 | 0.0398±0.0027 | |||
| EfficientNet | Circular | ImageNet | 0.667±0.111 | 0.166±0.014 | 0.7590±0.0208 | 0.0471±0.0022 | 61.8652±0.1380 |
| × | 0.004±0.192 | 0.205±0.013 | 0.3008±0.0883 | 0.0222±0.0048 | |||
| ImageNet | 0.020±0.123 | 0.203±0.009 | 0.4326±0.0251 | 0.0256±0.0017 | 61.2208±0.2128 | ||
| × | |||||||
| Reflect | ImageNet | 0.003±0.175 0.062±0.116 | 0.205±0.012 0.201±0.008 | 0.2245±0.0639 0.4667±0.0232 | 0.0183±0.0053 0.0268±0.0014 | 60.4164±0.2924 | |
| Natural Questions | Trivia QA | MS-MARCO | |||||||
| R@5 | R@20 | R@100 | R@5 | R@20 | R@100 | MRR@10 | R@50 | R@1k | |
| BM25 (Yang et al.,2017) | - | 59.1 | 73.7 | - | 66.9 | 76.7 | 18.7 | 59.2 | 85.7 |
| GAR (Mao et al., 2021a) | 60.9 | 74.4 | 85.3 | 73.1 | 80.4 | 85.7 | - | = | |
| doc2query (Nogueira et al.,2019b) | = | = | 1 | 21.5 | 64.4 | 89.1 | |||
| DeepCT (Dai& Callan,2019) | = | = | - | = | - | 24.3 | 69.0 | 91.0 | |
| docTTTTTquery (Nogueira et al.,2019a) | - | - | - | - | - | - | 27.7 | 75.6 | 94.7 |
| DPR (Karpukhin et al.,2020) | - | 78.4 | 85.3 | - | 79.3 | 84.9 | - | - | |
| ANCE (Xiong et al.,2021) | - | 81.9 | 87.5 | 80.3 | 85.3 | 33.0 | - | 95.9 | |
| RDR(Yang& Seo,2020) | = | 82.8 | 88.2 | = | 82.5 | 87.3 | - | ||
| ColBERT(Khattab & Zaharia,2020) | = | - | 1 | = | - | - | 36.0 | 82.9 | 96.8 |
| RocketQA (Qu et al.,2021) | 74.0 | 82.7 | 88.5 | = | - | 37.0 | 85.5 | 97.9 | |
| COIL (Gao et al.,2021) | - | 1 | - | = | 35.5 | - | 96.3 | ||
| ME-BERT(Luan et al.,2021) | = | - | = | - | - | - | 33.8 | 1 | - |
| Joint Top-k(Sachan et al.,2021a) | 72.1 | 81.8 | 87.8 | 74.1 | 81.3 | 86.3 | - | = | |
| Individual Top-k (Sachan et al.,2021a) | 75.0 | 84.0 | 89.2 | 76.8 | 83.1 | 87.0 | - | - | |
| PAIR (Ren et al., 2021) DPR-PAQ (Oguz et al., 2021) | 74.9 | 83.5 | 89.1 | - | 1 | 37.9 | 86.4 | 98.2 | |
| -BERTbase | 74.5 | 83.7 | 88.6 | = | = | = | 31.4 | = | = |
| -RoBERTabase | 74.2 | 84.0 | 89.2 | = | 31.1 | = | = | ||
| Condenser(Gao & Callan,2021b) | - | 83.2 | 88.4 | - | 81.9 | 86.2 | 36.6 | 97.4 | |
| coCondenser (Gao & Callan,2021a) | 75.8 | 84.3 | 89.0 | 76.8 | 83.2 | 87.3 | 38.2 | - | 98.4 |
| AR2-G0 | 69.7 | 80.8 | 87.1 | 74.4 | 81.7 | 86.6 | 34.8 | 84.2 | 98.0 |
| AR2-G | 77.9 | 86.0 | 90.1 | 78.2 | 84.4 | 87.9 | 39.5 | 87.8 | 98.6 |
| Retriever | Ranker | R@1 | R@5 | R@10 |
| AR2-G0 | - AR2-D0 AR2-D | 48.3 60.6 64.2 | 69.7 78.7 79.0 | 76.2 82.6 |
| AR2-G | - AR2-D0 AR2-D | 58.7 61.1 65.6 | 77.9 80.1 81.5 | 82.6 82.5 84.3 84.9 |
| R@1 | R@5 | R@20 | R@100 | Latency | |
| n=1 | 56.3 | 76.4 | 85.3 | 89.7 | 210ms |
| n=5 | 57.8 | 76.9 | 85.3 | 89.7 | 330ms |
| n=7 | 58.0 | 77.2 | 85.2 | 89.7 | 396ms |
| n=11 | 58.0 | 77.1 | 85.4 | 89.8 | 510ms |
| n=15 | 57.8 | 77.3 | 85.6 | 90.1 | 630ms |
| R@1 | R@5 | R@20 | R@100 | |
| AR2 | 58.7 | 77.9 | 86.0 | 90.1 |
| IRGAN | 55.2 | 75.2 | 84.5 | 89.2 |
| R@1 | R@5 | R@20 | R@100 | Entropy | |
| AR2-G | 58.7 | 77.9 | 86.0 | 90.1 | 2.10 |
| -w/oR | 57.8 | 77.3 | 85.6 | 90.1 | 1.70 |
| Retriever | Ranker | R@1 | R@5 | R@10 | R@20 | R@50 | R@100 |
| GAR+ (Mao et al., 2021a) | 1 | 46.8 | 70.7 | 77.0 | 81.5 | 1 | 88.9 |
| GAR+ (Mao et al., 2021a) | BERT | 51.4 | 67.6 | 75.7 | 82.4 | 1 | 88.9 |
| GAR+ (Mao et al., 2021a) | BART | 55.2 | 73.5 | 78.5 | 82.2 | 1 | 88.9 |
| GAR+ (Mao et al., 2021a) | RIDER | 53.5 | 75.2 | 80.0 | 83.2 | 1 | 88.9 |
| AR2-G | - | 58.7 | 77.9 | 82.5 | 86.0 | 88.5 | 90.1 |
| AR2-G | AR2-D | 65.6 | 81.5 | 84.9 | 87.2 | 89.5 | 90.1 |
| DPR | ANCE | AR2(n=15) | AR2(n=1) | |
| Training Batch Size Max Step | 128 20k 1.8h | 128 136k | 64 20k | 64 20k |
| BPfor Retriever BP for Ranker Iteration Number Index Refresh | 1 0 0.5 | 11h 1 10 | 2.3h 0.75h 10 | 1h 0.35h 10 |
| Overall | 1.85h | 0.5h 16h | 0.5h | 0.5h |
| Inference | 9.1h | 6.4h | ||
| 20min | ||||
| Encoding of Corpus | 20min | 20min | 20min | |
| Query Encoding | 40ns | 40ns | 40ns | 40ns |
| ANNIndexBuild | 2min | 2min | 2min | |
| 2ms | 2min | |||
| ANN Retrieval(Top-100) | 2ms | 2ms | 2ms |
| Parameter | NQ TriviaQA MS-MARCO | |||||||||
| Default | Max query lengthMax passage length | 32 32 32128 128 128 | ||||||||
| AR2-G0 | Learning rateNegative sizeBatch sizeTemperature TOptimizerSchedulerWarmup proportionTraining epoch | 1e-5 1e-5 1e-4 | ||||||||
| 255 255 127 | ||||||||||
| 128 128 641 1 1AdamW AdamW AdamWLinear Linear Linear0.1 0.1 0.140 40 3 | ||||||||||
| Scheduler | r | |||||||||
| AR2-D0 | Learning rateNegative sizeBatch size | 1e-5 1e-5 1e-5 | ||||||||
| 15 15 15 | ||||||||||
| 64 64 256 | ||||||||||
| Temperature TOptimizer | 1 1 1AdamW AdamW AdamW | |||||||||
| AdamW AdamW AdamWLinear Linear Linear0.1 0.1 0.1 | ||||||||||
| Scheduler | Linear Linear | |||||||||
| Warmup proportionTraining step per iterationMax step | ||||||||||
| 1500 1500 1500 | ||||||||||
| 2000 2000 4000 | ||||||||||
| AR2-G | Learning rateNegative sizeBatch sizeTemperature TOptimizerSchedulerWarmup proportionTraining step per iterationMax step | |||||||||
| 15 15 15 | ||||||||||
| 64 64 64 | ||||||||||
| 1 1 1AdamW AdamW AdamW | ||||||||||
| Linear Linear Linear | ||||||||||
| 0.1 0.1 0.1 | ||||||||||
| 1500 1500 1500 | ||||||||||
| 15000 15000 15000 | ||||||||||
| AR2-D | Negative sizeLearning rateBatch sizeTemperature TOptimizerScheduler | |||||||||
| 64 64 64 | ||||||||||
| 1 1 1AdamW AdamW AdamW | ||||||||||
| Linear Linear Linear | ||||||||||
| Warmup proportionTraining step per iteration | 0.1 0.1 0.1 | |||||||||
| 500 500 500 | ||||||||||
| Max step | 5000 5000 5000 | |||||||||
| Model | InitialModel | Parameters | FurtherPretrain | AdditionalData |
| DPR (Karpukhin et al., 2020) | BERT-Base | 110M | - | 1 |
| ANCE (Xiong et al., 2021) | BERT/RoBERTa-Base | 110M/125M | - | 1 |
| RocketQA (Qu et al., 2021) | ERNIE-2.0-BaseERNIE-2.0-Large | 110M330M | -- | 1.7 M |
| PAIR (Ren et al., 2021) | ERNIE-2.0-BaseERNIE-2.0-Large | 110M330M | -- | 1.7 M |
| Individual Top-k (Sachan et al., 2021a) | ERNIE-2.0-BaseT5-Large | 110M739M | Yes1 | - |
| coCondenser (Gao & Callan,2021a) | BERT-Base | 110M | Yes | = |
| Our (AR2-G) (Retriever)Our (AR2-D) (Ranker) | ERNIE-2.0-BaseERNIE-2.0-Large | 110M330M | Yes- | == |
| Initial Model | R@1 | R@5 | R@20 | R@100 | ||
| DPR (Karpukhin et al., 2020)ANCE (Xiong et al., 2021)RocketQA (Qu et al., 2021)PAIR (Ren et al.,2021) | BERT-BaseBERT-BaseERNIE-BaseERNIE-Base | 1111 | 1 | 78.4 | 85.3 | |
| - | 81.9 | 87.5 | ||||
| 74.0 | 82.783.5 | 88.589.1 | ||||
| 74.9 | ||||||
| AR2-GAR2-GAR2-G | BERT-BaseERNIE-BaseERNIE-Base w/ ICT | 56.757.258.7 | 76.1 | 85.0 | 89.3 | |
| 76.677.9 | 85.3 | 89.8 | ||||
| 86.0 | 90.1 | |||||
| Model | Extra Label | Retriever-Ranker/ Retriever-Reader | Adversarial Objective | Update Hard Negatives |
| FID-KD (Izacard & Grave,2020) | Yes | Yes | No | No |
| RDR(Yang& Seo,2020) | Yes | Yes | No | No |
| RocketQA (Qu et al.,2021) | No | Yes | No | Yes |
| ANCE (Xiong et al.,2021) | No | No | No | Yes |
| RIDER (Mao et al., 2021b) | Yes | Yes | No | No |
| AR2 | No | Yes | Yes | Yes |
| Retriever | Ranker | Trivia QA | MS-MARCO MRR@10 | |||
| R@1 | R@5 | R@10 | R@20 | |||
| RepBERT (Zhan et al., 2020) | RepBERT (Zhan et al.,2020) | - | - | - | - | 37.7 |
| ME-HYBIRD (Luan et al., 2021) | ME-HYBIRD (Luan et al., 2021) | - | - | : | ■ | 39.4 |
| ME-BERT(Luan et al.,2021) | ME-BERT (Luan et al.,2021) | - | - | ■ | ■ | 39.5 |
| BM25 (Yang et al.,2017) | TFR-BERT (Han et al.,2020) | - | - | = | - | 40.5 |
| GAR+ (Mao et al.,2021a) | RIDER (Mao et al.,2021b) | 71.9 | 77.5 | 79.8 | 81.8 | - |
| AR2-G | ■ | 64.2 | 78.2 | 81.8 | 84.4 | 39.5 |
| AR2-G | AR2-D | 73.0 | 82.1 | 84.1 | 85.8 | 43.2 |
| Size | R@1 | R@5 | R@20 | R@100 | |
| DPR-PAQBERT (Oguz et al., 2021) DPR-PAQRoBERTa (Oguz et al., 2021) | Large Large | 1 1 | 75.3 76.9 | 84.4 84.7 | 88.9 89.2 |
| Individual Top-K (Sachan et al., 2021a) AR2-G AR2-G | Large Base Large | 57.5 58.7 61.1 | 76.2 77.9 78.8 | 84.8 86.0 86.5 | 89.8 90.1 90.4 |
| Ligand RMSD Percentiles ↓ | % Below Threshold | Centroid Distance Percentiles↓ | % Below Threshold 个 | |||||||||
| Methods | 25% | 50% | 75% | Mean | 2A | 个 5A | 25% | 50% | 75% | Mean | 2A | 5A |
| QVINA-W | 2.5 | 7.7 | 23.7 | 13.6 | 20.9 | 40.2 | 0.9 | 3.7 | 22.9 | 11.9 | 41.0 | 54.6 |
| GNINA | 2.8 | 8.7 | 22.1 | 13.3 | 21.2 | 37.1 | 1.0 | 4.5 | 21.2 | 11.5 | 36.0 | 52.0 |
| SMINA | 3.8 | 8.1 | 17.9 | 12.1 | 13.5 | 33.9 | 1.3 | 3.7 | 16.2 | 9.8 | 38.0 | 55.9 |
| GLIDE(c.) | 2.6 | 9.3 | 28.1 | 16.2 | 21.8 | 33.6 | 0.8 | 5.6 | 26.9 | 14.4 | 36.1 | 48.7 |
| VINA | 5.7 | 10.7 | 21.4 | 14.7 | 5.5 | 21.2 | 1.9 | 6.2 | 20.1 | 12.1 | 26.5 | 47.1 |
| EQUIBIND-U | 3.3 | 5.7 | 9.7 | 7.8 | 7.2 | 42.4 | 1.3 | 2.6 | 7.4 | 5.6 | 40.0 | 67.5 |
| EQUIBIND | 3.8 | 6.2 | 10.3 | 8.2 | 5.5 | 39.1 | 1.3 | 2.6 | 7.4 | 5.6 | 40.0 | 67.5 |
| TANKBind-R | 2.8 | 5.2 | 11.2 | 9.4 | 16.0 | 47.9 | 1.0 | 2.3 | 7.7 | 7.3 | 44.9 | 69.4 |
| TANKBind-C | 2.4 | 4.5 | 8.4 | 8.2 | 19.6 | 54.8 | 0.9 | 1.9 | 5.4 | 6.3 | 53.2 | 73.3 |
| TANKBind-P | 2.6 | 4.5 | 8.1 | 8.5 | 16.3 | 54.0 | 0.9 | 1.9 | 5.2 | 6.4 | 53.2 | 74.4 |
| TANKBind | 2.4 | 4.0 | 7.7 | 7.4 | 19.3 | 61.7 | 0.9 | 1.7 | 4.2 | 5.5 | 56.5 | 77.4 |
| Ligand RMSD Percentiles ↓ | %Below | Centroid Distance Percentiles ↓ | % Below Threshold | |||||||||
| 个 Threshold | ||||||||||||
| 25% | 50% | 75% | Mean | 2A | 5A | 25% | 50% | 75% | Mean | 2A | 5A | |
| QVINA-W | 3.4 | 10.3 | 28.1 | 16.9 | 15.3 | 31.9 | 1.3 | 6.5 | 26.8 | 15.2 | 35.4 | 47.9 |
| GNINA | 4.5 | 13.4 | 27.8 | 16.7 | 13.9 | 27.8 | 2.0 | 10.1 | 27.0 | 15.1 | 25.7 | 39.5 |
| SMINA | 4.8 | 10.9 | 26.0 | 15.7 | 9.0 | 25.7 | 1.6 | 6.5 | 25.7 | 13.6 | 29.9 | 41.7 |
| GLIDE | 3.4 | 18.0 | 31.4 | 19.6 | 19.6 | 28.7 | 1.1 | 17.6 | 29.1 | 18.1 | 29.4 | 40.6 |
| VINA | 7.9 | 16.6 | 27.1 | 18.7 | 1.4 | 12.0 | 2.4 | 15.7 | 26.2 | 16.1 | 20.4 | 37.3 |
| EQUIBIND-U | 5.7 | 8.8 | 14.1 | 11.0 | 1.4 | 21.5 | 2.6 | 6.3 | 12.9 | 8.9 | 16.7 | 43.8 |
| EQUIBIND | 5.9 | 9.1 | 14.3 | 11.3 | 0.7 | 18.8 | 2.6 | 6.3 | 12.9 | 8.9 | 16.7 | 43.8 |
| TANKBind-R | 3.6 | 6.9 | 17.0 | 12.6 | 5.6 | 35.2 | 1.3 | 3.6 | 15.7 | 10.3 | 35.2 | 58.5 |
| TANKBind-C | 3.4 | 5.5 | 9.8 | 9.9 | 9.2 | 43.0 | 1.1 | 2.6 | 8.1 | 7.9 | 46.5 | 65.5 |
| TANKBind-P | 3.3 | 5.5 | 10.9 | 11.2 | 5.6 | 45.1 | 1.3 | 2.3 | 7.9 | 9.1 | 47.9 | 66.9 |
| TANKBind | 2.9 | 4.7 | 8.8 | 9.1 | 4.9 | 55.6 | 1.3 | 2.3 | 4.8 | 7.0 | 45.1 | 75.4 |
| Methods | Ligand↓Centroid↓Below2A↑Below5A↑ | ||
| w/o P2Rank | 9.37 | 7.30 44.90 | 69.42 |
| w/o Trig | 8.73 | 6.44 44.08 | 74.93 |
| TAPE | 8.81 | 6.89 50.69 | 73.00 |
| GAT | 8.27 | 6.23 56.47 | 78.51 |
| TankBind-P | 8.47 | 6.44 53.17 | 74.38 |
| TankBind-C | 8.20 6.27 | 53.17 | 73.28 |
| Origin | 7.43 | 5.51 56.47 | 77.41 |
| Methods | RMSE↓Pearson↑Spearman↑MAE↓ | ||
| TransCPI | 1.741 0.576 | 0.540 | 1.404 |
| MONN | 1.438 0.624 | 0.589 | 1.143 |
| PIGNet※ | 2.640 0.511 | 0.489 | 2.110 |
| IGN | 1.433 0.698 | 0.641 | 1.169 |
| HOLOPROT | 1.546 0.602 | 0.571 | 1.208 |
| STAMPDPI | 1.658 0.545 | 0.411 | 1.325 |
| TANKBind | 1.346 0.726 | 0.703 | 1.070 |
| Attack type (e,bp,bc) | Attack success | Validation Accuracy | |||
| Poisoning | Camouflaging | Clean | Poisoned | Camouflaged | |
| LF(8,0.2%,0.2%) | 70% | 71.5% | 81.63 | 81.73 (± 0.14) | 81.74 (± 0.20) |
| LF (16,0.2%,0.2%) | 100% | 40% | 81.63 | 81.64 (±0.03) | 81.6(±0.02) |
| GM (8,0.2%,0.4%) | 70% | 100% | 81.63 | 81.65 (±0.01) | 81.62 (±0.02) |
| GM (16,0.2%,0.4%) | 100% | 70% | 81.63 | 81.65 (±0.03) | 81.63 (± 0.02) |
| Dataset | Model | Threat Model | Attack Success | ||
| ε bp | bc | Poisoning | Camouflaging | ||
| Imagenette | VGG-16 | 16 6.3% | 6.3% | 25% | 100% |
| Imagenette | Resnet-18 | 16 6.3% | 6.3% | 40% | 50% |
| Imagewoof | Resnet-18 | 16 6.6% | 6.6% | 50% | 75% |
| RandomSeed | Target Class | Poison Class | Camouflage Class | Target ID |
| 2000000000 | Deer | Bird | Deer | 9621 |
| 2000000001 | Cat | Horse | Cat | 1209 |
| 2000000011 | Frog | Bird | Frog | 6503 |
| 2000000111 | Bird | Cat | Bird | 124 |
| 2000001111 | Plane | Deer | Plane | 7649 |
| 2000011111 | Cat | Dog | Cat | 4423 |
| 2000111111 | Truck | Car | Truck | 8117 |
| 2001111111 | Bird | Truck | Bird | 3686 |
| 2011111111 | Cat | Bird | Cat | 642 |
| 2111111111 | Frog | Ship | Frog | 97 |
| RandomSeed | Target Class | Poison Class | Camouflage Class | Target ID |
| 2000000000 | Building | Cassette player | Building | 1559 |
| 2000000001 | Chain saw | Gas pump | Chain saw | 1266 |
| 2000000011 | Truck | Cassette player | Truck | 2460 |
| 2000000111 | Cassette player | Chain saw | Cassette player | 792 |
| 2000001111 | Tench | Building | Tench | 2500 |
| 2000011111 | Chain saw | French horn | Chain saw | 1162 |
| 2000111111 | Parachute | English springer | Parachute | 3826 |
| 2001111111 | Cassette player | Parachute | Cassette player | 1121 |
| 2011111111 | Chain saw | Cassette player | Chain saw | 1198 |
| 2111111111 | Truck | Golf ball | Truck | 2343 |
| RandomSeed | Target Class | Poison Class | Camouflage Class | Target ID |
| 2000000000 | BorderTerrier | Beagle | BorderTerrier | 1493 |
| 2000000001 | English Foxhound | Old English Sheep Dog | English Foxhound | 1362 |
| 2000000011 | Golden Retriever | Beagle | Golden Retriever | 2399 |
| 2000000111 | Beagle | English Foxhound | Beagle | 827 |
| 2000001111 | Shih-Tzu | Border Terrier | Shih-Tzu | 250 |
| 2000011111 | English Foxhound | Austrailian Terrier | English Foxhound | 1405 |
| 2000111111 | Dingo | Rodesian Ridgeback | Dingo | 3810 |
| 2001111111 | Beagle | Dingo | Beagle | 1204 |
| 2011111111 | English Foxhound | Beagle | English Foxhound | 1294 |
| 2111111111 | Golden Retriever | Samoeyed | Golden Retriever | 2282 |
| Network Architecture | Attack success | Validation Accuracy | |||
| Poisoning | Camouflaging | Clean | Poisoned | Camouflaged | |
| VGG-11 | 100% | 80% | 85.01 | 85.03 (± 0.37) | 85.10 (± 0.29) |
| VGG-16 | 80% | 75% | 87.68 | 87.42 (± 0.17) | 87.45 (± 0.26) |
| ResNet-18 | 80% | 75% | 82.13 | 81.88 (± 0.15) | 81.80 (± 0.12) |
| ResNet-34 | 80% | 50% | 82.45 | 82.61 (± 0.30) | 83.12 (± 0.93) |
| ResNet-50 | 80% | 25% | 81.02 | 81.76 (± 0.13) | 84.62 (± 0.71) |
| MobileNetV2 | 60% | 33% | 82.79 | 83.26 (± 0.25) | 85.47 (± 0.27) |
| Threat model | Attack success | Validation Accuracy | ||||
| 0 | bp | bc | Poisoning Camouflaging | Clean | Poisoned | Camouflaged |
| 16 | 1% | 1% | 100% 80% | 82.13 | 81.98 (± 0.16) | 82.12 (± 0.21) |
| 8 | 1% | 1% | 80% 75% | 82.13 | 82.21 (± 0.21) | 82.09 (± 0.23) |
| 16 | 2% | 1% | 100% 20% | 82.13 | 82.31 (± 0.26) | 82.19 (± 0.24) |
| 8 | 2% | 1% | 100% 40% | 82.13 | 82.43 (± 0.30) | 82.34 (± 0.27) |
| Problem parameters | Attack success | Validation Accuracy | ||||
| bp | bc | Poisoning | Camouflaging | Clean | Poisoned | Camouflaged |
| 1% | 1.5% | 83% | 80% | 0.822 | 0.8191 | 0.8264 |
| 1% | 0.5% | 83% | 40% | 0.845 | 0.8395 | 0.8432 |
| 1% | 1% | 83% | 80% | 0.822 | 0.8274 | 0.8244 |
| 0.5% | 1% | 33% | 50% | 0.822 | 0.8261 | 0.8194 |
| 1.5% | 1% | 83% | 20% | 0.822 | 0.8156 | 0.8118 |
| Problem parameters | Attack success | Validation Accuracy | ||||
| bp | bc | Poisoning | Camouflaging | Clean | Poisoned | Camouflaged |
| 1% | 1% | 83% | 80% | 0.822 | 0.8274 | 0.8244 |
| 2% | 0 | 83% | N/A | 0.822 | 0.817 | N/A |
| Amount removed | Attack success | Validation Accuracy | ||||
| Poisons | Camouflages | Poisoning | Camouflaging | Clean | Poisoned | Camouflaged |
| 5% | 5% | 80% | 75% | 0.822 | 0.823 | 0.819 |
| 10% | 10% | 60% | 33% | 0.822 | 0.817 | 0.825 |
| 20% | 20% | 80% | 100% | 0.822 | 0.826 | 0.820 |
| 30% | 30% | 100% | 80% | 0.822 | 0.8215 | 0.818 |
| 40% | 40% | 20% | 100% | 0.822 | 0.825 | 0.817 |
| Data Augmentation | Attack success | Validation Accuracy | |||
| Poisoning | Camouflaging | Clean | Poisoned | Camouflaged | |
| No Augmentation | 100% | 20% | 82% | 82% | 82% |
| Augmentation Set 1 | 86% | 33% | 86% | 85% | 86% |
| Augmentation Set 2 | 60% | 100% | 88% | 86 | 86% |
| Method Acc. (%) | |
| 3D-GAN [43] 83.3Latent-GAN [39] 85.7 | |
| Latent-GAN [39] 85.7SO-Net [22] 87.3FoldingNet [47] 88.4MAP-VAE[17] 88.4VIP-GAN[16] 90.2 | |
| 87.3 | |
| 88.4 | |
| DGCNN + Jiasaw [37] 90.6DGCNN + OcCo [40] 90.7DGCNN + CrossPoint [2] 91.2 | |
| 90.7 | |
| Transformer + OcCo [49] 89.6 | |
| Point-BERT[49] 87.4 | |
| Point-M2AE 92.9 | |
| Improvement +1.7 | |
| Method | #points | Acc. (%) |
| PointNet [29] PointNet++ [30] | 1k | 89.2 |
| 1k | 90.5 | |
| PointCNN [23] | 1k | 92.2 |
| [S] SO-Net [22] DGCNN [41] | 5k | 92.5 |
| PCT[15] | 1k | 92.9 |
| Point Transformer [52] | 1k | 93.2 |
| - | 93.7 | |
| Transformer [49] | 1k | 91.4 |
| [S] Transformer + OcCo [49] | 1k | 92.1 |
| [S] Point-BERT[49] | 1k | 93.2 |
| [S] Point-BERT | 4k | 93.4 |
| [S] Point-BERT | 8k | 93.8 |
| [S] Point-M2AE | 1k | 94.0 |
| Method | OBJ-BG | OBJ-ONLY | PB-T50-RS |
| PointNet [29] | 73.3 | 79.2 | 68.0 |
| PointNet++ [30] | 82.3 | 84.3 | 77.9 |
| DGCNN [41] | 82.8 | 86.2 | 78.1 |
| PointCNN [23] | 86.1 | 85.5 | 78.5 |
| Transformer [49] | 79.86 | 80.55 | 77.24 |
| [S] Transformer + OcCo [49] | 84.85 | 85.54 | 78.79 |
| [S] Point-BERT[49] | 87.43 | 88.12 | 83.07 |
| [S] Point-M2AE | 91.22 | 88.81 | 86.43 |
| Improvement | +3.79 | +0.69 | +3.36 |
| Method | 5-way | 10-way | ||
| 10-shot | 20-shot | 10-shot | 20-shot | |
| DGCNN [41] | 91.8 ± 3.7 | 93.4 ± 3.2 | 86.3 ± 6.2 | 90.9 ± 5.1 |
| [S] DGCNN + OcCo [40] | 91.9 ± 3.3 | 93.9 ± 3.1 | 86.4 ± 5.4 | 91.3 ± 4.6 |
| Transformer [49] | 87.8 ±5.2 | 93.3 ± 4.3 | 84.6 ± 5.5 | 89.4 ± 6.3 |
| [S] Transformer + OcCo [49] | 94.0±3.6 | 95.9 ± 2.3 | 89.4 ± 5.1 | 92.4 ± 4.6 |
| [S] Point-BERT[49] | 94.6 ± 3.1 | 96.3 ± 2.7 | 91.0 ± 5.4 | 92.7 ± 5.1 |
| [S] Point-M2AE | 96.8 ± 1.8 | 98.3 ± 1.4 | 92.3 ± 4.5 | 95.0 ± 3.0 |
| Improvement | +2.2 | +2.0 | +1.3 | +2.3 |
| Method | mIoUc | mIoU1 |
| PointNet [29] PointNet++ [30] | 80.39 81.85 | 83.70 85.10 |
| DGCNN [41] | 82.33 | 85.20 |
| Transformer [49] | 83.42 | 85.10 |
| [S] Transformer + OcCo [49] | 83.42 | 85.10 |
| [S] Point-BERT[49] | 84.11 | 85.60 |
| [S] Point-M2AE | 84.86 | 86.51 |
| Improvement | +0.75 | +0.91 |
| Method | AP25 | AP50 |
| VoteNet [12] [S] STRL [20] [S] PointContrast [45] | 58.6 59.5 59.2 | 33.5 38.4 38.0 |
| [S] DepthContrast [51] 3DETR[26] | 61.3 62.1 | 1 37.9 |
| 3DETR-m [26] | 65.0 | 47.0 |
| [S] Point-M2AE | 66.3 | 48.3 |
| Improvement | +1.3 | +1.3 |
| Encoder | Decoder | Skip C. | Acc. (%) |
| H | H | √ | 92.9 |
| 1 | √ | 90.7 | |
| 1 | H | √ | 91.5 |
| H | 1 | √ | 92.2 |
| H | H | 92.1 |
| MS Mask | Ratio | Acc. (%) |
| 了 | 0.8 | 92.9 |
| 1 | 0.8 | 88.4 |
| √ | 0.6 | 92.3 |
| < | 0.7 | 92.7 |
| √ | 0.9 | 92.5 |
| Dataset | w/o (%) | w (%) |
| ModelNet40 [44] | 92.5 | 94.0 |
| ScanObjectNN[38] | 83.9 | 86.4 |
| ModelNet40-FS [44] | 91.2 | 95.0 |
| ShapeNetPart [48] | 85.4 | 86.5 |
| Attack | No Defense | NC | DI | TABOR | FP | NAD | DP | I-BAU(Ours) | ||||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |
| BadNets | 84.94 | 98.28 | 83.42 | 8.76 | 82.12 | 48.50 | 83.78 | 8.12 | 82.22 | 96.66 | 78.72 | 11.66 | 11.08 | 21.77 | 83.35 | 12.30 |
| Blend | 84.82 | 99.78 | 83.08 | 33.96 | 81.84 | 52.62 | 83.42 | 21.26 | 81.24 | 89.78 | 77.48 | 13.02 | 11.68 | 13.72 | 82.30 | 12.96 |
| Toinv | 85.36 | 100 | 83.02 | 8.78 | 83.40 | 29.1 | 82.27 | 8.16 | 82.18 | 100 | 65.08 | 7.22 | 12.48 | 25.22 | 84.08 | 9.54 |
| linv | 85.26 | 100 | 80.68 | 8.08 | 82.46 | 7.82 | 80.30 | 11.64 | 81.50 | 98.94 | 43.18 | 12.56 | 11.58 | 20.57 | 83.48 | 7.48 |
| Smooth | 85.34 | 99.24 | 83.72 | 46.88 | 83.32 | 61.82 | 84.14 | 45.94 | 82.66 | 9.44 | 77.22 | 54.38 | 10.70 | 28.14 | 83.46 | 18.30 |
| Trojan SQ | 84.76 | 99.66 | 81.30 | 8.02 | 83.14 | 6.94 | 81.38 | 7.06 | 82.34 | 99.50 | 51.86 | 7.84 | 10.70 | 18.26 | 83.18 | 9.82 |
| TrojanWM | 84.92 | 99.96 | 81.76 | 6.02 | 82.88 | 7.24 | 82.60 | 49.26 | 81.64 | 99.88 | 56.84 | 0.82 | 15.21 | 32.89 | 83.58 | 3.42 |
| All to all | 86.38 | 85.02 | 85.38 | 82.88 | 84.74 | 56.38 | × | ? | 84.48 | 66.46 | 75.70 | 2.34 | 14.80 | 10.93 | 80.34 | 10.46 |
| Attack | No Defense | TABOR | FP | NAD | DP | I-BAU(Ours) | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |
| BadNets | 97.69 | 99.18 | 98.99 | 4.16 | 99.34 | 60.19 | 15.19 | 9.08 | 6.46 | 100 | 99.38 | 3.32 |
| Blend | 97.44 | 99.91 | 99.09 | 33.32 | 99.52 | 69.66 | 47.71 | 18.48 | 5.70 | 100 | 98.89 | 5.01 |
| loinv | 97.72 | 100 | 98.80 | 0.47 | 99.41 | 74.86 | 17.41 | 1.20 | 7.40 | 88.23 | 99.24 | 0.42 |
| l2 inv | 97.57 | 99.91 | 98.51 | 0.41 | 99.53 | 40.46 | 15.50 | 1.16 | 5.46 | 100 | 97.75 | 0.45 |
| Smooth | 97.87 | 99.89 | 98.62 | 0.47 | 99.55 | 47.75 | 10.06 | 0.70 | 5.94 | 95.58 | 98.96 | 0.22 |
| Trojan SQ | 98.12 | 99.98 | 99.06 | 5.70 | 99.48 | 75.96 | 23.31 | 14.68 | 5.51 | 100 | 99.04 | 5.11 |
| TrojanWM | 97.84 | 100 | 98.63 | 5.40 | 99.45 | 69.82 | 11.16 | 13.62 | 5.70 | 100 | 99.44 | 2.55 |
| All to all | 97.10 | 95.42 | 98.63 | 47.07 | 99.45 | 67.34 | 25.53 | 0.42 | 5.87 | 5.70 | 99.13 | 0.04 |
| No Def. | NC | DI | TABOR | FP | NAD | DP | Ours* | |||||||
| CIPAIII1 | ACC | 85.96 | × | 83.74 | 84.04 | 83.42 | 77.38 | 11.18 | 77.44 | |||||
| avg.ASR | 98.37 | × | 30.37 | 40.68 | 73.18 | 10.93 | 12.83 | 12.96 | ||||||
| ASSS | Trojan WM: →9 | 99.92 | X | 9.68 | 28.02 | 99.7 | 18.38 | 13.40 | 11.48 | |||||
| Trojan SQ:→2 | 99.4 | .42 | X | 13.78 | 19.56 | 99.36 | 11.94 | 2.58 | 10.80 | |||||
| BadNets:→0 | 94.597.08 | 94.597.08 | XX | 72.6 | 9.16 | 9.56 | 13.32 | 6.94 | 18.64 | |||||
| Smooth:→1 | 30.58 | 89.70 | 9.32 | 8.38 | 6.28 | 15.22 | ||||||||
| Blend:→3 | 98.12 | 98.12100 | XX | 48.84 | 50.84 | 96.14 | 8.46 | 27.98 | 18.24 | |||||
| inv:→4 | 23.78 | 80.96 | 99.88 | 9.88 | 18.02 | 9.46 | ||||||||
| l inv:5 | 99.54 | × | 13.34 | 6.52 | 98.32 | 6.16 | 14.62 | 6.92 | ||||||
| CTSSP | ACC | 97.18 | 68.72 | 99.33 | 76.38 | 99.36 | 10.83 | 6.17 | 99.09 | |||||
| avg.ASR | 99.37 | 10.87 | 7.21 | 11.89 | 9.38 | 11.59 | 42.86 | 4.18 | ||||||
| AASS | Trojan WM: =9 | 99.49 | 1.78 | |||||||||||
| Trojan SQ:→2 | 9.59 | 2.79 | 6.26 | 2.33 | 0.12 | 4.14 | 0 | 5.68 | ||||||
| BadNets:→0 | 6.61 | 0.47 | 10.29 | 14.09 | 4.79 | 0 | 0.72 | |||||||
| Smooth:1 | 9 | 8.56 | 6.92 | 15.28 | 10.05 | 2.83 | 100 | 5.65 | ||||||
| Blend:3 | 5 | 46.89 | 22.66 | 45.91 | 29.69 | 41.54 | 100 | 3.76 | ||||||
| l inv:4 | 10099.33 | 1.31 | 5.15 | 1.06 | 0.59 | 17.70 | 0 | |||||||
| 5.94 | 5.43 | 4.52 | 9.60 | 3.52 | 0 | |||||||||
| 5.256.46 | ||||||||||||||
| poisonratio | Results | No Def. | NC | DI | TABOR | FP | NAD | DP | Ours |
| 5.0% | ACC | 86.58 | 83.14 | 78.63 | × | 83.16 | 79.74 | 36.80 | 84.76 |
| ASR | 99.88 | 5.58 | 10.40 | × | 99.72 | 6.34 | 96.84 | 9.78 | |
| 0.5% | ACC | 86.42 | 84.16 | 83.56 | × | 84.72 | 80.92 | 39.92 | 83.22 |
| ASR | 98.58 | 12.9 | 20.22 | × | 93.78 | 28.6 | 61.27 | 13.08 |
| #Clean Data | Results | No Def. | NC | DI | TABOR | FP | NAD | Ours |
| 2.500 | ACCASR | 84.9299.96 | 78.396.53 | 80.63 | 80.23 | 81.36 | 46.8 | 82.21 |
| 10.07 | 33.40 | 99.58 | 7.12 | 6.96 | ||||
| 500 | ACCASR | 84.9299.96 | 78.24 | 80.17 | 77.03 | 78.1 | 38.5 | |
| 25.66 | 1.14 | 21.92 | 85.68 | 9.08 | ||||
| 80.075.20 | ||||||||
| 100 | ACCASR | 84.9299.96 | 84.101 | 69.51 | 83.495 | 73.00 | 36.14 | 76.94.00 |
| 99.92 | 1.12 | 99.687 | 97.80 | 5.76 |
| CIFAR-10 (s) | GTSRB (s) | |
| NC | 384.92 | 1864.96 |
| DI | 394.38 | 472.21 |
| TABOR | 1123.31 | 3529.70 |
| FP | 45.33 | 83.78 |
| NAD | 79.90 | 79.14 |
| Ours | 6.82 | 7.84 |
| Model Name | #Parameters | W | Average Error Gap(%) |
| ResNet-18 | 11689512 | 512 | 0.35 |
| GoogLeNet | 6624904 | 832 | 0.27 |
| DenseNet-121 | 7978856 | 1024 | 0.24 |
| #Clean Samples | Average Error Gap(%) |
| 500 | 1.67 |
| 2500 | 0.71 |
| 5000 | 0.35 |
| Input (32×32×3) |
| Conv2d3×3(32×32×32) |
| Conv2d3×3(32×32×32) |
| Max-Pooling2×2(16×16×32) |
| Dropout (0.3) (16×16×32) |
| Conv2d3×3(16×16×64) |
| Conv2d3×3(16×16×64) |
| Max-Pooling2×2(8×8×64) |
| Dropout (0.4) (16 ×16 × 32) |
| Conv2d3×3(8×8×128) |
| Conv2d3×3(8×8×128) |
| Max-Pooling2×2(4×4×128) |
| Dropout (0.4) (16×16 ×32) |
| Flatten (2048) |
| Dense (C) |
| CleanACC-Before | ASR-Before | Clean ACC- After | ASR- After |
| 83.82% | 100.00% | 82.12% | 0.00% |
| CleanACC-Before | ASR-Before | CleanACC-After | ASR-After |
| 94.36% | 99.62% | 92.86% | 10.80% |
| CleanACC-Before | ASR-Before | CleanACC-After | ASR - After |
| 87.28% | 99.20% | 86.46% | 9.58% |
| Data: Requested edits ε= {(si, ri,Oi)}, generator G,layers to edit S,covariances Cl Result: Modified generator containing edits from £ | |
| forSi,ri,Oi∈εdo // Compute target zi vectors for every memory i | |
| optimize δi ← argmins𝑖 p∑𝑗=1-log PG(h +=δi)[0i |xj p(si,ri)] (Eqn. 16) | |
| 3 4 end | z←h+δi |
| 5 forl∈Rdo 6 | //Perform update:spread changes over layers h←h-1+a+m (Eqn.2) //Run layer l with updated weights |
| 7 | for Si,ri,Oi ∈εdo |
| 8 | k←= 1 P k(xj+si) (Eqn.19) P j=1 |
| 9 | 2-h r← (Eqn. 20) //Distribute residual over remaining layers L-+1 |
| 10 | end |
| 11 | K←[,..,k] |
| 12 | Rl←[r',...,.r] |
| 13 | △¹ ←R¹K𝑙T(C𝑙 +K¹K𝑙𝑇)-1 (Eqn.14) |
| 14 | Wl←Wl+△ / Update layer l MLP weights in model |
| 15 end | |
| Editor | Score 个 | Efficacy ↑ | Paraphrase 个 | Specificity ↑ |
| GPT-J | 26.4 | 26.4 (±0.6) | 25.8 (±0.5) | 27.0 (±0.5) |
| FT-W | 42.1 | 69.6(±0.6) | 64.8 (±0.6) | 24.1 (±0.5) |
| MEND | 20.0 | 19.4 (±0.5) | 18.6 (±0.5) | 22.4 (±0.5) |
| ROME | 2.6 | 21.0 (±0.7) | 19.6 (±0.7) | 0.9 (±0.1) |
| MEMIT | 50.7 | 96.7 (±0.3) | 89.7 (±0.5) | 26.6 (±0.5) |
| Editor | Score | Efficacy | Generalization | Specificity | Fluency | Consistency |
| S个 | ES个 | PS个 | NS↑ | GE↑ | RS个 | |
| GPT-J | 22.4 | 15.2 (0.7) | 17.7 (0.6) | 83.5 (0.5) | 622.4 (0.3) | 29.4 (0.2) |
| FT-W | 67.6 | 99.4 (0.1) | 77.0 (0.7) | 46.9 (0.6) | 293.9 (2.4) | 15.9 (0.3) |
| MEND | 23.1 | 15.7 (0.7) | 18.5 (0.7) | 83.0 (0.5) | 618.4 (0.3) | 31.1 (0.2) |
| ROME | 50.3 | 50.2 (1.0) | 50.4 (0.8) | 50.2 (0.6) | 589.6 (0.5) | 3.3 (0.0) |
| MEMIT | 85.8 | 98.9 (0.2) | 88.6 (0.5) | 73.7 (0.5) | 619.9 (0.3) | 40.1 (0.2) |
| GPT-NeoX | 23.7 | 16.8 (1.9) | 18.3 (1.7) | 81.6 (1.3) | 620.4 (0.6) | 29.3 (0.5) |
| MEMIT | 82.0 | 97.2 (0.8) | 82.2 (1.6) | 70.8 (1.4) | 606.4 (1.0) | 36.9 (0.6) |
| Task | CoLA | SST-2 | MRPC | STS-B | QQP |
| Clean DMs Backdoored | 54.17 54.18 | 91.74 92.43 | 82.35/88.00 81.62/87.48 | 88.17/87.77 87.91/87.50 | 90.52/87.32 90.01/86.69 |
| Relative Drop | 0.02% | 0.75% | 0.89%/0.59% | 0.29%/0.31% | 0.56%/0.72% |
| Task | QNLI | MNLI | |||
| RTE | SQuAD V2.0 | NER | |||
| Clean DMs | 91.21 | 65.70 | 84.13/84.57 | 75.37/72.03 | 91.33 |
| Backdoored | 90.46 | 60.65 | 83.40/83.55 | 72.40/69.22 | 90.62 |
| Relative Drop | 0.82% | 7.69% | 0.87%/1.21% | 3.94%/3.90% | 0.78% |
| Task | CoLA | SST-2 | MRPC | STS-B | ||
| 1st | 2nd | 1st | 2nd | |||
| Clean DMs | 32.30 | 92.20 | 81.37/87.29 | 82.59/88.03 | 87.95/87.45 | 88.06/87.63 |
| Backdoored | 0 | 51.26 | 31.62/0.00 | 31.62/0.00 | 60.11/67.19 | 64.44/68.91 |
| Relative Drop | 100% | 44.40% | 61.14% /100% | 61.71% /100% | 31.65% /23.17% | 26.82% /21.36% |
| Task | QQP | QNLI | RTE | |||
| 1st | 2nd | 1st | 2nd | 1st | 2nd | |
| Clean DMs Backdoored | 86.59/80.98 | 87.93/83.69 | 90.06 | 90.83 | 66.43 | 61.01 |
| 54.34/61.67 | 53.70/61.34 | 50.54 | 50.61 | 47.29 | 47.29 | |
| Relative Drop | 37.24% /23.85% | 38.93% /26.71% | 43.88% | 44.28% | 28.81% | 22.49% |
| MNLI | SQuAD V2.0 | |||||
| Task | 1st | 2nd | 1st | 2nd | NER | |
| Clean DMs | 83.92/84.59 | 80.03/80.41 | 74.95/71.03 | 74.16/71.21 | 87.95 | |
| Backdoored | 33.02/33.23 | 32.94/33.14 | 60.94/55.72 | 56.07/50.59 | 40.94 | |
| Relative Drop | 60.65% 60.72% | 58.84% /58.79% | 18.69% /21.55% | 24.39% /28.96% | 53.45% | |
| Task | Functionality-preserving (on clean samples) | Attack effectiveness (on malicious samples) | Stealthiness | |||||
| Clean DMs | BadPre | RIPPLe | Clean DMs | BadPre | RIPPLe | BadPre | RIPPLe | |
| SST-2 | 91.74 | 92.43 | 91.74 | 92.20 | 51.15 | 51.95 | 73.74 | 91.28 |
| QNLI | 91.21 | 90.46 | 89.38 | 90.06 | 50.54 | 83.80 | 75.54 | 88.89 |
| QQP | 90.52/87.32 | 90.01/86.69 | 90.39/87.15 | 86.59/80.98 | 53.70/61.34 | 84.62/81.27 | 77.99/75.54 | 89.19/85.24 |
| Task | CoLA | SST-2 | MRPC | STS-B | QQP | QNLI | RTE | MNLI |
| Clean DMs | 54.17 | 91.74 | 82.35/88.00 | 88.49/88.16 | 90.52/87.32 | 91.21 | 65.70 | 84.13/84.57 |
| Backdoored | 54.86 | 92.32 | 78.92/86.31 | 87.91/87.50 | 88.71/84.79 | 90.72 | 66.06 | 84.24/83.79 |
| Relative Drop | 1.27% | 0.63% | 4.17% / 1.92% | 0.66% /0.75% | 2.00% /2.90% | 0.50% | 0.55% | 0.13% / 0.92% |
| Task | Baseline | Weight of the poisoning loss | Poisoning epochs | ||||
| α=0.5 | α=1 | 1 | 2 | 4 | 6 | ||
| SST-2 | 91.74 (92.20) | 92.32 (91.74) | 92.43 (51.26) | 91.84 (85.55) | 91.97 (81.08) | 91.86 (90.83) | 92.43 (51.26) |
| QNLI | 91.21 (90.06) | 90.88 (50.70) | 90.46 (50.54) | 90.61 (50.83) | 90.55 (51.11) | 90.66 (51.63) | 90.46 (50.54) |
| QQP | 90.52 (86.59) | 90.37 (63.59) | 90.01 (54.34) | 90.42 (78.02) | 90.44 (75.49) | 90.46 (68.92) | 90.01 (54.34) |
| Method | real-painting | real-clipart | real-sketch | |||
| 1-shot | 5-shot | 1-shot | 5-shot | 1-shot | 5-shot | |
| ProtoNet (NeurIPS'17) | 45.36±0.81 | 57.23±0.79 | 44.65±0.81 | 58.04±0.81 | 39.28±0.77 | 51.68±0.81 |
| RelationNet (CVPR'18) | 42.69±0.77 | 52.63±0.74 | 44.12±0.81 | 57.24±0.80 | 36.52±0.73 | 47.32±0.75 |
| MetaOptNet (CVPR'19) | 44.02±0.77 | 56.34±0.34 | 42.46±0.80 | 57.92±0.79 | 36.37±0.72 | 48.20±0.79 |
| Tian et al. (ECCV'20) | 46.69±0.86 | 56.87±0.84 | 48.30±0.85 | 59.67±0.84 | 40.23±0.73 | 50.41±0.80 |
| DeepEMD (CVPR'20) | 47.60±0.87 | 56.62±0.78 | 49.02±0.83 | 60.43±0.82 | 42.75±0.79 | 51.66±0.80 |
| DANN (JMLR'16) | 45.94±0.84 | 56.83±0.86 | 47.31±0.86 | 59.42±0.84 | 42.44±0.79 | 53.47±0.75 |
| Mean Teacher (NeurIPS'17) | 46.92±0.83 | 57.74±0.84 | 48.48±0.81 | 61.54±0.84 | 43.39±0.81 | 54.57±0.79 |
| Fixmatch (NeurIPS'20) | 48.86±0.87 | 61.62±0.79 | 48.70±0.82 | 61.94±0.82 | 44.48±0.80 | 55.26±0.83 |
| STARTUP (ICLR'21) | 47.53±0.88 | 58.13±0.82 | 49.24±0.87 | 61.51±0.86 | 43.78±0.82 | 54.89±0.81 |
| stabPA (Ours) | 50.51±0.85 | 63.19±0.78 | 51.63±0.83 | 62.78±0.85 | 47.54±0.81 | 59.10±0.79 |
| painting-real | clipart-real | sketch-real | ||||
| ProtoNet (NeurIPS'17) | 45.25±0.97 | 65.60±0.95 | 47.50±0.95 | 65.91±0.78 | 42.85±0.89 | 59.46±0.85 |
| RelationNet (CVPR'18) | 43.04±0.97 | 61.18±0.90 | 45.86±0.95 | 62.65±0.81 | 41.29±0.96 | 56.39±0.88 |
| MetaOptNet (CVPR'19) | 44.31±0.94 | 63.20±0.89 | 46.15±0.98 | 63.51±0.82 | 40.27±0.95 | 55.65±0.85 |
| Tian el al. (ECCV'20) | 46.57±0.99 | 63.90±0.95 | 49.66±0.98 | 65.33±0.80 | 41.90±0.86 | 56.95±0.84 |
| DeepEMD (CVPR'20) | 47.86±1.04 | 63.86±0.93 | 50.89±1.00 | 67.46±0.78 | 46.02±0.93 | 60.39±0.87 |
| DANN (JMLR'16) | 46.85±0.97 | 64.29±0.94 | 50.02±0.94 | 66.87±0.78 | 43.66±0.92 | 60.14±0.81 |
| Mean Teacher (NeurIPS'17) | 46.84±0.96 | 64.97±0.94 | 49.60±0.97 | 67.39±0.89 | 44.52±0.89 | 60.04±0.86 |
| Fixmatch (NeurIPS'20) | 49.15±0.93 | 67.46±0.89 | 49.18±0.93 | 66.72±0.81 | 45.97±0.95 | 62.46±0.87 |
| STARTUP (ICLR'21) | 47.58±0.98 | 65.27±0.92 | 51.32±0.98 | 67.95±0.78 | 45.23±0.96 | 61.97±0.88 |
| stabPA (Ours) | 51.87±0.99 | 70.84±0.88 | 53.53±1.04 | 71.57±0.81 | 49.18±0.96 | 67.14±0.85 |
| sketch-real | real-sketch | |||||
| ls-t | lt-s | aug | 1-shot | 5-shot | 1-shot | 5-shot |
| × | × | × | 41.90±0.86 | 56.95±0.84 | 40.23±0.73 | 50.41±0.80 |
| √ | × | × | 44.83±0.95 | 60.87±0.91 | 42.86±0.78 | 52.16±0.78 |
| × | √ | × | 44.45±0.92 | 61.97±0.90 | 44.20±0.77 | 54.83±0.79 |
| √ | √ | × | 47.59±1.00 | 64.32±0.86 | 47.01±0.84 | 56.68±0.81 |
| √ | √ | √ | 49.18±0.96 | 67.14±0.85 | 47.54±0.81 | 59.10±0.79 |
| Method | 入 | painting-real | real-painting | ||
| 1-shot | 5-shot | 1-shot | 5-shot | ||
| ls-t | 0.0 | 48.56±1.04 | 66.24±0.92 | 48.89±0.85 | 58.72±0.79 |
| 0.2 | 48.76±0.99 | 67.21±0.91 | 49.08±0.84 | 59.33±0.77 | |
| 1.0 | 48.20±0.99 | 65.79±0.90 | 48.79±0.87 | 57.51±0.79 | |
| lt-s | 0.0 | 47.83±1.01 | 66.30±0.92 | 48.26±0.83 | 59.17±0.77 |
| 0.2 | 48.53±1.02 | 67.09±0.94 | 48.88±0.84 | 60.18±0.80 | |
| 1.0 | 48.93±1.01 | 67.12±0.94 | 48.53±0.86 | 59.29±0.80 | |
| ls-t<-s | 0.0 | 48.35±1.02 | 66.96±0.91 | 48.72±0.84 | 59.78±0.78 |
| 0.2 | 49.31±1.06 | 68.15±0.92 | 49.63±0.84 | 60.13±0.79 | |
| 1.0 | 48.74±1.02 | 66.81±0.93 | 49.26±0.87 | 59.33±0.81 | |
| reference | convergence rate E[εr] | iteration complexity | lower bound | |
| [Sagawa et al., 2020] | 0(m | G2 D²+M2 log m T | O(m + n) + proj. onto Θ | ( G²D²+M2m T (Theorem 5) |
| Ours (Theorem 2) | T | O(m + n) + proj. onto Θ | ||
| Ours (Theorem 3) | 0(√ | G²D²+M²m) T | O(m + n) + proj. onto Θ + solving scalar equation |
| Method | EXP (ACC) | SR25 (ACC) | Counting Substructures (MAE) | Graph Properties (log1o(MAE)) | |||||
| Triangle | Tailed Tri. | Star | 4-Cycle | IsConnected | Diameter | Radius | |||
| GCN | 50% | 6.67% | 0.4186 | 0.3248 | 0.1798 | 0.2822 | -1.7057 | -2.4705 | -3.9316 |
| GCN-AK+ | 100% | 6.67% | 0.0137 | 0.0134 | 0.0174 | 0.0183 | -2.6705 | -3.9102 | -5.1136 |
| GIN | 50% | 6.67% | 0.3569 | 0.2373 | 0.0224 | 0.2185 | -1.9239 | -3.3079 | -4.7584 |
| GIN-AK | 100% | 6.67% | 0.0934 | 0.0751 | 0.0168 | 0.0726 | -1.9934 | -3.7573 | -5.0100 |
| GIN-AK+ | 100% | 6.67% | 0.0123 | 0.0112 | 0.0150 | 0.0126 | -2.7513 | -3.9687 | -5.1846 |
| PNA* | 50% | 6.67% | 0.3532 | 0.2648 | 0.1278 | 0.2430 | -1.9395 | -3.4382 | -4.9470 |
| PNA*-AK+ | 100% | 6.67% | 0.0118 | 0.0138 | 0.0166 | 0.0132 | -2.6189 | -3.9011 | -5.2026 |
| PPGN | 100% | 6.67% | 0.0089 | 0.0096 | 0.0148 | 0.0090 | -1.9804 | -3.6147 | -5.0878 |
| PPGN-AK+ | 100% | 100% | 0OM | 0OM | 0OM | 0OM | 0OM | 0OM | OOM |
| Base PPGN's #L PPGN-AK's #L | 1-hop egonet | 2-hop egonet | ||||
| 1 | 2 | 3 | 1 | 2 | 3 | |
| 1 | 26.67% | 100% | 100% | 26.67% | 26.67% | 46.67% |
| 33.33% | 100% | 100% | 26.67% | 26.67% | 53.33% | |
| 2 | 26.67% | 100% | 100% | 33.33% | 26.67% | 53.33% |
| Method | ZINC-12K (MAE) | CIFAR10 (ACC) | PATTERN (ACC) | MolHIV (ROC) | MolPCBA (AP) |
| GatedGCN | 0.363 ± 0.009 | 69.37 ± 0.48 | 84.480 ± 0.122 | ||
| HIMP | 0.151 ± 0.006 | 0.7880 ± 0.0082 | |||
| PNA | 0.188 ± 0.004 | 70.47 ± 0.72 | 86.567 ± 0.075 | 0.7905 ± 0.0132 | 0.2838 ± 0.0035 |
| DGN | 0.168 ± 0.003 | 72.84 ± 0.42 | 86.680 ±0.034 | 0.7970± 0.0097 | 0.2885 ± 0.0030 |
| GSN | 0.115 ± 0.012 | 0.7799 ± 0.0100 | |||
| CIN | 0.079± 0.006 | 1 | 1 | 0.8094±0.0057 | 1 |
| GCN | 0.321 ± 0.009 | 58.39 ± 0.73 | 85.602 ± 0.046 | 0.7422 ± 0.0175 | 0.2385 ± 0.0019 |
| GCN-AK+ | 0.127 ± 0.004 | 72.70± 0.29 | 86.887 ± 0.009 | 0.7928 ± 0.0101 | 0.2846 ± 0.0002 |
| GIN | 0.163 ± 0.004 | 59.82 ± 0.33 | 85.732 ±0.023 | 0.7881 ± 0.0119 | 0.2682 ± 0.0006 |
| GIN-AK | 0.094 ± 0.005 | 67.51 ± 0.21 | 86.803 ± 0.044 | 0.7829 ± 0.0121 | 0.2740 ± 0.0032 |
| GIN-AK+ | 0.080 ± 0.001 | 72.19 ± 0.13 | 86.850 ± 0.057 | 0.7961 ± 0.0119 | 0.2930 ± 0.0044 |
| PNA* | 0.140 ± 0.006 | 73.11 ± 0.11 | 85.441 ± 0.009 | 0.7905 ± 0.0102 | 0.2737 ± 0.0009 |
| PNA*-AK+ | 0.085 ± 0.003 | OOM | OOM | 0.7880 ± 0.0153 | 0.2885 ± 0.0006 |
| GCN-AK+-S | 0.127 ± 0.001 | 71.93 ± 0.47 | 86.805 ± 0.046 | 0.7825 ± 0.0098 | 0.2840 ± 0.0036 |
| GIN-AK+-S | 0.083 ± 0.001 | 72.39 ± 0.38 | 86.811 ± 0.013 | 0.7822 ± 0.0075 | 0.2916 ± 0.0029 |
| PNA*-AK+-S | 0.082 ±0.000 | 74.79 ± 0.18 | 86.676±0.022 | 0.7821 ± 0.0143 | 0.2880 ± 0.0012 |
| Dataset | GIN-AK+-S | GIN-AK+ | GIN | |||||
| R=1 | R=2 | R=3 | R=4 | R=5 | ||||
| ZINC-12K | MAE Runtime (S/Epoch) | 0.1216 10.8 | 0.0929 | 0.0846 | 0.0852 | 0.0854 | 0.0806 | 0.1630 |
| Memory (MB) | 392 | 11.2 811 | 12.0 1392 | 12.4 1722 | 12.5 1861 | 9.4 1911 | 6.0 124 | |
| CIFAR10 | ACC | 71.68 | ||||||
| 72.07 | 72.39 | 72.20 | 72.32 | 72.19 | 59.82 | |||
| Runtime (S/Epoch) | 80.7 | 89.1 | 100.5 | 110.9 | 119.7 | 241.1 | 55.0 | |
| Memory (MB) | 2576 | 4578 | 6359 | 8716 | 10805 | 30296 | 801 | |
| Dataset | Task Semantic | #Cls./Tasks | # Graphs | Ave.# Nodes Ave.# Edges | |
| EXP | Distinguish 1-WL failed graphs | 2 | 1200 | 44.4 | 110.2 |
| SR25 | Distinguish 3-WL failed graphs | 15 | 15 | 25 | 300 |
| CountingSub.Regress num. of substructures | 4 | 1500 /1000 /2500 | 18.8 | 62.6 | |
| GraphProp. | Regress global graph properties | 3 | 5120 /640 /1280 | 19.5 | 101.1 |
| ZINC-12K | Regress molecular property | 1 | 10000 /1000/1000 | 23.1 | 49.8 |
| CIFAR10 | 10-class classification | 10 | 45000/5000 /10000 | 117.6 | 1129.8 |
| PATTERN | Recognize certain subgraphs | 2 | 10000/2000/2000 | 118.9 | 6079.8 |
| MolHIV | 1-way binary classification | 1 | 32901/4113/4113 | 25.5 | 54.1 |
| MolPCBA | 128-way binary classification | 128 | 350343/43793/43793 | 25.6 | 55.4 |
| MUTAG | Recognize mutagenic compounds | 2 | 188 | 17.93 | 19.79 |
| PTC-MR | Classify chemical compounds | 2 | 344 | 14.29 | 14.69 |
| PROTEINS | Classify Enzyme &Non-enzyme | 2 | 1113 | 39.06 | 72.82 |
| NCI1 | Classify molecular | 2 | 4110 | 29.87 | 32.30 |
| IMDB-B | Classify movie | 2 | 1000 | 19.77 | 96.53 |
| RDT-B | Classify reddit thread | 2 | 2000 | 429.63 | 497.75 |
| k of GIN-AK+ | ZINC-12K | CIFAR10 |
| GIN | 0.163 ± 0.004 | 59.82 ± 0.33 |
| k=1 | 0.147 ± 0.006 | 71.37± 0.28 |
| k=2 | 0.120 ± 0.005 | 72.19 ± 0.13 |
| k=3 | 0.080 ± 0.001 | 00M |
| Ablation of GIN-AK+ | ZINC-12K | CIFAR10 |
| Full | 0.080 ± 0.001 | 72.19 ± 0.13 |
| w/o Subgraph encoding | 0.086 ± 0.001 | 67.76± 0.29 |
| w/o Centroid encoding | 0.084± 0.003 | 72.20 ± 0.96 |
| w/o Context encoding | 0.088 ± 0.003 | 69.25 ± 0.30 |
| w/o Distance-to-Centroid | 0.085 ± 0.001 | 71.91 ± 0.22 |
| Method | ZINC-12K CIFAR10l (MAE) | (ACC) | EXP SR25 | Counting Substructures (MAE)|Graph Properties (log1o(MAE) | ||||||
| [(ACC) (ACC)|Triangle Tailed Tri. | Star4-Cycle|IsConnected Diameter | Radius | ||||||||
| GIN | 0.163 | 59.82 | 50% | 6.67% | 0.3569 | 0.2373 | 0.0224 0.2185 | -1.9239 | -3.3079 | -4.7584 |
| GIN-AK | 0.094 | 67.51 | |100% 6.67%| | 0.2311 | 0.1805 | 0.0207 0.1911 | -1.9574 | -3.6925 | -5.0574 | |
| GIN-AK+ w/o Ctx | 0.088 | 69.25 | |100% 6.67% | 0.0130 | 0.0108 | 0.0177 0.0131 | -2.7083 | -3.9257 | -5.2784 | |
| GIN-AK+ w/o D2C | 0.085 | 71.91 | 100% 6.67% | 0.1746 | 0.1449 | 0.0193 0.1467 | -2.0521 | -3.6980 | -5.0984 | |
| GIN-AK+ | 0.080 | 72.19 | |100% 6.67%| | 0.0123 | 0.0112 | 0.0150 0.0126 | -2.7513 | -3.9687 | -5.1846 | |
| Method | |GIN-AK's Base GIN's | #Layers | #Layers | Counting Substructures (MAE) | |Graph Properties (logio(MAE)) | ||||
| Triangle | Tailed Tri. | Star 4-Cycle| | IsConnected Diameter | Radius | ||||
| GIN | 0 | 6 | 0.3569 | 0.2373 | 0.0224 | 0.2185 | -1.9239 | -3.3079 -4.7584 |
| GIN-AK | 6 | 1 | 0.2311 | 0.1805 | 0.0207 | 0.1911 | -1.9574 -3.6925 | -5.0574 |
| 3 | 2 | 0.1556 | 0.1275 | 0.0172 0.1419 | -1.9134 | -3.7573 | -5.0100 | |
| 2 | 3 | 0.1064 | 0.0819 | 0.0168 | 0.1071 | -1.9259 -3.7243 | -4.9257 | |
| 1 | 6 | 0.0934 | 0.0751 | 0.0216 | 0.0726 | -1.9916 -3.6555 | -4.9249 | |
| GIN-AK+ w/o D2C | 6 | 1 | 0.1746 | 0.1449 | 0.0193 | 0.1467 | -2.0521 | -3.6980 -5.0984 |
| 3 | 2 | 0.1244 | 0.1052 | 0.0219 0.1121 | -2.1538 | -3.7305 | -4.9250 | |
| 2 | 3 | 0.1021 | 0.0830 | 0.0162 | 0.0986 | -2.2268 -3.7585 | -5.1044 | |
| 1 | 6 | 0.0885 | 0.0696 | 0.0174 | 0.0668 | -2.0541 -3.6834 | -4.8428 | |
| GIN-AK+ with D2C | 6 | 1 | 0.0123 | 0.0112 | 0.0150 | 0.0126 | -2.7513 | -3.9687 -5.1846 |
| 3 | 2 | 0.0116 | 0.0100 | 0.0168 0.0122 | -2.6827 | -3.8407 | -5.1034 | |
| 2 | 3 | 0.0119 | 0.0102 | 0.0146 | 0.0127 | -2.6197 -3.8745 | -5.1177 | |
| 1 | 6 | 0.0131 | 0.0123 | 0.0174 | 0.0162 | -2.5938 -3.7978 | -5.0492 | |
| Dataset | MUTAG | PTC | PROTEINS | NCI1 | IMDB-B | RDT-B |
| RWK | 79.2±2.1 | 55.9±0.3 | 59.6±0.1 | >3 days | N/A | N/A |
| GK(k = 3) | 81.4±1.7 | 55.7±0.5 | 71.4±0.31 | 62.5±0.3 | N/A | N/A |
| PK | 76.0±2.7 | 59.5±2.4 | 73.7±0.7 | 82.5±0.5 | N/A | N/A |
| WL kernel | 90.4±5.7 | 59.9±4.3 | 75.0±3.1 | 86.0±1.8 | 73.8±3.9 | 81.0±3.1 |
| DCNN | N/A | N/A | 61.3±1.6 | 56.6±1.0 | 49.1±1.4 | N/A |
| DGCNN | 85.8±1.8 | 58.6±2.5 | 75.5±0.9 | 74.4±0.5 | 70.0±0.9 | N/A |
| IGN | 83.9±13.0 | 58.5±6.9 | 76.6±5.5 | 74.3±2.7 | 72.0±5.5 | N/A |
| GIN | 89.4±5.6 | 64.6±7.0 | 76.2±2.8 | 82.7±1.7 | 75.1±5.1 | 92.4±2.5 |
| PPGNs | 9 0.6±8.7 | 66.2±6.6 | 77.2±4.7 | 83.2±1.1 | 73.0±5.8 | N/A |
| Natural GN | 89.4±1.6 | 66.8±1.7 | 71.7±1.0 | 82.4±1.3 | 73.5±2.0 | N/A |
| GSN | 92.2 ± 7.5 | 68.2 ± 7.2 | 76.6± 5.0 | 83.5 ± 2.0 | 77.8 ± 3.3 | N/A |
| SIN | N/A | N/A | 76.4 ± 3.3 | 82.7 ± 2.1 | 75.6 ± 3.2 | 92.2 ± 1.0 |
| CIN | 92.7 ± 6.1 | 68.2 ± 5.6 | 77.0 ± 4.3 | 83.6 ± 1.4 | 75.6 ± 3.7 | 92.4 ± 2.1 |
| GIN-AK+ | 91.3 ± 7.0 | 67.7 ± 8.8 | 77.1 ± 5.7 | 85.0 ± 2.0 | 75.0 ± 4.2 | 94.8 ± 0.8 |
| Feature | Train | Val | Test | Total |
| # Repositories | 19 | 14 | 14 | 47 |
| #Files | 2655 | 1060 | 1308 | 4757 |
| #Holes | 92721 | 48548 | 48288 | 189557 |
| Data Split | SR Codex(%) | SR Oracle(%) | Rel.个 over Codex(%) |
| Train | 59.78 | 80.29 | 34.31 |
| Val | 62.10 | 79.05 | 27.28 |
| Test | 58.73 | 79.63 | 35.58 |
| Method | Success Rate(%) (hole-wise) | Rel. ↑(%) (hole-wise) | Success Rate(%) (repo-wise) | Rel.↑(%) (repo-wise) |
| Codex (Chen et al.,2021) | 58.73 | - | 60.64 | = |
| Oracle | 79.63 | 35.58 | 80.24 | 32.31 |
| Random | 58.13 | -1.02 | 58.95 | -2.79 |
| Random NN | 58.98 | 0.43 | 60.04 | -0.99 |
| File-level BM25 | 63.14 | 7.51 | 64.28 | 6.00 |
| Identifier Usage (Random) | 64.93 | 10.55 | 67.83 | 11.85 |
| Identifier Usage (NN) | 64.91 | 10.52 | 67.94 | 12.03 |
| Fixed Prompt Proposal | 65.78 | 12.00 | 68.01 | 12.15 |
| RLPG-BM25 | 66.41 | 13.07 | 68.15 | 12.39 |
| RLPG-H | 68.51 | 16.65 | 69.26 | 14.21 |
| RLPG-R | 67.80 | 15.44 | 69.28 | 14.26 |
| Prompt Source | File Ranking |
| Current Parent Class | file with the target hole.Returns a single file. file that contains the parent class that occurs closest to the target hole. |
| Import | Returns a single file. files with the corresponding import usage ranked based on the proximity |
| Sibling | to the hole.Returns a ranked list of files. files with import usage common to the current file and the sibling file, |
| Similar Name | ranked based on the proximity to the hole.The total number of common imports between the current and the sibling file is used as a tie-breaker. Returnsa ranked list of files. files with import usage common to the current file and the similar name |
| Child Class | file,ranked based on the proximity to the hole.The total number of common imports between the current and the similar name file is used as a tie-breaker.Returns a ranked list of files. files with import usage common to the current file and the child file, |
| ranked based on the proximity to the hole.The total number of common imports between the current and the child class file is used as a tie-breaker. Returns a ranked list of files. | |
| Import of Sibling | import files ranked based on the frequency of usage in all the sibling files. Returns a ranked list of files. |
| Import of Similar Name | import file ranked on the basis of frequency of usage in all the similar name files.Returns a ranked list of files. |
| Import of Parent Class | import file ranked on the basis of frequency of usage in all the parent |
| Importof Child Class | class files.Returns a ranked list of fles. import file ranked on the basis of frequency of usage in all the child class |
| Prompt Proposal ID | Prompt Source | Prompt Context Type |
| 0,1,2,3,4 | Current | MN, I, TI, SL,FD |
| 5,6,7 | Current | PL (taking 25%,50% and 75% contribution to the total context length) |
| 8,9,10,11,12,13 | Parent Class | MNB,MN,I, TI, SL,FD |
| 14,15,16,17,18,19 | Import | MNB,MN,I,TI, SL,FD |
| 20,21,22,23,24,25 | Sibling | MNB,MN,I, TI, SL,FD |
| 26,27,28,29,30,31 | Similar Name | MNB,MN,I,TI, SL,FD |
| 32,33,34,35,36,37 | Child Class | MNB,MN,I, TI, SL,FD |
| 38,39,40,41,42,43 | Import of Sibling | MNB,MN,I,TI,SL,FD |
| 44,45,46,47,48,49 | Import of Similar Name | MNB,MN,I,TI, SL,FD |
| 50,51,52,53,54,55 | Import of Parent Class | MNB,MN,I, TI, SL,FD |
| 56,57,58,59,60,61 | Import of Child Class | MNB,MN,I,TI, SL,FD |
| 62 | Codex |
| Method | Success Rate(%) (hole-wise) | Rel. ↑(%) (hole-wise) | Success Rate(%) (repo-wise) | Rel. ↑(%) (repo-wise) |
| Codex (Chen et al., 2021) | 58.73 | 1 | 60.64 | 1 |
| Post Lines (immediate line after the hole) | 65.78 | 12.00 | 68.01 | 12.15 |
| Post Lines (skipping three lines after the hole) | 65.11 | 10.86 | 66.42 | 9.53 |
| Method | Success Rate(%) (hole-wise) | Rel. ↑(%) (hole-wise) | Success Rate(%) (repo-wise) | Rel. ↑(%) (repo-wise) |
| Codex (Chen et al.,2021) | 58.73 | 60.64 | ||
| RLPG-H(𝑙 = 1) | 68.51 | 16.65 | 69.26 | 14.21 |
| RLPG-R (𝑙 = 1) | 67.80 | 15.44 | 69.28 | 14.26 |
| RLPG-H(l = 2) | 67.07 | 14.20 | 67.87 | 11.91 |
| RLPG-R ( = 2) | 66.57 | 13.35 | 67.88 | 11.94 |
| RLPG-H(l = 5) | 66.60 | 13.40 | 67.91 | 11.98 |
| RLPG-R (l = 5) | 65.78 | 12.01 | 67.69 | 11.62 |
| RLPG-H(l = 10) | 65.53 | 11.58 | 67.24 | 10.88 |
| RLPG-R (l = 10) | 63.59 | 8.27 | 65.98 | 8.79 |
| Method | Success Rate(%) (hole-wise) | Rel. ↑(%) (hole-wise) | Success Rate(%) (repo-wise) | Rel. ↑(%) (repo-wise) |
| Codex (Chen et al., 2021) | 57.77 | 1 | 58.90 | 1 |
| Oracle | 61.90 | 7.15 | 67.18 | 14.07 |
| Repo name | #Total Holes | Oracle | Codex | Fixed prompt proposal | RLPG-H | RLPG-R |
| largemail | 1653 | 75.38 | 55.11 | 62.73 | 63.94 | 63.28 |
| ftpserverremoteadmin | 7323 | 86.44 | 66.11 | 76.09 | 76.21 | 76.76 |
| myt5lib | 838 | 91.65 | 53.58 | 61.34 | 73.51 | 74.46 |
| seamlets | 4890 | 92.74 | 62.25 | 62.72 | 71.55 | 74.27 |
| gloodb | 10000 | 91.07 | 57.50 | 57.50 | 70.32 | 72.31 |
| jjskit | 9043 | 80.36 | 65.61 | 72.18 | 72.00 | 72.44 |
| mobileexpensetracker | 2298 | 75.94 | 57.88 | 67.28 | 66.84 | 66.97 |
| gfsfa | 10000 | 80.55 | 57.33 | 57.33 | 59.28 | 65.24 |
| swe574-group3 | 2029 | 76.79 | 54.46 | 66.19 | 65.16 | 64.91 |
| strudem-sicsa | 6131 | 77.83 | 64.96 | 72.55 | 73.25 | 73.32 |
| soap-dtc | 1370 | 81.24 | 64.82 | 70.73 | 71.61 | 72.70 |
| openprocesslogger | 7191 | 81.06 | 62.19 | 71.77 | 72.22 | 72.62 |
| tapestry-sesame | 397 | 72.54 | 45.84 | 61.21 | 60.71 | 63.98 |
| exogdx | 735 | 84.76 | 63.81 | 75.51 | 75.92 | 76.60 |
| designpatternjavapedro | 1069 | 78.30 | 54.82 | 64.36 | 63.99 | 68.57 |
| quidsee | 3020 | 81.66 | 60.79 | 69.50 | 70.36 | 70.26 |
| healpix-rangeset | 4734 | 63.54 | 48.71 | 54.67 | 54.94 | 55.07 |
| sol-agent-platform | 10000 | 73.76 | 58.22 | 65.72 | 65.65 | 65.94 |
| rsbotownversion | 10000 | 75.23 | 57.89 | 65.58 | 66.22 | 66.31 |
| Repo name | #Total Holes | Oracle | Codex | Fixed prompt proposal | RLPG-H | RLPG-R | |
| tyrondmath-mech-eshopinfinispan-storage-serviceteammates-shakthijavasummerframeworktinwiki jlooglejcontenedorsohocmsaffinity_propagation_javajata4testswinagilenavigablep2pspringlime | 721 | 83.91 | 60.33 | 71.15 | 71.57 | 72.6873.1777.7572.4665.55 | |
| 2225 | 83.46 | 62.20 | 72.76 | 73.53 | |||
| 373 | 82.31 | 71.85 | 78.55 | 76.94 | |||
| 7665 | 82.02 | 63.74 | 72.38 | 72.47 | |||
| 10000 | 79.27 | 55.92 | 65.30 | 65.74 | |||
| 10000 | 73.67 | 69.27 | 69.27 | 69.12 | 69.58 | ||
| 3145 | 84.55 | 73.16 | 77.87 | 77.17 | 77.3668.3267.62 | ||
| 5464772 | 81.26 | 58.99 | 67.77 | 67.95 | |||
| 76.68 | 57.90 | 67.10 | 67.49 | ||||
| 1466 | 79.54 | 59.14 | 70.33 | 70.26 | 70.2657.47 | ||
| 1921 | 71.06 | 44.09 | 54.92 | 55.91 | |||
| 2595 | 79.69 | 63.01 | 72.29 | 72.49 | 72.68 | ||
| 1322879 | 75.72 | 59.76 | 65.43 | 65.13 | 65.2874.40 | ||
| 83.50 | 62.34 | 74.18 | 74.86 | ||||
| Repo Name | #Total Holes | Oracle | Codex | Fixed PP | RLPG-H | RLPG-R | Random | Random NN | Iden Usage (Random) | Iden Usage (NN) | File-Level BM25 | RLPG- BM25 |
| dovetaildb | 10000 | 76.89 | 57.12 | 66.45 | 66.06 | 66.25 | 57.45 | 57.58 | 61.39 | 60.77 | 59.39 | 66.09 |
| project-pt-diaoc | 10000 | 82.01 | 52.67 | 52.81 | 65.08 | 61.25 | 51.58 | 52.93 | 55.54 | 56.21 | 57.04 | 58.29 |
| realtimegc | 2513 | 77.64 | 57.58 | 67.01 | 67.85 | 68.48 | 57.78 | 58.89 | 63.51 | 63.99 | 61.84 | 66.69 |
| fswuniceubtemplates | 2070 | 77.44 | 55.7 | 58.89 | 66.81 | 65.8 | 55.22 | 55.89 | 65.7 | 66.43 | 59.28 | 66.71 |
| qwikioffice-java | 1138 | 76.45 | 70.21 | 70.21 | 69.86 | 70.56 | 46.13 | 48.15 | 60.37 | 62.92 | 64.41 | 58.17 |
| glperaudsimon | 1766 | 78.65 | 53.57 | 62.51 | 62.4 | 61.66 | 55.66 | 57.76 | 69.42 | 68.4 | 69.14 | 61.55 |
| xiaonei-java-api | 839 | 73.42 | 57.57 | 62.1 | 62.69 | 63.29 | 57.09 | 57.21 | 71.28 | 72.35 | 63.77 | 63.29 |
| ircrpgbot | 6591 | 83.67 | 69.67 | 77.24 | 76.71 | 76.65 | 69.55 | 70.54 | 74.68 | 74.43 | 69.32 | 75.75 |
| robotsimulator2009w | 7514 | 75.63 | 56.28 | 67.55 | 67.53 | 67.55 | 56.4 | 56.18 | 64.61 | 64.71 | 62.96 | 66.12 |
| gwt-plugindetect | 73 | 84.93 | 60.27 | 68.49 | 65.75 | 68.49 | 58.9 | 57.53 | 63.01 | 63.01 | 50.68 | 75.34 |
| apiitfriends | 1385 | 85.05 | 65.05 | 74.8 | 75.67 | 75.31 | 65.7 | 68.59 | 70.25 | 70.11 | 66.93 | 73.57 |
| wicketbits | 754 | 83.02 | 59.81 | 72.94 | 72.81 | 73.08 | 60.21 | 61.94 | 81.96 | 79.31 | 84.48 | 73.47 |
| hucourses | 590 | 84.41 | 70.68 | 77.46 | 77.63 | 77.97 | 70 | 72.2 | 70.68 | 72.54 | 53.39 | 75.08 |
| xfuze | 3055 | 84.09 | 62.82 | 73.62 | 72.73 | 73.62 | 63.67 | 65.17 | 77.25 | 75.97 | 77.32 | 74.01 |
| ·RLPG | = | ring FileName(){ | VS | Codex |
| ring GetAlgorithmFilename(){ |
| Method | Data Aug. | M=5 | M=20 | M=100 | Train TFLOPs | Mem. (Mb) | |||
| AAA | Acc | AAA | Acc | AAA | Acc | ||||
| iid | X | 62.7±0.7 72.9±0.7 | 1 | 62.7±0.7 72.9±0.7 | - | 62.7±0.7 72.9±0.7 | 8 16 | 4 4 | |
| jid+± DER++ | 50.7±1.1 | 31.8±0.9 | 55.6±1.2 | 39.3±1.0 | 60.1±1.3 | 52.3±1.1 | 24 | (4,7) | |
| ER | 40.0±0.8 45.6±1.1 | 19.7±0.3 28.4±1.0 | 45.2±1.3 55.9±1.2 | 26.7±1.0 40.3±0.6 | 55.4±1.4 60.3±1.3 | 38.7±0.8 49.4±1.3 | 17 | (4,7) | |
| iCaRLt | 47.0±0.8 49.1±1.0 | 30.6±0.8 33.4±1.0 | 55.1±0.7 54.4±0.7 | 41.7±0.6 39.2±0.8 | 59.3±0.6 56.9±0.7 | 45.1±0.6 42.3±0.8 | (21,47) | (8,11) | |
| MIRt | X | 39.3±1.0 44.9±0.9 | 19.7±0.5 29.8±0.8 | 44.7±1.1 49.7±1.0 | 29.7±0.6 41.8±0.6 | 53.8±1.7 54.6±1.4 | 43.3±1.0 49.3±0.6 | 41 | (4,7) |
| SS-IL+ | X √ | 42.6±1.7 41.1±1.6 | 29.6±0.4 31.6±0.5 | 44.8±1.8 47.0±1.2 | 35.1±0.9 38.3±0.4 | 48.1±2.2 48.1±1.7 | 41.1±0.4 47.5±0.7 | 19 | (8,11) |
| ER-ACE (ours) | 53.1±1.0 52.6±0.9 | 35.6±1.0 35.1±0.8 | 58.0±0.7 56.4±1.0 | 42.6±0.7 43.4±1.6 | 61.9±0.9 61.7±0.9 | 52.2±0.7 53.7±1.1 | 17 | (4,7) | |
| ER-AML | X | 49.4±1.0 | 30.9±0.8 | 57.0±1.0 | 39.2±1.0 | 63.3±1.0 | 52.2±1.1 | 17 | (4,7) |
| (ours) GDUMB | 50.4±1.3 0±0.0 | 36.4±1.4 35.0±0.6 | 56.8±1.0 0±0.0 | 47.7±0.7 45.8±0.9 | 62.0±0.9 0±0.0 | 55.7±1.3 61.3±1.7 | (43,853) | (11,14) | |
| Method | AAA | Acc. | Train TFLOPs | Mem. (Mb.) | AAA | Acc. | Train TFLOPs | Mem. (Mb.) |
| iid | 19.8±0.3 | 9 | 4 | 16.7±0.5 | 59 | 4 | ||
| iid++ | 28.3±0.3 | 17 | 4 | 25.0±0.8 | 118 | 4 | ||
| DER++ | 23.3±0.5 | 15.1±0.4 | 25 | 36 | 21.7±0.6 | 12.9±0.3 | 176 | 217 |
| ER | 24.2±0.6 | 19.8±0.4 | 17 | 35 | 26.2±0.8 | 18.2±0.5 | 118 | 216 |
| iCaRLt | 26.3±0.3 | 17.3±0.2 | 294 | 39 | 24.4±0.4 | 17.1±0.1 | 2097 | 220 |
| MIRt | 23.6±0.8 | 20.6±0.5 | 41 | 35 | 27.2±0.7 | 20.2±0.8 | 294 | 216 |
| SS-ILt | 31.5±0.5 | 25.0±0.3 | 19 | 39 | 29.7±0.6 | 23.5±0.5 | 137 | 220 |
| ER-ACE (ours) | 32.7±0.5 | 25.8±0.4 | 17 | 35 | 30.2±0.6 | 22.7±0.6 | 118 | 216 |
| ER-AML (ours) | 30.2±0.6 | 24.3±0.4 | 28 | 35 | 27.0±0.7 | 19.3±0.6 | 200 | 216 |
| Method | M= 20 | M=100 |
| ER | 32.1±1.5 | 42.7±2.2 |
| DER++ | 31.0±1.4 | 41.7±1.4 |
| ER-AML | 45.6±1.2 | 55.2±1.1 |
| ER-ACE | 44.5±0.5 | 50.2±1.1 |
| ER | (3.2 ± 1.8) × 10-2 |
| ER-AML-Tripletw.All Negs | (3.0±0.6)× 10-2 |
| ER-AML-Triplet w. Incoming Negs | (2.5 ±0.6)× 10-² |
| Accuracy ↑ | ||||
| M=5 | M=20 | M=50 | M=100 | |
| iid online | 60.8±1.0 | 60.8±1.0 | 60.8 ± 1.0 | 60.8±1.0 |
| iid++ online | 72.0 ± 0.1 | 72.0±0.1 | 72.0±0.1 | 72.0±0.1 |
| iid offline | 79.2 ± 0.4 | 79.2 ± 0.4 | 79.2 ± 0.4 | 79.2 ± 0.4 |
| fine-tuning | 18.4± 0.3 | 18.4± 0.3 | 18.4±0.3 | 18.4±0.3 |
| ER | 19.0 ± 0.1 | 26.7± 0.3 | 36.1 ± 0.6 | 41.5 ± 0.6 |
| ER-AML Triplet | 33.0 ± 0.3 | 40.1± 0.4 | 46.0± 0.5 | 49.8 ± 0.5 |
| ER-AML SupCon 33.0 ±0.2 | 41.9 ± 0.1 | 48.3±0.2 | 51.9 ±0.3 | |
| Accuracy (↑is better) M=20 M=50 | Forgetting (↓is better) M=20 M=50 | |
| ER | 26.7±0.3 36.1±0.6 | 47.1± 0.8 37.6± 0.9 |
| ER-AML(all negatives) | 28.5± 0.3 41.4± 0.4 | 56.7± 0.6 35.0±0.4 |
| ER-AML(incoming negatives) | 41.9 ± 0.1 48.3±0.2 | 33.6±0.2 25.8±0.3 |
| Method | Avg. unique classes per MB 1 2 3 4 5 |
| ER 23.1 DER++ 20.3 31.1 | 25.7 26.3 31.1 34.4 31.4 37.3 34.4 |
| ER-ACE 32.8 | 36.2 36.8 41.7 44.5 |
| ER-AML 34.0 | 40.4 46.0 47.6 47.9 |
| Method | % ofData Used 5% 10% 25% 50% |
| ER 17.3 | 22.5 28.0 33.2 32.8 |
| DER++ 17.4 SS-IL 15.2 | 19.9 24.8 21.7 28.6 |
| 31.9 | |
| ER-ACE 20.5 ER-AML 18.1 | 25.4 31.2 36.1 24.3 31.0 38.7 |
| Dataset | Model | Image similarity ↑ | Text Similarity ↑ |
| Cat | Baseline | 0.768 | 0.292 |
| Fixed background | 0.724 | 0.281 | |
| Fixed pose | 0.752 | 0.278 | |
| No prompt templates | 0.722 | 0.282 | |
| Sculpture | Baseline | 0.806 | 0.278 |
| Fixed background | 0.786 | 0.263 | |
| Fixed pose | 0.799 | 0.262 | |
| No prompt templates | 0.604 | 0.287 |
| Model | FID↓ | Diversity ↑ | ||
| AFHQ Cat | Face Sketch | AFHQ Cat | Face Sketch | |
| CDC (Ojha et al., 2021) | 45.13 | 72.74 | 0.52 | 0.45 |
| MineGAN (Wang et al., 2020) | 79.31 | 62.27 | 0.21 | 0.40 |
| TGAN (Wang et al., 2018) | 87.11 | 69.44 | 0.28 | 0.39 |
| TGAN + ADA (Karras et al., 2020) | 52.70 | 56.76 | 0.34 | 0.41 |
| Textual Inversion | 79.72 | 75.61 | 0.62 | 0.53 |
| Type | Model | N-M | ddk | H MACs Param | BLEU | COMET-22 | |||
| Multi-Branch Architecture | Weighted Transformer (Ahmed et al.,2017) | 6-6 | 1024 | 211M | 28.90 | ||||
| Multi-Unit Transformer (Yan et al.,2020) | 6-6 | 130M | 29.30 | ||||||
| MAT (Fan et al., 2020) | 6-6 | 206M | 29.90 | ||||||
| Multi-Path Transformer (Lin et al., 2022) | 6-6 | 193M | 29.68 | ||||||
| Lightweight Architecture | Evolved Transformer (So et al., 2019) Delight (Mehta et al., 2021) | 64M | 28.20 | ||||||
| Weight Sharing | 640 | 54M | 28.00 | ||||||
| Universal Transformer (Dehghani et al.,2019) | 1024 | = | 65M | 28.90 | |||||
| SubFormer (Reid et al., 2021) | - | = | 63M | 28.50 | |||||
| SubFormer-big (Reid et al., 2021) | 197M | 29.30 | |||||||
| ODE Transformer (RK4) (Li et al., 2022) ODE Transformer (RK4) (Li et al., 2022) | 6-6 24-6 | 512 512 | = | 62M 118M | 29.03 29.80 | ||||
| Other Comparisons | 64 | ||||||||
| RealFormer (He et al., 2021) DMAN (Fan et al., 2021) | 18-18 | 512 512 | 8 8 | 151M | 29.35 | ||||
| Mega-Softmax (Ma et al.,2022) | 6-6 6-6 | 512 | 63M 67M | 29.10 29.01 | |||||
| Our System | Transformer | 64 | 1 | ||||||
| PartialFormer (w/o Head Scaling) | 24-6 | 512 | 8-8 | 11.1B | 118M | 29.05 | 83.60 | ||
| PartialFormer | 24-6 | 512 | 64 24-16 | 8-8 | 8.8B | 66M | 28.86 30.09 | 83.35 84.17 | |
| 24-6 | 512 | 64 | 12.2B | 115M | |||||
| Transformer | 6-6 | 512 | 64 45 | 8-8 | 9.9B | 62M | 27.43 | 82.19 | |
| Transformer PartialFormer (w/o Head Scaling) | 24-6 24-6 | 360 360 | 45 | 8-8 8-8 | 6.3B 5.2B | 62M 36M | 28.00 27.88 | 82.72 82.49 | |
| 360 | 45 | 24-16 | 6.8B | 61M | 29.23 | ||||
| PartialFormer PartialFormer | 24-6 24-6 | 360 | 45 30-16 | 6.9B | 68M | 29.56 | 83.74 83.94 |
| Model | N | ddk | HParam BLEU |
| Weighted Transformer (2017) | 6 | - 211M 41.40 | |
| Evolved Transformer (2019) | = | 64M 40.60 | |
| Delight (2021) | -640 | 54M 40.50 | |
| ODE Transformer (2022) | 6 | = | 69M 42.56 |
| ODE Transformer (2022) | 24 | 123M 43.28 | |
| Multi-Path Transformer (2022) | = | 168M 42.44 | |
| Transformer | 24 512 64 | 8-8 | 120M 42.33 |
| PartialFormer (w/o Head Scaling) 24 512 648-8 | 68M 41.68 | ||
| PartialFormer | 24 512 64 24-18 | 119M 43.10 | |
| PartialFormer | 24 512 64 24-24 | 127M 43.29 | |
| Transformer | 6 512 64 | 8-8 | 63M 40.79 |
| Transformer | 24 360 45 | 8-8 | 64M 40.96 |
| PartialFormer(w/o Head Scaling) 24 360 45 | 8-8 | 38M 40.44 | |
| PartialFormer | 24 360 45 24-18 | 63M 42.16 | |
| PartialFormer | 24 360 45 24-24 | 67M 42.39 |
| Model | N | ddk | H Param BLEU |
| Delight (Mehta et al., 2021) | - 640- | ■ | 53M 34.70 |
| Subformer (Reid et al.,2021) | ■ | - ■ - | 48M 34.70 |
| ODE Transformer (Li et al.,2022) | 6 1024 64 16-16 | 226M 35.28 | |
| Transformer | 24 512 64 | 8-8 | 111M 35.00 |
| PartialFormer (w/o Head Scaling) | 24 512 | 64 8-8 | 59M 35.07 |
| PartialFormer | 24 | 320 4024-24 | 48M 35.30 |
| Model | Fi←→En | De←→En | Lv← →En | Avg. | |||
| Fi→En En→FiDe→En En-→DeLv-→En En→Lv | |||||||
| Transformer | 26.07 | 22.14 | 35.04 | 28.59 | 17.59 | 16.23 | 24.27 |
| PartialFormer | 27.48 | 23.35 | 35.60 | 29.91 | 19.65 | 17.37 | 25.56 |
| # Model | Param BLEU |
| 1 Transformer (N = 24,d = 360) | 62M 28.00 |
| 2 Pure Attention (N= 24,d = 360) | 31M 25.70 |
| 3 PartialFormer | 68M 29.56 |
| 4 w/o Partial-level Gated FFN | 52M 27.51 |
| 5 w/o Residual-like Attention Calculation | 66M 29.26 |
| 6 w/o Head Scaling | 36M 27.88 |
| 7 PartialFormer (encoder only) | 67M 29.15 |
| 8 PartialFormer (decoder only) | 63M 28.80 |
| Model | Param BLEU |
| PartialFormer (w/o Head Scaling) | 36M 27.88 |
| + Simple Head Scaling | 68M 29.33 |
| + Complex Head Scaling | 68M 29.56 |
| Model | Seting | H d | dk Param | BLEU |
| PartialFormer | Basic | |30-16 360 | 45 68M | 29.56 |
| Varying Encoder H | |24-16 360 45 16-16 360 45 | 61M 51M | 29.23 29.02 | |
| Varying Decoder H | [16-24 360 45 16-30360 45 | 56M 60M | 28.85 29.20 | |
| |30-16 360 30 | 49M | 28.70 | ||
| Varying dh | 30-16 360 60 30-16 360 90 | 86M 124M | 29.68 | |
| 30.00 | ||||
| Varying d | |30-16 180 45 30-16 270 45 30-16 450 45 | 35M 51M 84M | 27.61 28.80 29.41 |
| Model | Param BLEU |
| PG-FFNs | 68M 29.56 |
| PG-FFNs with Sigmoid activation | 68M 29.21 |
| PG-FFNs with Tanh activation | 68M 29.03 |
| Model | Param Speed (Tok./s)Memory BLEU | |||
| Transformer | 62M | 4325 | 3.0G | 28.00 |
| PartialFormer (w/o head scaling) | 66M | 3634 | 3.2G | 28.86 |
| PartialFormer | 68M | 3023 | 3.3G | 29.56 |
| Dataset | Sentence | BPE | Vocab | ||
| Train | Dev | Test | |||
| WMT'14 En-De | 4.5M | 2999 26815 | 3003 | 32K 32K | 34040 37288 |
| WMT'14 En-Fr WMT'16 En-Ro | 36M 0.6M | 1999 | 3003 1999 | 20K | 19064 |
| WMT'17 En-De | 5.9M | 7998 | 3004 | 32K | 35488 |
| WMT'17 De-En | 5.9M | 7998 | 3004 | 32K | 35448 |
| WMT'17 En-Fi | 2.7M | 4225 | 3002 | 32K | 32584 |
| WMT'17Fi-En | 2.7M | 4225 | 3002 | 32K | 32584 |
| WMT'17 En-Lv WMT'17Lv-En | 4.5M 4.5M | 2003 2003 | 2001 2001 | 20K 20K | 32368 32368 |
| Hyper-parameter WMT'14 En-De WMT'16En-Ro WMT'14 En-Fr | |||
| GPUs | 8 | 4 | 8 |
| Batch Size | 4096 | 4096 | 4096 |
| Update Frequency | 2 | 1 | 8 |
| Optimer | Adam | Adam | Adam |
| Adamβ | (0.9,0.997) | (0.9, 0.997) | (0.9, 0.997) |
| LR | 0.0020 | 0.0020 | 0.0020 |
| LR scheduler | inverse sqrt | inverse sqrt | inverse sqrt |
| InitialLR | 1e-7 | 1e-7 | le-7 |
| Total updates | 50K | 25K | 100K |
| Warmup updates | 16000 | 8000 | 16000 |
| Weight decay | 0.0000 | 0.0000 | 0.0000 |
| Label smoothing | 0.1 | 0.1 | 0.1 |
| Dropout | 0.1 | 0.1 | 0.1 |
| Attention dropout | 0.1 | 0.1 | 0.1 |
| ReLU dropout | 0.1 | 0.1 | 0.1 |
| Hyper-parameterI En-{De,Lv} {De,Lv}-En | En-Fi | Fi-En | ||
| GPUs | 8 | 8 | 8 | 8 |
| Batch Size | 4096 | 4096 | 4096 | 4096 |
| Update Frequency | 2 | 1 | 1 | 4 |
| Optimer | Adam | Adam | Adam | Adam |
| Adamβ | (0.9, 0.997) | (0.9, 0.997) | (0.9,0.997) (0.9,0.997) | |
| LR | 0.0020 | 0.0020 | 0.0020 | 0.0020 |
| LR scheduler | inverse sqrt | inverse sqrt | inverse sqrt | inverse sqrt |
| Initial LR | 1e-7 | 1e-7 | 1e-7 | le-7 |
| Total updates | 50K/17K | 50K/17K | 40K | 10K |
| Warmup updates | 16000 | 16000 | 16000 | 16000 |
| Weight decay | 0.0000 | 0.0000 | 0.0000 | 0.0000 |
| Label smoothing | 0.1 | 0.1 | 0.1 | 0.1 |
| Dropout | 0.1 | 0.1 | 0.1 | 0.1 |
| Attention dropout | 0.1 | 0.1 | 0.1 | 0.1 |
| ReLU dropout | 0.1 | 0.1 | 0.1 | 0.1 |
| Model | Param BLEU |
| DELIGHT (Mehta et al., 2021) EdgeFormer (Ge et al., 2022) Lite Transformer (Wu et al.,2020) PartialFormer | 23M 26.70 - 26.90 - 26.50 27M 27.50 |
| Evolved Transformer (So et al., 2019) DELIGHT (Mehta et al., 2021) ODE Transformer (Li et al., 2022) PartialFormer | 48M 27.70 37M 27.60 37M 28.24 36M 28.35 |
| AG | AL | Param | BLEU |
| RPR | MHSA | 62M | 35.76 |
| Model | Param | BLEU |
| PartialFormer | 67M | 29.56 |
| PartialFormer + Weight Sharing | 67M | 29.71 |
| GLU-based PartialFormer | 67M | 29.67 |
| Model | Depth 0 (M) Test PPL |
| Adaptive Input | 8 147M 21.11 |
| PartialFormer | 16 143M 19.87 |
| Prompt | Accuracy |
| What is this piece of news regarding? | 40.9 |
| What is this article about? | 52.4 |
| What is the best way to describe this article? | 68.2 |
| What is the most accurate label for this news article? | 71.2 |
| Meta prompts |
| Write a paraphrase for the following sentence: <seed-prompt> Paraphrase: <seed-prompt> Paraphrase: |
| Write a likely paraphrase of the text: <seed-prompt> Paraphrase: |
| Write a sentence similar to the following one: <seed-prompt> Paraphrase: |
| Paraphrase the following sentence: <seed-prompt> Paraphrase: |
| Write a variation of this sentence: <seed-prompt> How would you say the following sentence in a different way? <seed-prompt> |
| Task | # Step 0 | # Step 1 # Step 2 |
| Word-Level Translation | 12 59 | 118 |
| Antonyms | 12 85 | 176 |
| GLUE Cola | 4 27 | 144 |
| Newspop | 13 43 | 119 |
| AG News | 4 23 | 108 |
| IMDB | 10 45 | 178 |
| DBpedia | 8 23 | 103 |
| Emotion | 4 14 | 94 |
| Tweet Offensive | 5 41 | 119 |
| All Manually Created Prompts | Examples of Similar Automatically Created Prompts |
| What label best describes this news article? | What's the most accurate label for this news article? |
| What is this piece of news regarding? Which newspaper section would this article likely appear in? | What does this piece of news concern? |
| What topic is this news article about? | In what section of the newspaper could this article be published? What category does this article fall into? |
| Model | Task | Perplexity-score corr. Pearson Spearman | Perplexity-acc corr. Pearson Spearman | Avg Acc | Acc 50% | ||
| OPT175B | Antonyms | **-0.41 | **-0.53 | ||||
| GLUE Cola | -0.15 | -0.14 | -0.04 | 1 -0.02 | 1 47.7 | 1 57.1 | |
| Newspop | *-0.24 | **-0.26 | *-0.20 | -0.18 | 66.4 | 72.9 | |
| AG News | **-0.63 | **-0.68 | **-0.77 | **-0.81 | 57.5 | 68.7 | |
| IMDB | **0.35 | **0.40 | 0.14 | *0.20 | 86.2 | 91.0 | |
| DBpedia | **-0.50 | **-0.44 | **-0.51 | **-0.42 | 46.7 | 55.2 | |
| Emotion | -0.14 | -0.19 | **-0.30 | **-0.32 | 16.4 | 23.0 | |
| Tweet Offensive | *-0.19 | 0.07 | 0.18 | *0.23 | 51.3 | 55.8 | |
| Bloom 176B | Antonyms | **-0.37 | **-0.23 | 1 | 1 | ||
| GLUE Cola | 0.07 | 0.11 | **-0.25 | **-0.26 | 55.5 | 65.6 | |
| Newspop | **-0.50 | **-0.42 | **-0.59 | **-0.51 | 78.9 | 87.8 | |
| AG News | **-0.62 | **-0.54 | **-0.44 | **-0.44 | 50.3 | 59.4 | |
| IMDB | 0.04 | 0.09 | -0.08 | -0.14 | 89.3 | 92.2 | |
| DBpedia | **-0.47 | *-0.27 | **-0.35 | *-0.21 | 27.2 | 33.4 | |
| Emotion | **-0.33 | **-0.42 | **-0.48 | **-0.55 | 29.3 | 31.7 | |
| Tweet Offensive | 0.14 | *0.24 | *-0.20 | -0.03 | 41.6 | 46.2 | |
| OPT 30B | Antonyms | **-0.54 | **-0.70 | 1 | 1 | 1 | 1 |
| GLUE Cola | -0.05 | 0.03 | -0.13 | 0.02 | 32.2 | 35.5 | |
| Newspop | *-0.23 | *-0.25 | *-0.18 | -0.12 | 60.3 | 66.6 | |
| AG News | **-0.66 | **-0.71 | **-0.81 | **-0.80 | 49.3 | 60.7 | |
| IMDB | -0.06 | *0.17 | 0.04 | **0.22 | 81.6 | 86.1 | |
| DBpedia | **-0.41 | **-0.34 | *-0.21 | *-0.25 | 35.9 | 42.4 | |
| Emotion | 0.00 | -0.03 | 0.18 | 0.13 | 12.3 | 16.2 | |
| Tweet Offensive | **-0.44 | **-0.39 | -0.11 | -0.05 | 54.6 | 60.2 | |
| OPT 1.3B | Antonyms | **-0.45 | **-0.53 | 1 | |||
| GLUE Cola | **-0.39 | **-0.36 | -0.09 | *-0.19 | 60.3 | 1 65.9 | |
| Newspop | **0.33 | *0.21 | -0.07 | -0.07 | 37.6 | 40.3 | |
| AG News | **-0.33 | **-0.29 | **-0.56 | **-0.49 | 31.9 | 37.6 | |
| IMDB | -0.11 | -0.07 | **0.24 | **0.22 | 86.0 | 89.1 | |
| DBpedia | -0.16 | -0.14 | -0.02 | -0.01 | 8.7 | 9.2 | |
| Emotion | 0.08 | 0.08 | **-0.29 | **-0.30 | 7.0 | 9.1 | |
| Tweet Offensive | **-0.42 | **-0.35 | **-0.50 | **-0.38 | 58.6 | 62.6 | |
| Lang | OPT175B | Bloom 176B Pearson | Spear. | |
| Pearson | Spear. | |||
| ita | -0.44 | -0.57 | -0.37 | -0.63 |
| spa | -0.47 | -0.61 | -0.51 | -0.66 |
| cat | -0.47 | -0.58 | -0.24 | -0.31 |
| fra | -0.48 | -0.57 | -0.48 | -0.64 |
| deu | -0.44 | -0.60 | -0.46 | -0.65 |
| fin | -0.44 | -0.62 | -0.34 | -0.56 |
| por | -0.45 | -0.62 | -0.46 | -0.61 |
| eus | -0.47 | -0.61 | -0.45 | -0.61 |
| tur | -0.44 | -0.62 | -0.33 | -0.62 |
| jpn | -0.33 | -0.26 | ||
| arb | -0.36 | -0.47 | ||
| rus | 1 | -0.54 | -0.69 | |
| kor | -0.42 | -0.58 | ||
| ell | -0.40 | -0.51 | ||
| prompt | ppl |
| Is this example correct English usage? | 25.79 |
| Is this example using English correctly? | 25.46 |
| Is this example correct English? | 25.33 |
| Is this the example in correct English? | 25.00 |
| Is English in this example correct? | 24.90 |
| Task | Lang | Before filtering | After filtering | ||
| Pearson | Spearman | Pearson | Spearman | ||
| AG News | - | -0.63 | -0.68 | -0.62 | -0.54 |
| WLT | ita | -0.44 | -0.58 | -0.44 | -0.57 |
| spa | -0.47 | -0.61 | -0.47 | -0.61 | |
| cat | -0.45 | -0.57 | -0.47 | -0.58 | |
| fra | -0.47 | -0.57 | -0.48 | -0.57 | |
| deu | -0.43 | -0.60 | -0.44 | -0.60 | |
| fin | -0.41 | -0.60 | -0.44 | -0.62 | |
| por | -0.43 | -0.61 | -0.45 | -0.62 | |
| eus | -0.45 | -0.60 | -0.47 | -0.61 | |
| tur | -0.43 | -0.61 | -0.44 | -0.62 | |
| prompt | ppl |
| The following two words are antonyms:“good"and “ | 10.24 |
| The antonym of the word“good’is“ | 10.32 |
| The word that has the opposite meaning of the word “good"is “ | 10.43 |
| The word“good’is the antithesis of the word “ | 10.85 |
| The word “good"is the opposite of the word “ | 11.15 |
| OPT | Bloom | |||||
| Task | low-ppl | manual | △ | low-ppl | manual | △ |
| GLUE Cola | 51.7 | 48.5 | 3.1 | 64.5 | 60.9 | 3.6 |
| Newspop | 80.6 | 70.4 | 10.2 | 90.0 | 80.0 | 10.0 |
| AG News | 68.4 | 61.9 | 6.5 | 51.0 | 63.5 | -12.5 |
| IMDB | 90.4 | 88.9 | 1.4 | 91.3 | 88.8 | 2.5 |
| DBpedia | 46.0 | 51.7 | -5.7 | 31.2 | 30.2 | 1.0 |
| Emotion | 21.6 | 22.6 | -1.1 | 35.8 | 32.1 | 3.6 |
| Tweet Offensive | 48.4 | 50.6 | -2.3 | 48.6 | 40.8 | 7.8 |
| Task Manual Prompts | |
| Antonyms | The antonym of the word “good”is“ The opposite meaning of the word “good”is “ “Good”is the opposite of“ “Good”is the negation of“ The following are opposites of each other:“good"and “ The word “good"contradicts the word “ The antonym of the word good is The opposite meaning of the word good is Good is the opposite of |
| GLUE Cola | The following are opposites of each other: good and The word good contradicts the word Does the this sentence make sense and use correct English? Is this example grammatically correct and sensible? Does this sentence make sense and is it grammatically correct? |
| Newspop | Does this example use correct English? What is the article about? What is this news about? What is the topic of this news piece? What does this article discuss? What is the topic of this sentence? What category does the article belong to? Pick one category forthis news piece. Pick the category that fits the text. The article refers to which category? What topic does the article belong to? What category fits this article? What topic does this news piece belong to? Choose the correct category for this article. |
| AG News | What label best describes this news article? What is this piece of news regarding? Which newspaper section would this article likely appear in? What topic is this news article about? This movie review expresses what sentiment? Did the reviewer find this movie good or bad? |
| IMDB | Is this review positive or negative? How does the viewer feel about the movie? What sentiment does the writer express for the movie? What sentiment is expressed for the movie? What is the sentiment expressed in this text? Did the reviewer enjoy the movie? What is the sentiment expressed by the reviewer for the movie? How does the reviewer feel about the movie? What category does the paragraph belong to? Pick one category for the text. |
| DBpedia | Pick the category that fits the text. The text refers to which category? What category does the title belong to? What category fits this text? What topic does this text belong to? Choose the correct category for the text. What is the emotion expressed in this message? |
| Emotion | What emotion does this message express? How will you feel about the message? What emotion does the writer express for the message? Is this tweet offensive? Can the tweet be removed for being offensive? |
| Tweet Offensive | Is the author's tweet offensive? Task: Identify if the tweetor text is offensive. Is this an offensive tweet? The translation of the word “dog"to French is“ The translation of the word dog to French is The word “dog”in French is“ |
| Word-Level Translation | “dog”(In French: “ Translate the word dog into French: The translation of dog to French is “dog”(French: “ The word dog in French is Translate the word“dog”into French: “ dog (In French: dog (French: The translation of“dog”to French is “ |
| Task | Lowest Perplexity Prompts | Perplexity |
| Antonyms | The following two words are antonyms:“good”and “ | 10.24 |
| The antonym of the word“good"is“ | 10.32 | |
| The word that has the opposite meaning of the word “good" is “ | 10.43 | |
| The word“good”is the antithesis of the word“ | 10.85 | |
| The word “good” is the opposite of the word “ | 11.15 | |
| GLUE Cola | Is this an example of the proper use of the English language? | 11.63 |
| Does the sentence make sense and does it follow the rules of grammar? | 11.76 | |
| Is this sentence an example of the correct use of the English language? | 12.10 | |
| Does this sentence make sense and is it grammatically correct? | 12.15 | |
| Is this sentence grammatically correct and does it make sense? | 12.68 | |
| Newspop | What is the main subject of the article? | 10.01 |
| What is the main topic of the article? | 10.01 | |
| What is the subject matter of the article? | 10.17 | |
| What is the subject of the article? | 10.21 | |
| What is the main idea of this article? | 10.21 | |
| AG News | In what section of the newspaper would you expect to find this article? | 7.51 |
| In which section of the newspaper would you expect to find this article? | 7.52 | |
| In which section of the newspaper would this article be most likely to appear? | 7.60 | |
| In what section of the newspaper do you expect to find this article? | 7.80 | |
| In what section of the newspaper would this article most likely appear? | 7.87 | |
| IMDB | What is the opinion of the review? Is it positive or negative? | 7.19 |
| Is this a positive or negative review? | 7.31 | |
| What do you think of the movie? | 7.33 | |
| What do you think of the film? | 7.35 | |
| Is that a positive or a negative? | 7.35 | |
| DBpedia | What is the category to which the text refers? | 8.99 |
| What is the subject of the text? | 9.15 | |
| What category does the title belong to? | 9.18 | |
| Which category does the text refer to? | 9.19 | |
| What is the subject of this text? | 9.20 | |
| Emotion | How do you feel when you hear this message? | 12.72 |
| What is the writer's emotional reaction to this news? | 13.18 | |
| What is the emotion expressed in this message? | 13.20 | |
| How does this message make you feel? | 13.32 | |
| How do you feel about this message? | 13.50 | |
| Tweet Offensive | If someone said this to you, would you be offended? | 13.00 |
| If someone said that to you, would you be offended? | 13.10 | |
| Would you be offended if someone said that to you? | 13.73 | |
| Would it offend you if someone said that to you? | 14.79 | |
| If someone told you that, would you be offended? | 14.93 | |
| Word-Level Translation | The word for“dog"in French is“ | 7.73 |
| The French word for“dog”is“ | 8.16 | |
| The French translation of the word“dog”is“ | 8.24 | |
| The translation of the word “dog”in French is “ | 8.35 | |
| The translation of the word “dog” into French is “ | 8.91 |
| Model | Type | Params(M) | GMACs | Train. Epoch | Top-1(%) | Latency (ms) | ||
| NPU | CPU | A100 | ||||||
| MobileNetV2×1.0 | CONV | 3.5 | 0.3 | 300 | 71.8 | 1.3 | 8.0 | 5.0 |
| MobileNetV2×1.4 | CONV | 6.1 | 0.6 | 300 | 74.7 | 1.6 | 10.7 | 7.3 |
| ResNet50 | CONV | 25.5 | 4.1 | 300 | 78.5 | 3.0 | 29.4 | 9.0 |
| EfficientNet-B0 | CONV | 5.3 | 0.4 | 350 | 77.1 | 2.7 | 14.5 | 10.0 |
| EfficientNet-B3 | CONV | 12.0 | 1.8 | 350 | 81.6 | 6.6 | 52.6 | 35.0 |
| EffcientNet-B5 | CONV | 30.0 | 9.9 | 350 | 83.6 | 23.0 | 258.8 | 141.0 |
| DeiT-T | Attention | 5.9 | 1.2 | 300/1000 | 74.5/76.6 | 9.2 | 16.7 | 7.1 |
| DeiT-S | Attention | 22.5 | 4.5 | 300/1000 | 81.2/82.6 | 11.8 | 41.0 | 15.5 |
| PVT-Small | Attention | 24.5 | 3.8 | 300 | 79.8 | 24.4 | 89.5 | 23.8 |
| T2T-ViT-14 | Attention | 21.5 | 4.8 | 310 | 81.5 | - | - | 21.0 |
| Swin-Tiny | Attention | 29 | 4.5 | 300 | 81.3 | - | ■ | 22.0 |
| CSwin-T | Attention | 23 | 4.3 | 300 | 82.7 | - | - | 28.7 |
| PoolFormer-s12 | Pool | 12 | 2.0 | 300 | 77.2 | 6.1 | 59.0 | 14.5 |
| PoolFormer-s24 | Pool | 21 | 3.6 | 300 | 80.3 | 6.2 | 126.7 | 28.2 |
| PoolFormer-s36 | Pool | 31 | 5.2 | 300 | 81.4 | 6.7 | 192.6 | 41.2 |
| ResMLP-S24 | SMLP | 30 | 6.0 | 300 | 79.4 | 7.6 | 40.2 | 17.4 |
| Convmixer-768 | Hybrid | 21.1 | 20.7 | 300 | 80.2 | 11.6 | 29.3 | - |
| LeViT-256 | Hybrid | 18.9 | 1.1 | 1000 | 81.6 | 11.9 † | 13.5 | 4.5 |
| NASViT-A5 | Hybrid | - | 0.76 | 360 | 81.8 | - | ■ | - |
| MobileViT-XS | Hybrid | 2.3 | 0.7 | 300 | 74.8 | 7.2 | 26.5 | 11.7 |
| MobileFormer-508M | Hybrid | 14.0 | 0.51 | 450 | 79.3 | 13.2 | 22.2 | 14.6 |
| EfficientFormer-L1 | MetaBlock | 12.3 | 1.3 | 300/1000 | 79.2/80.2 | 1.6 | 11.5 | 6.2 |
| EfficientFormer-L3 | MetaBlock | 31.3 | 3.9 | 300 | 82.4 | 3.0 | 28.2 | 13.9 |
| EfficientFormer-L7 | MetaBlock | 82.1 | 10.2 | 300 | 83.3 | 7.0 | 67.7 | 30.7 |
| Backbone | Detection & Instance Segmentation | Semantic mIoU(%) | |||||
| Apbox | AP5 | APb | APmask | APmask | |||
| ResNet18 | 34.0 | 54.0 | 36.7 | 31.2 | 51.0 | APmask 32.7 | 32.9 |
| PoolFormer-S12 | 37.3 | 59.0 | 40.1 | 34.6 | 55.8 | 36.9 | 37.2 |
| EfficientFormer-L1 | 37.9 | 60.3 | 41.0 | 35.4 | 57.3 | 37.3 | 38.9 |
| ResNet50 | 38.0 | 58.6 | 41.4 | 34.4 | 55.1 | 36.7 | 36.7 |
| PoolFormer-S24 | 40.1 | 62.2 | 43.4 | 37.0 | 59.1 | 39.6 | 40.3 |
| EfficientFormer-L3 | 41.4 | 63.9 | 44.7 | 38.1 | 61.0 | 40.4 | 43.5 |
| ResNet101 | 40.4 | 61.1 | 44.2 | 36.4 | 57.7 | 38.8 | 38.8 |
| PoolFormer-S36 | 41.0 | 63.1 | 44.8 | 37.7 | 60.1 | 40.0 | 42.0 |
| EfficientFormer-L7 | 42.6 | 65.1 | 46.1 | 39.0 | 62.2 | 41.7 | 45.1 |
| Method | Animals | Landscapes | User Study | ||||||
| SFID↓ | CSFID↓ | LPIPS↓ | SFID↓ | CSFID↓ | LPIPS↓ | Text个 | Realism↑ | Content | |
| VQGAN-CLIP | 30.01 | 65.51 | 0.462 | 33.31 | 82.92 | 0.571 | 2.78 | 2.05 | 2.16 |
| CLIP-GD | 12.50 | 53.05 | 0.468 | 18.13 | 62.19 | 0.458 | 2.61 | 2.24 | 2.28 |
| DiffusionCLIP | 25.09 | 66.50 | 0.379 | 29.85 | 76.29 | 0.568 | 2.50 | 2.54 | 3.06 |
| FlexIT | 32.71 | 57.87 | 0.215 | 18.04 | 60.04 | 0.243 | 2.22 | 3.15 | 3.89 |
| Ours | 9.98 | 41.07 | 0.372 | 16.86 | 54.48 | 0.417 | 3.68 | 4.28 | 4.11 |
| Method | Style↑ | Realism↑ | Content个 |
| SANet | 2.75 | 4.08 | 4.37 |
| WCT2 | 2.59 | 4.64 | 4.90 |
| STROTSS | 3.92 | 2.91 | 3.17 |
| SplicingViT | 3.50 | 2.08 | 2.15 |
| Ours | 4.23 | 4.25 | 4.51 |
| Settings | Animals | ||
| SFID↓ | CSFID↓ | LPIPS↓ | |
| VGG instead of ViT (g) | 9.72 | 43.08 | 0.518 |
| No resampling (f) | 11.88 | 59.09 | 0.316 |
| L2 Max instead of lsem (e) | 13.18 | 49.47 | 0.324 |
| LPIPs instead of lcont (d) | 11.15 | 58.67 | 0.400 |
| No lcont (c) | 9.90 | 33.07 | 0.477 |
| No lsem (b) | 15.00 | 53.43 | 0.347 |
| Ours (a) | 9.98 | 41.07 | 0.372 |
| Ours | Splicing Vit | STROTSS | WCT2 | SANET | |
| time | 37s | 25m 30s | 53s | 0.18s | 0.12s |
| Question: Amy is 1O years old. Jake is 8 years old. Alex's age is right in the middle. How old is Alex? Selected Chain-of-Thought: Amy is 1O years old.Jake is 8 years old.Alex's age is in the middle of Amy and Jake,so Alex is(8+10)/2=9 years old. The answer is 9. |
| Mixed-formats of training data: Format 1: Input: [CoT prompting examples] +‘\n’ + [Question] +‘\n’+‘A:' Output: Amy is 10 years old. Jake is 8 years old. Alex's age is in the middle of Amy and Jake, so Alex |
| is (8 + 10)/2=9 years old. The answer is 9. Format 2: Input: [Standard prompting examples] +‘\n'+ [Question] +‘\n’ +‘A:' |
| Output: The answer is 9. Format 3: Input: [Question] +‘\n’ +‘A: Let's think step by step.' |
| Output: Amy is 10 years old. Jake is 8 years old. Alex's age is in the middle of Amy and Jake,so Alex |
| is (8 + 10)/2=9 years old. The answer is 9. |
| Format 4: Input: [Question] +‘\n’ +‘A:' Output: The answer is 9. |
| Prompting Method | GSM8K | DROP | ARC-c | OpenBookQA | ANLI-A2 | ANLI-A3 | |
| Previous SOTA | 82.3a | 84.9b | 88.7c | 91.0d | 64.9d | 66.0d | |
| w/o LMSI | Standard-Prompting | 17.9 | 60.0 | 87.1 | 84.4 | 55.8 | 55.8 |
| CoT-Prompting | 56.5 | 70.6 | 85.2 | 86.4 | 58.9 | 60.6 | |
| Self-Consistency | 74.4 | 78.2 | 88.7 | 90.0 | 64.5 | 63.4 | |
| LMSI | Standard-Prompting | 32.2 | 71.7 | 87.2 | 92.0 | 64.8 | 66.9 |
| CoT-Prompting | 73.5 | 76.2 | 88.3 | 93.0 | 65.3 | 67.3 | |
| Self-Consistency | 82.1 | 83.0 | 89.8 | 94.4 | 66.5 | 67.9 |
| Self-training data | AQUA | SVAMP | StrategyQA | ANLI-A1 | RTE | MNLI-M/MM | |
| w/o LMSI | - | 35.8 | 79.0 | 75.3 | 68.8 | 79.1 | 72.0/74.0 |
| LMSI | GSM8K + DROP +... | 39.0 | 82.8 | 77.8 | 79.2 | 80.1 | 81.8/82.2 |
| Results on GSM8K | |
| Standard Prompting CoT Prompting | |
| w/o LMSI | 17.9 56.5 |
| LMSI w/o CoT formats 23.6 | 61.6 |
| LMSI | 32.2 73.5 |
| Questions used for Self-Training | Results on GSM8K CoT-Prompting Self-Consistency | ||
| w/o LMSI | = | 56.5 | 74.4 |
| LMSI | Generated Questions | 66.2 | 78.1 |
| LMSI | Training-set Questions | 73.5 | 82.1 |
| Results on GSM8K | |||
| 8 billion | 62 billion | 540 billion | |
| w/o LMSI | 5.0 | 29.7 | 56.5 |
| Distilled from LMSI 540 billion | 33.4 | 57.4 | 1 |
| Prompting Method | GSM8K | DROP | ARC-c | OpenBookQA | ANLI-A2 | ANLI-A3 | |
| w/o LMSI | CoT-Prompting Self-Consistency | 5.4/7.1 6.4/9.9 | 11.1/16.8 16.8/26.5 | 49.9 54.9 | 53.6 54.0 | 35.9 37.4 | 33.8 36.8 |
| LMSI | CoT-Prompting Self-Consistency | 6.1/8.6 7.9/10.2 | 11.4/17.1 18.1/28.1 | 50.9 54.7 | 53.8 55.2 | 35.4 38.1 | 34.4 37.4 |
| Ablation Study on Different Components in ACM-SGC and ACM-GCN (%) | |||||||||||||
| Baseline | ModelComponents | Cornell | Wisconsin | Texas | Film | Chameleon | Squirrel | Cora | CiteSeer | PubMed | Rank | ||
| Models | LP HP Identity Mixing| | Acc ± Std | Acc ± Std | Acc ± Std | Acc ± Std | Acc ± Std | Acc ± Std | Acc ± Std | Acc ± Std | Acc ± Std | |||
| 70.98±8.39 70.38±2.85 83.28±5.43 25.26±1.18 64.86±1.8147.62±1.2785.12±1.64 79.66±0.7585.5±0.76 | 12.89 | ||||||||||||
| ACM-SGC-1 w/ | √ | √ √ | 83.28±5.8191.88±1.6190.98±2.46 36.76±1.0165.27±1.947.27±1.3786.8±1.0880.98±1.68 87.21±0.42 | 10.44 | |||||||||
| √ √ | √ | 93.93±3.695.25±1.84 93.93±2.54 38.38±1.13 63.83±2.0746.79±0.7586.73±1.2880.57±0.9987.8±0.58 | 9.44 | ||||||||||
| √ √ √ | √ | 88.2 ±4.3993.5±2.95 92.95±2.94 37.19±0.87 62.82±1.84 4.94±0.93 85.22±1.35 80.75±1.68 88.11±0.21 | 11.00 | ||||||||||
| √ | √ | 93.77±1.9193.25±2.9293.61±1.5539.33±1.2563.68±1.6246.4±1.1386.63±1.1380.96±0.9387.75±0.88 | 10.00 | ||||||||||
| ACM-GCN w/ | √ | 82.46±3.11 75.5±2.9283.11±3.2 35.51±0.99 64.18±2.62 44.76±1.39 87.78±0.96 81.39±1.2388.9±0.32 | 11.44 | ||||||||||
| √ √ | √ | 82.13 ±2.59 86.62±4.6189.19 ±3.04 38.06±1.3569.21±1.6857.2±1.018.93±1.5581.96±0.9190.01±0.8 | 7.22 | ||||||||||
| √ √ | √ | √ | 94.26±2.2396.13±2.294.1±2.9541.51±0.99 67.44±2.14 53.97±1.3988.95±0.981.72±1.22 90.88±0.55 | 4.44 | |||||||||
| √ | √ | 91.64±295.37±3.3195.25±2.3740.47±1.49 68.93±2.04 54.78±1.2789.13±1.7781.96±2.0391.01±0.7 | 3.11 | ||||||||||
| √ √ | √ | √ | 94.75±2.6296.75±1.695.08±3.241.62±1.15 69.04±1.7458.02±1.8688.95±1.381.80±1.2690.69±0.53 | 2.78 | |||||||||
| ACMII-GCN w/ | √ √ | √ | 82.46±3.03 91.00±1.7590.33±2.69 38.39±0.75 67.59±2.1453.67±1.7189.13±1.1481.75±0.85 89.87±0.39 | 7.44 | |||||||||
| √ | √ | √ | 94.26±2.57 96.00±2.15 94.26 ±2.96 40.96±1.2 66.35±1.76 50.78±2.0789.06±1.0781.86±1.22 90.71±0.67 | 4.67 | |||||||||
| √ √ | √ | 91.48±1.43 96.25±2.09 93.77±2.9140.27±1.076.52±2.65 52.9±1.6488.83±1.1681.54±0.9590.6±0.47 | 6.67 | ||||||||||
| √ √ | √ | √ | 95.9±1.8396.62±2.4495.25±3.1541.84±1.1568.38±1.36 54.53±2.0989.00±0.7281.79±0.9590.74±0.5 | 2.78 | |||||||||
| Comparison of Average Running Time Per Epoch(ms) | |||||||||||||
| 2.53 | 2.83 | 3.47 | 3.43 | ||||||||||
| √ √ √ | √ | 4.01 | 4.57 | 2.5 4.24 | 3.18 4.55 | 3.48 4.76 | 4.65 5.09 | 5.39 | 4.69 | 4.04 4.75 | |||
| ACM-SGC-1 w/ | √ | √ | √ | 3.88 | 4.01 | 4.04 | 4.43 | 4.06 | 4.5 | 4.38 | 3.82 | 4.16 | |
| √ √ | √ | 3.31 | 3.49 | 3.18 | 3.7 | 3.53 | 4.83 | 3.92 | 3.87 | 4.24 | |||
| √ √ | √ | √ | 5.53 | 5.96 | 5.43 | 5.21 | 5.41 | 6.96 | 6 | 5.9 | 6.04 | ||
| 3.67 | 3.74 | 3.59 | 4.18 | 5.08 | |||||||||
| √ √ √ | √ | 6.63 | 8.06 | 7.89 | 4.86 8.11 | 4.96 7.8 | 6.41 9.39 | 4.24 7.82 | 7.38 | 8.74 6.8 | |||
| ACM-GCN w/ | √ | √ | √ | 5.73 | 5.91 | 5.93 | 6.86 | 6.35 | 7.15 | 7.34 | 6.65 | 6.16 | |
| √ √ | √ | 5.16 | 5.25 | 5.2 | 5.93 | 5.64 | 8.02 | 5.73 | 5.65 | ||||
| √ √ | √ | √ | 8.25 | 8.11 | 7.89 | 7.97 | 8.41 | 11.9 | 8.84 | 8.38 | |||
| √ √ | √ | 6.62 | 7.35 | 7.39 | 7.62 | 7.33 | 9.69 | 7.49 | 7.58 | ||||
| ACMII-GCNw/ | √ | √ | √ | 6.3 | 6.05 | 6.26 | 6.87 | 6.44 | 6.5 | 6.14 | 7.21 | 7.97 6.6 6.33 | |
| √ √ √ √ | √ √ | √ | 5.24 7.59 | 5.27 8.28 | 5.46 8.06 | 5.72 8.85 | 5.65 8 | 7.87 10 | 5.48 8.27 | 5.65 8.5 | 8.68 | ||
| Methods | Real-Objects Single→Fused | Real-Characters Single-→Fused | Real-Scenes Single-→Fused | Mean Change | |
| Text-alignment | Upper Bound | 0.811 | 0.767 | 0.834 | 0.804 |
| P+ [4] | 0.771-→0.771 (-) | 0.686-→0.686(-) | 0.759-→0.759 (-) | 0.739-→0.739 (-) | |
| Custom Diffusion [11] | 0.745->0.747(+0.002) | 0.674-0.650 (-0.024) | 0.748-→0.738 (-0.010) | 0.722-0.712(-0.010) | |
| LoRA [3] | 0.720->0.795 (+0.075) | 0.654->0.700 (+0.046) | 0.717->0.760 (+0.043) | 0.697-→0.752(+0.055) | |
| Mix-of-Show (Ours) | 0.724-→0.745 (+0.021) | 0.632-→0.662(+0.030) | 0.716-→0.736(+0.020) | 0.691-0.714(+0.024) | |
| Image-alignment | Lower Bound | 0.721 | 0.471 | 0.595 | 0.596 |
| P+ [4] | 0.790→0.790 (-) | 0.670-→0.670 (-) | 0.796→0.796 (-) | 0.752-→0.752(-) | |
| Custom Diffusion [11] | 0.842-→0.808 (-0.034) | 0.714-→0.694 (-0.020) | 0.804-→0.750 (-0.054) | 0.787-→0.751(-0.036) | |
| LoRA [3] | 0.864-→0.778 (-0.086) | 0.761-→0.555(-0.206) | 0.824-→0.769 (-0.055) | 0.816-→0.701 (-0.115) | |
| Mix-of-Show (Ours) | 0.868-→0.846 (-0.022) | 0.802->0.770 (-0.032) | 0.858-→0.838 (-0.020) | 0.843-→0.818(-0.025) |
| Methods | Real-Objects Single-→Fused | Real-Characters Single-→Fused | Real-Scenes Single-→Fused | Mean Change | |
| Image-alignment | Lower Bound | 0.721 | 0.471 | 0.595 | 0.596 |
| LoRA+ Weight Fusion | 0.864- 0.778 (-0.086) | 0.761-0.555(-0.206) | 0.824->0.769 (-0.055) | 0.816-→0.701 (-0.115) | |
| ED-LoRA + Weight Fusion | 0.868-→0.798 (-0.070) | 0.802-→0.634 (-0.168) | 0.858->0.816 (-0.042) | 0.843->0.749 (-0.094) | |
| ED-LoRA + Gradient Fusions | 0.868-→0.846(-0.022) | 0.802-→0.770 (-0.032) | 0.858-→0.838 (-0.020) | 0.843-0.818 (-0.025) |
| Human Evaluation | Image Alignment | Text Alignment |
| ED-LoRA+Weight Fusion | 33.5% | 47.5% |
| ED-LoRA + Gradient Fusion | 66.5% | 52.5% |
| Templates in PIOI |
| Then, [B] and [A] went to the [PLACE]. [B] gave a [OBJECT] to [A] |
| Then,[B] and [A] had a lot of fun at the [PLACE]. [B] gave a [OBJECT] to [A] |
| Then, [B] and [A] were working at the [PLACE]. [B] decided to give a [OBJECT] to [A] |
| Then, [B] and [A] were thinking about going to the [PLACE]. [B] wanted to give a [OBJECT] to [A] |
| Then, [B] and [A] had a long argument, and afterwards [B] said to [A] |
| After [B] and [A] went to the [PLACE], [B] gave a [OBJECT] to [A] |
| When [B] and [A] got a [OBJECT] at the [PLACE], [B] decided to give it to [A] |
| When [B] and [A] got a [OBJECT] at the [PLACE], [B] decided to give the [OBJECT] to [A] |
| While [B] and [A] were working at the [PLACE], [B] gave a [OBJECT] to [A] |
| While [B] and [A] were commuting to the [PLACE], [B] gave a [OBJECT] to [A] |
| After the lunch, [B] and [A] went to the [PLACE]. [B] gave a [OBJECT] to [A] |
| Afterwards, [B] and [A] went to the [PLACE]. [B] gave a [OBJECT] to [A] |
| Then, [B] and [A] had a long argument. Afterwards [B] said to [A] |
| The [PLACE] [B] and [A] went to had a [OBJECT]. [B] gave it to [A] |
| Friends [B] and [A] found a [OBJECT] at the [PLACE]. [B] gave it to [A] |
| Distribution | Logit difference | IO probability | Proportion ofS logit greater than IO |
| PIOI | 3.55 | 0.49 | 0.7% |
| Additional occurrence of S(natural sentence) | 3.64 | 0.59 | 0.4% |
| Additional occurrence of IO(natural sentence) | 1.23 | 0.36 | 23.4% |
| [A] had a good day. |
| [A] was enjoying the situation. |
| [A] was tired. |
| [A] enjoyed being with a friend. |
| [A] was an enthusiast person. |
| Original positional signal | Inverted position signal | |
| Original S token signal | 3.55 (baseline) | -0.99 |
| Random S token signal | 2.45 | -1.96 |
| S←→IO inverted token signal | 1.77 | -3.16 |
| U | Class | KU{u} | F(C\(KU{u) | F(C\K) |
| (9,9) | Name Mover | [(9,9)] | 2.78 | 3.14 |
| (10,0) | Name Mover | [(9,9), (10,0)] | 2.43 | 2.78 |
| (9,6 | Name Mover | [(9,9),(10,0),(9,6)] | 2.77 | 2.43 |
| (10,7) | Negative Name Mover | All Negative Name Mover Heads | 5.11 | 3.84 |
| (11,10) | Negative Name Mover | All Negative Name Mover Heads | 5.11 | 4.06 |
| (7,3) | S-Inhibition | All S-Inhibition Heads | 0.33 | 1.15 |
| (7,9) | S-Inhibition | All S-Inhibition Heads | 0.33 | 1.12 |
| (8,6) | S-Inhibition | All S-Inhibition Heads | 0.33 | 1.10 |
| (8,10) | S-Inhibition | All S-InhibitionHeads | 0.33 | 0.55 |
| (5,5) | Induction | Induction Headsand NegativeHeads | 1.06 | 3.95 |
| (5,8) | Induction | All Induction Heads | 1.06 | 2.58 |
| (5,9) | Induction | All Induction Heads | 4.40 | 5.11 |
| (6,9) | Induction | Induction Heads and Negative Heads | 4.76 | 5.11 |
| (0,1) | Duplicate Token | All Duplicate Token Heads | 1.14 | 2.52 |
| (0,10) | Duplicate Token | AllDuplicate Token Heads | 1.14 | 2.29 |
| (3,0 | Duplicate Token | All Duplicate Token Heads | 1.14 | 1.65 |
| (2,2) | Previous Token | All Previous Token Heads | 2.03 | 2.80 |
| (2,9) | Previous Token | All Previous Token Heads | 2.03 | 2.42 |
| (4, 11) | Previous Token | All Previous Token Heads | 2.03 | 2.27 |
| (10,10) | Backup Name Mover | All NMs and previous Backup NMs | 2.40 | 2.63 |
| (10,2) | Backup Name Mover | All NMs and previous Backup NMs | 0.89 | 1.09 |
| (11,2) | Backup Name Mover | All NMs and previous Backup NMs | 0.72 | 0.89 |
| (10,6) | Backup Name Mover | All NMs and previous Backup NMs | 2.63 | 2.77 |
| (10,1) | Backup Name Mover | All NMs and previous Backup NMs | 1.34 | 1.47 |
| (9,7) | Backup Name Mover | All NMs and previous Backup NMs | 0.85 | 1.02 |
| (11,9) | Backup Name Mover | All NMs and previous Backup NMs | 1.02 | 1.13 |
| (11,3) | Backup Name Mover | [(9,9),(10,0),(9,6),(10,10),(11,3)] | 2.53 | 2.59 |
| K found by greedy optimization |
| (9,9),(9,6), (5,8), (5,5), (2,2), (2,9) |
| (9,9),(11,10),(10,7),(8,6),(5,8),(4,11) |
| (10,7), (5,5), (2,2), (4,11) |
| (9,9),(11,10),(10,7),(11,2),(3,0),,(5,8),(2,2) |
| Network | Arch. | Task | Datasets | Annotated | #Units |
| AlexNet | CNN | Class. | ImageNet Places365 | conv. 1-5 | 1152 1376 |
| ResNet152 | CNN | Class. | ImageNet Places365 | conv. 1 res. 1-4 | 3904 3904 |
| BigGAN | CNN | Gen. | ImageNet Places365 | res.0-5 | 3744 4992 |
| DINO | ViT | BYOL | ImageNet | MLP 1-12 (first 100) | 1200 |
| Model | CE | ND | p(d|E) | pmi(d; E) |
| AlexNet-ImageNet | .01 | .24 | .34 | .38 |
| AlexNet-Places | .02 | .21 | .31 | .37 |
| ResNet-ImageNet | .01 | .25 | .27 | .35 |
| ResNet-Places | .03 | .22 | .30 | .31 |
| Generalization | Train + Test | BERTScore (f) | |
| within network | AlexNet-ImageNet | .39 | |
| AlexNet-Places | .47 | ||
| ResNet152-ImageNet | .35 | ||
| ResNet152-Places | .28 | ||
| BigGAN-ImageNet | .49 | ||
| BigGAN-Places | .52 | ||
| Train Test | |||
| across arch. | AlexNet | ResNet152 | .28 |
| ResNet152 | AlexNet | .35 | |
| CNNs | ViT | .34 | |
| across datasets | ImageNet | Places | .30 |
| Places | ImageNet | .33 | |
| Classifiers | BigGAN | .34 | |
| across tasks | BigGAN | Classifiers | .27 |
| Model | Dataset | IAA |
| AlexNet | ImageNet | .25 |
| Places365 | .27 | |
| ResNet152 | ImageNet | .21 |
| Places365 | .17 | |
| BigGAN | ImageNet | .26 |
| Places365 | .24 | |
| DINO | ImageNet | .23 |
| Model | Layer | #Units | #Words | Len. | % Noun | % Adj | % Prep |
| AlexNet-ImageNet | conv1 | 64 | 185 | 4.8 | 37.5 | 24.3 | 12.2 |
| conv2 | 192 | 384 | 5.5 | 37.8 | 19.4 | 13.2 | |
| conv3 | 384 | 661 | 5.3 | 41.0 | 16.4 | 13.0 | |
| conv4 | 256 | 608 | 5.5 | 43.1 | 11.9 | 12.5 | |
| conv5 | 256 | 693 | 5.5 | 46.0 | 10.2 | 10.4 | |
| AlexNet-Places365 | conv1 | 96 | 153 | 4.3 | 38.4 | 26.8 | 12.7 |
| conv2 | 256 | 297 | 4.8 | 37.8 | 26.0 | 12.7 | |
| conv3 | 384 | 412 | 4.7 | 40.2 | 24.8 | 10.5 | |
| conv4 | 384 | 483 | 4.4 | 43.7 | 19.9 | 10.3 | |
| conv5 | 256 | 486 | 4.1 | 45.8 | 17.6 | 10.6 | |
| ResNet152-ImageNet | conv1 | 64 | 285 | 4.7 | 43.8 | 11.8 | 10.3 |
| layer1 | 256 | 653 | 5.5 | 43.1 | 10.5 | 12.5 | |
| layer2 | 512 | 936 | 5.1 | 44.0 | 12.7 | 12.6 | |
| layer3 | 1024 | 1222 | 4.2 | 49.6 | 10.9 | 11.3 | |
| layer4 | 2048 | 1728 | 4.6 | 47.8 | 8.6 | 7.8 | |
| ResNet152-Places365 | conv1 | 64 | 283 | 5.2 | 47.3 | 11.1 | 14.6 |
| layer1 | 256 | 633 | 5.3 | 46.3 | 9.4 | 13.3 | |
| layer2 | 512 | 986 | 5.8 | 46.0 | 8.3 | 13.8 | |
| layer3 | 1024 | 1389 | 4.8 | 48.2 | 6.7 | 12.7 | |
| layer4 | 2048 | 1970 | 5.3 | 46.3 | 5.5 | 11.9 | |
| BigGAN-ImageNet | layer0 | 1536 | 1147 | 3.9 | 52.4 | 7.8 | 8.2 |
| layer1 | 768 | 853 | 3.5 | 53.0 | 9.4 | 8.9 | |
| layer2 | 768 | 618 | 3.2 | 52.6 | 12.3 | 9.5 | |
| layer3 | 384 | 495 | 3.7 | 49.9 | 14.3 | 10.9 | |
| layer4 | 192 | 269 | 3.3 | 47.9 | 18.0 | 13.4 | |
| layer5 | 96 | 69 | 2.6 | 53.6 | 22.8 | 14.6 | |
| BigGAN-Places365 | layer0 | 2048 | 1062 | 4.2 | 53.3 | 5.4 | 8.3 |
| layer1 | 1024 | 708 | 3.9 | 55.0 | 6.1 | 11.5 | |
| layer2 | 1024 | 410 | 4.6 | 52.7 | 8.1 | 16.3 | |
| layer3 | 512 | 273 | 5.2 | 50.4 | 7.6 | 15.0 | |
| layer4 | 256 | 192 | 4.6 | 47.5 | 9.3 | 14.9 | |
| layer5 | 128 | 123 | 4.2 | 46.7 | 13.5 | 13.0 | |
| DINO-ImageNet | layer0 | 100 | 320 | 4.4 | 45.7 | 12.7 | 4.8 |
| layer1 | 100 | 321 | 4.2 | 49.8 | 9.1 | 6.8 | |
| layer2 | 100 | 285 | 3.9 | 53.3 | 6.2 | 7.5 | |
| layer3 | 100 | 312 | 3.9 | 54.4 | 6.2 | 7.1 | |
| layer4 | 100 | 304 | 3.9 | 53.5 | 4.4 | 7.0 | |
| layer5 | 100 | 287 | 3.5 | 55.1 | 5.5 | 5.2 | |
| layer6 | 100 | 377 | 3.9 | 51.3 | 8.2 | 5.4 | |
| layer7 | 100 | 374 | 3.8 | 52.0 | 6.4 | 6.2 | |
| layer8 | 100 | 330 | 3.4 | 53.0 | 7.0 | 8.8 | |
| layer9 | 100 | 350 | 3.1 | 56.1 | 6.3 | 9.6 | |
| layer10 | 100 | 369 | 3.9 | 50.3 | 9.3 | 8.2 | |
| layer11 | 100 | 294 | 3.3 | 52.4 | 7.5 | 9.4 | |
| Total | 20272 | 4597 | 4.5 | 48.7 | 9.4 | 10.9 |
| Gen. | Train + Test | # Units | #Words | Len. | % Noun | % Adj | %Prep | |
| within netwok | AlexNet-ImageNet | 115 | 100 | 3.5 | 45.7 | 16.4 | 11.9 | |
| AlexNet-Places | 137 | 46 | 2.5 | 49.3 | 28.7 | 9.6 | ||
| ResNet-ImageNet | 390 | 121 | 2.8 | 52.2 | 23.8 | 11.7 | ||
| ResNet-Places | 390 | 376 | 4.3 | 46.5 | 8.7 | 10.9 | ||
| BigGAN-ImageNet | 374 | 112 | 2.2 | 59.8 | 17.5 | 10.4 | ||
| BigGAN-Places | 499 | 245 | 3.8 | 54.2 | 6.0 | 9.0 | ||
| Train | Test | |||||||
| across arch. | AlexNet | ResNet | 7808 | 326 | 3.0 | 46.1 | 21.0 | 8.9 |
| ResNet | AlexNet | 2528 | 275 | 2.7 | 48.0 | 27.1 | 6.4 | |
| CNNs | ViT | 1200 | 200 | 2.6 | 55.0 | 18.2 | 13.0 | |
| across dataset | ImageNet | Places | 10272 | 271 | 2.2 | 58.8 | 14.0 | 13.8 |
| Places | ImageNet | 8800 | 309 | 3.1 | 47.8 | 26.9 | 7.8 | |
| across task | Classifiers | BigGAN | 8736 | 202 | 2.1 | 53.0 | 25.3 | 6.1 |
| BigGAN | Classifiers | 10336 | 336 | 3.2 | 54.3 | 14.2 | 16.8 | |
| Total | 51585 | 1002 | 2.7 | 51.9 | 19.8 | 11.1 | ||
| Backbone | Method | Input size | #Layer | #Param | FLOPs | Throughput | Top-1 Acc(%) |
| DeiT | DeiT-S | 224 | 12 | 22.0M | 4.57G | 1589.4 | 79.8 (79.9) |
| DeiT-S + AttnScale | 224 | 12 | 22.0M | 4.57G | 1416.7 | 80.7 (↑0.9) | |
| DeiT-S+FeatScale | 224 | 12 | 22.0M | 4.57G | 1509.9 | 80.9 (↑ 1.1) | |
| DeiT-S | 224 | 24 | 43.3M | 9.09G | 836.4 | 80.5 (81.0) | |
| DeiT-S + AttnScale | 224 | 24 | 43.3M | 9.10G | 722.0 | 81.1(↑0.6) | |
| CaiT | DeiT-S+FeatScale | 224 | 24 | 43.4M | 9.10G | 772.5 | 81.3(↑0.8) |
| CaiT-S | 224 | 24 | 46.9M | 9.33G | 371.9 | 82.6(82.7) | |
| CaiT-S + AttnScale | 224 | 24 | 46.9M | 9.34G | 339.0 | 83.2 (↑0.6) | |
| Swin | CaiT-S+FeatScale | 224 | 24 | 46.9M | 9.34G | 358.2 | 83.2 (↑0.6) |
| Swin-S | 224 | 24 | 49.6M | 8.74G | 593.2 | 83.0(83.0) | |
| Swin-S + AttnScale Swin-S +FeatScale | 224 224 | 24 24 | 49.6M 49.6M | 8.75G 8.75G | 553.4 550.3 | 83.4(↑0.4) 83.5 (↑0.5) |
| Category | Method | #Param | Input size | #Layer | Top-1 Acc (%) |
| CNN | ResNet-152 (He et al.,2016) | 230M | 224 | 152 | 78.1 * |
| DenseNet-201 (Huang et al.,2017) | 77M | 224 | 201 | 77.6 * | |
| CNN+ Transformer | CVT-21(Wu et al.,2021) LV-ViT-S (Jiang et al.,2021b) | 32M 26M | 224 224 | 21 | 82.5 * 83.3* |
| ViT-S/16 (Dosovitskiy et al.,2020) | 49M | 224 | 16 | 78.1 | |
| ViT-B/16 (Dosovitskiy et al., 2020) | 86M | 12 | |||
| DeiT-S (Touvron et al., 2021a) | 22M | 224 | 12 | 79.8 | |
| DeiT-S Distilled (Touvron et al.,2021a) | 22M | 224 | 12 | 79.8 | |
| Swin-S (Liu et al., 2021) | 50M | 224 | 12 | 81.2 | |
| T2T-ViT-24(Yuan et al.,2021) | 64M | 224 | 12 | 83.0 | |
| Deep ViT-24B (Zhou et al.,2021a) | 36M | 224 224 | 24 24 | 82.3 | |
| CaiT-S (Touvron et al.,2021b) | 80.1 | ||||
| DeiT-S + DiversePatch (Gong et al.,2021) | 47M | 224 | 24 | 82.7 | |
| DeiT-S+AttnScale | 44M | 224 | 24 | 82.2 | |
| Ours | DeiT-S+FeatScale | 43M 43M | 224 | 24 | 81.1 |
| CaiT-S + AttnScale | 47M | 224 | 24 | 81.3 | |
| CaiT-S+FeatScale | 47M | 224 224 | 24 | 83.2 | |
| Swin-S + AttnScale | 50M | 224 | 24 | 83.2 | |
| Swin-S +FeatScale | 50M | 224 | 24 24 | 83.4 83.5 |
| Backbone | Method | Input size | #Layer | #Param | FLOPs | Throughput | Top-1 Acc (%) |
| CaiT | CaiT-XXS | 224 | 24 | 12.0M | 2.53G | 589.3 | 77.5 (77.6) |
| CaiT-XXS + AttnScale | 224 | 24 | 12.0M | 2.53G | 548.1 | 77.8(↑0.3) | |
| CaiT-XXS+FeatScale | 224 | 24 | 12.0M | 2.53G | 573.5 | 77.8(↑0.3) | |
| CaiT-S | 224 | 24 | 46.9M | 8.74G | 371.9 | 82.6(82.7) | |
| CaiT-S + AttnScale | 224 | 24 | 46.9M | 8.75G | 339.0 | 82.8(↑0.2) | |
| CaiT-S+FeatScale | 224 | 24 | 46.9M | 8.75G | 358.2 | 82.9 (↑0.3) |
| Dataset | #Dialogues | #Utterances | Domains |
| ReDial | 10,006 | 182,150 | Movie |
| OpenDialKG | 13,802 | 91,209 | Movie, Book, Sports, Music |
| Datasets | ReDial | OpenDialKG | ||||
| Models | Recall@1 | Recall@10 | Recall@50 | Recall@ 1 | Recall@10 | Recall@ 25 |
| KBRD | 0.028 | 0.169 | 0.366 | 0.231 | 0.423 | 0.492 |
| KGSF | 0.039 | 0.183 | 0.378 | 0.119 | 0.436 | 0.523 |
| CRFR | 0.040 | 0.202 | 0.399 | 0.130 | 0.458 | 0.543 |
| BARCOR | 0.031 | 0.170 | 0.372 | 0.312 | 0.453 | 0.510 |
| UniCRS | 0.050 | 0.215 | 0.413 | 0.308 | 0.513 | 0.574 |
| MESE | 0.056* | 0.256* | 0.455* | 0.279 | 0.592* | 0.666* |
| text-embedding-ada-002 | 0.025 | 0.140 | 0.250 | 0.279 | 0.519 | 0.571 |
| ChatGPT | 0.034 | 0.172 | 1 | 0.105 | 0.264 | 1 |
| +MESE | 0.036 | 0.195 | 一 | 0.240 | 0.508 | |
| + text-embedding-ada-002 | 0.037 | 0.174 | 1 | 0.310 | 0.539 | 1 |
| Dataset | Irrelevant | Partially relevant | Highly relevant |
| ReDial | 8% | 20% | 72% |
| OpenDialKG | 20% | 16% | 64% |
| Setting | Single-turn | Multi-turn | ||
| Naturalness | Usefulness | Naturalness | Usefulness | |
| DialoGPT | 13% | 23% | 11% | 31% |
| iEvaLM | 36% | 43% | 55% | 38% |
| Tie | 51% | 34% | 34% | 31% |
| Human | 10% | 34% | 17% | 28% |
| iEvaLM | 39% | 33% | 35% | 40% |
| Tie | 51% | 33% | 48% | 32% |
| Model | KBRD | BARCOR | UniCRS | ChatGPT | |||||||||
| Evaluation Approach | Original | iEvaLM (attr) | iEvaLM (free) | Original | iEvaLM (attr) | iEvaLM (free) | Original | iEvaLM (attr) | iEvaLM (free) | Original | iEvaLM (attr) | iEvaLM (free) | |
| ReDial | R@1 | 0.028 | 0.039 (+39.3%) | 0.035 (+25.0%) | 0.031 | 0.034 (+9.7%) | 0.034 (+9.7%) | 0.050 | 0.053 (+6.0%) | 0.107 (+114.0%) | 0.037 | 0.191* (+416.2%) | 0.146 (+294.6%) |
| R@10 | 0.169 | 0.196 (+16.0%) | 0.198 (+17.2%) | 0.170 | 0.201 (+18.2%) | 0.190 (+11.8%) | 0.215 | 0.238 (+10.7%) | 0.317 (+47.4%) | 0.174 | 0.536* (+208.0%) | 0.440 (+152.9%) | |
| R@50 | 0.366 | 0.436 (+19.1%) | 0.453 (+23.8%) | 0.372 | 0.427 (+14.8%) | 0.467 (+25.5%) | 0.413 | 0.520 (+25.9%) | 0.602* (+45.8%) | 1 | 1 | 二 | |
| OpenDialKG | R@1 | 0.231 | 0.131 (-43.3%) | 0.234 (+1.3%) | 0.312 | 0.264 (-15.4%) | 0.314 (+0.6%) | 0.308 | 0.180 (-41.6%) | 0.314 (+1.9%) | 0.310 | 0.299 (-3.5%) | 0.400* (+29.0%) |
| R@10 | 0.423 | 0.293 (-30.7%) | 0.431 (+1.9%) | 0.453 | 0.423 (-6.7%) | 0.458 (+1.1%) | 0.513 | 0.393 (-23.4%) | 0.538 (+4.9%) | 0.539 | 0.604 (+12.1%) | 0.715* (+32.7%) | |
| R@25 | 0.492 | 0.377 (-23.4%) | 0.509 (+3.5%) | 0.510 | 0.482 (-5.5%) | 0.530 (+3.9%) | 0.574 | 0.458 (-20.2%) | 0.609* (+6.1%) | 1 | 1 | 1 | |
| Model | Evaluation Approach | ReDial | OpenDialKG |
| KBRD | Original | 0.638 | 0.824 |
| iEvaLM | 0.766 (+20.1%) | 0.862 (+4.6%) | |
| BARCOR | Original | 0.667 | 1.149 |
| iEvaLM | 0.795 (+19.2%) | 1.211 | |
| UniCRS | Original | 0.685 | (+5.4%) 1.128 |
| iEvaLM | 1.015 | 1.314 | |
| ChatGPT | (+48.2%) | (+16.5%) | |
| Original iEvaLM | 0.787 1.331* | 1.221 1.513* |
| Method | Unpersuasive | Partially persuasive | Highly persuasive |
| iEvaLM | 1% | 5% | 94% |
| Human | 4% | 7% | 89% |
| Evaluation Approach | KBRD | BARCOR | UniCRS | ChatGPT | |
| iEvaLM | Recall@ 10 | 0.180 | 0.210 | 0.330 | 0.460 |
| Persuasiveness | 0.810 | 0.860 | 1.050 | 1.330 | |
| Human | Recall@ 10 | 0.210 | 0.250 | 0.370 | 0.560 |
| Persuasiveness | 0.870 | 0.930 | 1.120 | 1.370 | |
| Models | Accuracy | Precision | Recall | F1 score | AUROC | AUPRC |
| Non-DL (KNN) | 0.6766±0.0000 | 0.6500±0.0000 | 0.6821±0.0000 | 0.6442±0.0000 | 0.8190±0.0000 | 0.5231±0.0000 |
| Random Init. | 0.4219±0.0865 | 0.4751±0.0925 | 0.4963±0.1026 | 0.4886±0.0967 | 0.7129±0.1206 | 0.3358±0.1194 |
| TS-SD | 0.6937±0.0533 | 0.6806±0.0496 | 0.6883±0.0525 | 0.6785±0.0495 | 0.8708±0.0305 | 0.6261±0.0790 |
| TS2vec | 0.6453±0.0260 | 0.6287±0.0339 | 0.6451±0.0218 | 0.6261±0.0294 | 0.8890±0.0054 | 0.6670±0.0118 |
| CLOCS | 0.4731±0.0229 | 0.4639±0.0432 | 0.4766±0.0266 | 0.4392±0.0198 | 0.8161±0.0068 | 0.4916±0.0103 |
| Mixing-up | 0.7183±0.0123 | 0.7001±0.0166 | 0.7183±0.0123 | 0.6991±0.0145 | 0.9127±0.0018 | 0.7654±0.0071 |
| TS-TCC | 0.7593±0.0242 | 0.7668±0.0257 | 0.7566±0.0231 | 0.7457±0.0210 | 0.8866±0.0040 | 0.7217±0.0121 |
| SimCLR | 0.4383±0.0652 | 0.4255±0.1072 | 0.4383±0.0652 | 0.3713±0.0919 | 0.7721±0.0559 | 0.4116±0.0971 |
| TF-C (Ours) | 0.7824±0.0237 | 0.7982±0.0496 | 0.8011±0.0322 | 0.7991±0.0296 | 0.9052±0.0136 | 0.7861±0.0149 |
| Scenarios | Models | Accuracy | Precision | Recall | F1 score | AUROC | AUPRC |
| SLEEPEEG √ EPILEPSY | Non-DL (KNN) | 0.8525±0.0000 | 0.8639±0.0000 | 0.6431±0.0000 | 0.6791±0.0000 | 0.6434±0.0000 | 0.6279±0.0000 |
| Random Init. | 0.8983±0.0656 | 0.9213±0.1369 | 0.7447±0.1135 | 0.7959±0.1208 | 0.8578±0.2153 | 0.6489±0.1926 | |
| TS-SD | 0.8952±0.0522 | 0.8018±0.2244 | 0.7647±0.1485 | 0.7767±0.1855 | 0.7677±0.2452 | 0.7940±0.1825 | |
| TS2vec | 0.9395±0.0044 | 0.9059±0.0116 | 0.9039±0.0118 | 0.9045±0.0067 | 0.9587±0.0086 | 0.9430±0.0103 | |
| CLOCS | 0.9507±0.0027 | 0.9301±0.0067 | 0.9127±0.0165 | 0.9206±0.0066 | 0.9803±0.0023 | 0.9609±0.0116 | |
| Mixing-up | 0.8021±0.0000 | 0.4011±0.0000 | 0.5000±0.0000 | 0.4451±0.0000 | 0.9743±0.0081 | 0.9618±0.0104 | |
| TS-TCC | 0.9253±0.0098 | 0.9451±0.0049 | 0.8181±0.0257 | 0.8633±0.0215 | 0.9842±0.0034 | 0.9744±0.0043 | |
| SimCLR | 0.9071±0.0344 | 0.9221±0.0166 | 0.7864±0.1071 | 0.8178±0.0998 | 0.9045±0.0539 | 0.9128±0.0205 | |
| TF-C (Ours) | 0.9495±0.0249 | 0.9456±0.0108 | 0.8908±0.0216 | 0.9149±0.0534 | 0.9811±0.0237 | 0.9703±0.0199 | |
| Non-DL (KNN) | 0.4473±0.0000 | 0.2847±0.0000 | 0.3275±0.0000 | 0.2284±0.0000 | 0.4946±0.0000 | 0.3308±0.0000 | |
| SLEEPEEG √ FD-B | Random Init. | 0.4736±0.0623 | 0.4829±0.0529 | 0.5235±0.1023 | 0.4911±0.0590 | 0.7864±0.0349 | 0.7528±0.0254 |
| TS-SD | 0.5566±0.0210 | 0.5710±0.0535 | 0.6054±0.0272 | 0.5703±0.0328 | 0.7196±0.0113 | 0.5693±0.0532 | |
| TS2vec | 0.4790±0.0113 | 0.4339±0.0092 | 0.4842±0.0197 | 0.4389±0.0107 | 0.6463±0.0130 | 0.4442±0.0162 | |
| CLOCS | 0.4927±0.0310 | 0.4824±0.0316 | 0.5873±0.0387 | 0.4746±0.0485 | 0.6992±0.0099 | 0.5501±0.0365 | |
| Mixing-up | 0.6789±0.0246 | 0.7146±0.0343 | 0.7613±0.0198 | 0.7273±0.0228 | 0.8209±0.0035 | 0.7707±0.0042 | |
| TS-TCC | 0.5499±0.0220 | 0.5279±0.0293 | 0.6396±0.0178 | 0.5418±0.0338 | 0.7329±0.0203 | 0.5824±0.0468 | |
| SimCLR | 0.4917±0.0437 | 0.5446±0.1024 | 0.4760±0.0885 | 0.4224±0.1138 | 0.6619±0.0219 | 0.5009±0.0477 | |
| TF-C (Ours) | 0.6938±0.0231 | 0.7559±0.0349 | 0.7202±0.0257 | 0.7487±0.0268 | 0.8965±0.0135 | 0.7871±0.0267 | |
| SLEEPEEG √ GESTURE | Non-DL (KNN) | 0.6833±0.0000 | 0.6501±0.0000 | 0.6833±0.0000 | 0.6443±0.0000 | 0.8190±0.0000 | 0.5232±0.0000 |
| Random Init. | 0.4219±0.0629 | 0.4751±0.0175 | 0.4963±0.0679 | 0.4886±0.0459 | 0.7129±0.0166 | 0.3358±0.1439 | |
| TS-SD | 0.6922±0.0444 | 0.6698±0.0472 | 0.6867±0.0488 | 0.6656±0.0443 | 0.8725±0.0324 | 0.6185±0.0966 | |
| TS2vec | 0.6917±0.0333 | 0.6545±0.0358 | 0.6854±0.0349 | 0.6570±0.0392 | 0.8968±0.0123 | 0.6989±0.0346 | |
| CLOCS | 0.4433±0.0518 | 0.4237±0.0794 | 0.4433±0.0518 | 0.4014±0.0602 | 0.8073±0.0109 | 0.4460±0.0384 | |
| Mixing-up | 0.6933±0.0231 | 0.6719±0.0232 | 0.6933±0.0231 | 0.6497±0.0306 | 0.8915±0.0261 | 0.7279±0.0558 | |
| TS-TCC | 0.7188±0.0349 | 0.7135±0.0352 | 0.7167±0.0373 | 0.6984±0.0360 | 0.9099±0.0085 | 0.7675±0.0201 | |
| SimCLR | 0.4804±0.0594 | 0.5946±0.1623 | 0.5411±0.1946 | 0.4955±0.1870 | 0.8131±0.0521 | 0.5076±0.1588 | |
| TF-C (Ours) | 0.7642±0.0196 | 0.7731±0.0355 | 0.7429±0.0268 | 0.7572±0.0311 | 0.9238±0.0159 | 0.7961±0.0109 | |
| Non-DL (KNN) | 0.4390±0.0000 | 0.3772±0.0000 | 0.5143±0.0000 | 0.3979±0.0000 | 0.6025±0.0000 | 0.4084±0.0000 | |
| SLEEPEEG √ EMG | Random Init. | 0.7780±0.0729 | 0.5909±0.0625 | 0.6667±0.0135 | 0.6238±0.0267 | 0.9109±0.1239 | 0.7771±0.1427 |
| TS-SD | 0.4606±0.0000 | ||||||
| TS2vec | 0.7854±0.0318 | 0.1545±0.0000 0.8040±0.0750 | 0.3333±0.0000 0.6785±0.0396 | 0.2111±0.0000 0.6766±0.0501 | 0.5005±0.0126 0.9331±0.0164 | 0.3775±0.0110 | |
| CLOCS | 0.6985±0.0323 | 0.5306±0.0750 | 0.5354±0.0291 | 0.5139±0.0409 | 0.7923±0.0573 | 0.8436±0.0372 0.6484±0.0680 | |
| Mixing-up | 0.3024±0.0534 | 0.1099±0.0126 | 0.2583±0.0456 | 0.1541±0.0204 | 0.4506±0.1718 | 0.3660±0.1635 | |
| TS-TCC | 0.7889±0.0192 | 0.5851±0.0974 | 0.6310±0.0991 | 0.5904±0.0952 | 0.8851±0.0113 | 0.7939±0.0386 | |
| SimCLR | 0.6146±0.0582 | 0.5361±0.1724 | 0.4990±0.1214 | 0.4708±0.1486 | 0.7799±0.1344 | 0.6392±0.1596 | |
| TF-C (Ours) | 0.8171±0.0287 | 0.7265±0.0353 | 0.8159±0.0289 | 0.7683±0.0311 | 0.9152±0.0211 | 0.8329±0.0137 |
| Attack success rate | |||||
| Distant Attack | Nearby Attack | ||||
| ARCA | GBDA | ARCA | GBDA | ||
| GPT-2 | None | 67% ± 4% | 12% ± 3% | 84% ±3% | 16% ± 3% |
| LLaMA | None | 2% ±1% | 1% ± 1% | 61% ± 6% | 10% ±1% |
| Vicuna | Instruct | 6%± 2% | 0% ± 1% | 8%± 2% | 1% ±1% |
| Pass Rate given N× extra tokens | ||||
| Method | 1× | 2× | 5× | 10× |
| Brute Force | 100.0% | 100.0% | 100.0% | 100.0% |
| ARCA | 11.1% | 14.6% | 25.8% | 30.6% |
| GBDA | 3.1% | 6.2% | 8.8 % | 9.5% |
| Model | Attack Success Rate | Mean l2 Distortion |
| LLaMA Adapter | 100% | 3.91 ± 0.36 |
| Mini GPT-4 (Instruct) | 100% | 2.51 ± 1.45 |
| Mini GPT-4 (RLHF) | 100% | 2.71 ± 2.12 |
| LLaVA | 100% | 0.86 ± 0.17 |
| Model | Size | Elo |
| GPT-4 | = | 1348 ±1 |
| Guanaco 65B | 41 GB | 1022 ±1 |
| Guanaco 33B | 21 GB | 992 ±1 |
| Vicuna 13B | 26 GB | 974±1 |
| ChatGPT | = | 966±1 |
| Guanaco 13B | 10 GB | 916±1 |
| Bard | - | 902±1 |
| Guanaco 7B | 6 GB | 879±1 |
| LLaMA Size | Mean 5-shot MMLU Accuracy | Mean | |||||||
| 7B | 13B | 33B | 65B | ||||||
| Alpaca | FLAN v2 | Alpaca | FLAN v2 | Alpaca | FLAN v2 | Alpaca | FLAN v2 | ||
| BFloat16 | 38.4 | 45.6 | 47.2 | 50.6 | 57.7 | 60.5 | 61.8 | 62.5 | 53.0 |
| Float4 | 37.2 | 44.0 | 47.3 | 50.0 | 55.9 | 58.5 | 61.3 | 63.3 | 52.2 |
| NFloat4 + DQ | 39.0 | 44.5 | 47.5 | 50.7 | 57.3 | 59.2 | 61.8 | 63.9 | 53.1 |
| Data type | Mean PPL |
| Int4 | 34.34 |
| Float4 (E2M1) | 31.07 |
| Float4 (E3M0) | 29.48 |
| NFloat4 + DQ | 27.41 |
| Model/Dataset | Params | Model bits | Memory | ChatGPT vs Sys | Sys vs ChatGPT | Mean | 95% CI |
| GPT-4 Bard | = | = | = | 119.4% | 110.1% | 114.5% | 2.6% |
| 1 | - | - | 93.2% | 96.4% | 94.8% | 4.1% | |
| Guanaco Alpaca FLAN v2 | 65B | 4-bit | 41 GB | 96.7% | 101.9% | 99.3% | 4.4% |
| 65B | 4-bit | 41 GB | 63.0% | 77.9% | 70.7% | 4.3% | |
| 65B | 4-bit | 41 GB | 37.0% | 59.6% | 48.4% | 4.6% | |
| Guanaco Open Assistant Alpaca FLAN v2 | 33B | 4-bit | 21 GB | 96.5% | 99.2% | 97.8% | 4.4% |
| 33B | 16-bit | 66 GB | 73.4% | 85.7% | 78.1% | 5.3% | |
| 33B 33B | 4-bit 4-bit | 21GB 21GB | 67.2% | 79.7% | 73.6% | 4.2% | |
| Vicuna Guanaco Alpaca HH-RLHF | 26.3% | 49.7% | 38.0% | 3.9% | |||
| 13B 13B | 16-bit 4-bit | 26 GB 10 GB | 91.2% 87.3% | 98.7% 93.4% | 94.9% 90.4% | 4.5% 5.2% | |
| 13B | 4-bit | 10 GB | 63.8% | 76.7% | 69.4% | ||
| 13B | 4-bit | 10 GB | 55.5% | 69.1% | 62.5% | 4.2% | |
| 13B | 4-bit | 10 GB | 50.6% | 69.8% | 60.5% | 4.7% | |
| 13B | 4-bit | 10 GB | 49.2% | 69.3% | 59.5% | 4.2% 4.7% | |
| Longform Self-Instruct FLAN v2 | 13B | 4-bit | 10 GB | 44.9% | 62.0% | 53.6% | 5.2% |
| 13B | 4-bit | 10 GB | 38.0% | 60.5% | 49.1% | 4.6% | |
| 13B | 4-bit | 10 GB | 32.4% | 61.2% | 47.0% | 3.6% | |
| 4-bit | 5GB | ||||||
| Guanaco Alpaca FLAN v2 | 7B 7B | 4-bit | 5GB | 84.1% 57.3% | 89.8% 71.2% | 87.0% | 5.4% |
| 7B | 4-bit | 5GB | 33.3% | 56.1% | 64.4% 44.8% | 5.0% 4.0% |
| Dataset | 7B | 13B | 33B | 65B |
| LLaMA no tuning | 35.1 | 46.9 | 57.8 | 63.4 |
| Self-Instruct | 36.4 | 33.3 | 53.0 | 56.7 |
| Longform | 32.1 | 43.2 | 56.6 | 59.7 |
| Chip2 | 34.5 | 41.6 | 53.6 | 59.8 |
| HH-RLHF | 34.9 | 44.6 | 55.8 | 60.1 |
| Unnatural Instruct | 41.9 | 48.1 | 57.3 | 61.3 |
| Guanaco (OASST1) | 36.6 | 46.4 | 57.0 | 62.2 |
| Alpaca | 38.8 | 47.8 | 57.3 | 62.5 |
| FLAN v2 | 44.5 | 51.4 | 59.2 | 63.9 |
| Benchmark # Prompts Judge Model | Vicuna 80 Human raters | Vicuna 80 GPT-4 | Open Assistant 953 GPT-4 | Median Rank | |||
| Elo | Rank | Elo | Rank | Elo | Rank | ||
| GPT-4 | 1176 | 1 | 1348 | 1 | 1294 | 1 | 1 |
| Guanaco-65B | 1023 | 2 | 1022 | 1008 | 3 | 2 | |
| Guanaco-33B | 1009 | 4 | 992 | 1002 | 4 | 4 | |
| ChatGPT-3.5 Turbo | 916 | 7 | 966 | 235 | 1015 | 2 | 5 |
| Vicuna-13B | 984 | 5 | 974 | 4 | 936 | 5 | 5 |
| Guanaco-13B | 975 | 6 | 913 | 6 | 885 | 6 | 6 |
| Guanaco-7B | 1010 | 3 | 879 | 8 | 860 | 7 | 7 |
| Bard | 909 | 8 | 902 | 7 | 1 | - | 8 |
| Dataset Model | GLUE (Acc.) | Super-NaturalInstructions (RougeL) | |||||
| RoBERTa-large | T5-80M | T5-250M | T5-780M | T5-3B | T5-11B | ||
| BF16 | 88.6 | 40.1 | 42.1 | 48.0 | 54.3 | 62.0 | |
| BF16 replication | 88.6 | 40.0 | 42.2 | 47.3 | 54.9 | 1 | |
| LoRA BF16 | 88.8 | 40.5 | 42.6 | 47.1 | 55.4 | 60.7 | |
| QLoRA Int8 | 88.8 | 40.4 | 42.9 | 45.4 | 56.5 | 60.7 | |
| QLORA FP4 | 88.6 | 40.3 | 42.4 | 47.5 | 55.6 | 60.9 | |
| QLORA NF4 + DQ | 1 | 40.4 | 42.7 | 47.7 | 55.3 | 60.9 | |
| Parameters | Dataset | Batch size | LR | Steps | Source Length | Target Length |
| 7B | All | 16 | 2e-4 | 10000 | 384 | 128 |
| 7B | OASST1 | 16 | 2e-4 | 1875 | - | 512 |
| 7B | HH-RLHF | 16 | 2e-4 | 10000 | 1 | 768 |
| 7B | Longform | 16 | 2e-4 | 4000 | 512 | 1024 |
| 13B | All | 16 | 2e-4 | 10000 | 384 | 128 |
| 13B | OASST1 | 16 | 2e-4 | 1875 | - | 512 |
| 13B | HH-RLHF | 16 | 2e-4 | 10000 | = | 768 |
| 13B | Longform | 16 | 2e-4 | 4000 | 512 | 1024 |
| 33B | All | 32 | 1e-4 | 5000 | 384 | 128 |
| 33B | OASST1 | 16 | 1e-4 | 1875 | - | 512 |
| 33B | HH-RLHF | 32 | 1e-4 | 5000 | 1 | 768 |
| 33B | Longform | 32 | 1e-4 | 2343 | 512 | 1024 |
| 65B | All | 64 | 1e-4 | 2500 | 384 | 128 |
| 65B | OASST1 | 16 | 1e-4 | 1875 | - | 512 |
| 65B | HH-RLHF | 64 | 1e-4 | 2500 | 1 | 768 |
| 65B | Longform | 32 | 1e-4 | 2343 | 512 | 1024 |
| Dataset | Unnatural Instructions | Chip2 | Alpaca | FLAN v2 | Mean |
| Train on source and target | 36.2 | 33.7 | 38.1 | 42.0 | 37.5 |
| Train on target | 38.0 | 34.5 | 39.0 | 42.9 | 38.6 |
| Datapoints ↓Epochs→ | Chip | Unnatural Instructions | FLAN v2 | |||||||
| 1 | 2 | 3 | 1 | 2 | 3 | 1 | 2 | 3 | Mean | |
| 50000 | 34.50 | 35.30 | 34.70 | 38.10 | 42.20 | 38.10 | 43.00 | 43.50 | 44.10 | 39.28 |
| 100000 | 33.70 | 33.90 | 34.00 | 40.10 | 41.20 | 37.00 | 43.90 | 43.70 | 44.90 | 39.16 |
| 150000 | 34.40 | 34.80 | 35.10 | 39.70 | 41.10 | 41.50 | 44.60 | 45.50 | 43.50 | 40.02 |
| Mean | 34.20 | 34.67 | 34.60 | 39.30 | 41.50 | 38.87 | 43.83 | 44.23 | 44.17 | |
| total#numberof judgments | ChatGPT-3.5 Turbo Bard | Guanaco 13B | Guanaco 7B | ||||
| Model | Guanaco 65B | Guanaco 33B | Vicuna | ||||
| Guanaco 65B | - | 0.21 | 0.19 | 0.16 | 0.72 | 0.59 | 0.86 |
| Guanaco 33B | -0.21 | 0.17 | 0.10 | 0.51 | 0.41 | 0.68 | |
| Vicuna | -0.19 | -0.17 | 0.10 | 0.50 | 0.20 | 0.57 | |
| ChatGPT-3.5 Turbo | -0.16 | -0.10 | -0.10 | - | 0.35 | 0.19 | 0.40 |
| Bard | -0.72 | -0.51 | -0.50 -0.20 | -0.35 -0.19 | - -0.12 | 0.12 | 0.03 |
| Guanaco 13B Guanaco 7B | -0.59 | -0.41 | -0.57 | -0.40 | -0.03 | -0.20 | 0.20 |
| -0.86 | -0.68 | - | |||||
| Model | Params | Size |
| Guanaco | 65B | 41 GB |
| Guanaco | 33B | 21 GB |
| Vicuna | 13B | 26 GB |
| ChatGPT-3.5 Turbo | N/A | N/A |
| Bard | N/A | N/A |
| Guanaco | 13B | 10 GB |
| Guanaco | 7B | 5GB |
| LLaMA-65B | GPT-3 | OPT-175B | Guanaco-65B | |
| Gender | 70.6 | 62.6 | 65.7 | 47.5 |
| Religion | 79.0 | 73.3 | 68.6 | 38.7 |
| Race/Color | 57.0 | 64.7 | 68.6 | 45.3 |
| Sexual orientation | 81.0 | 76.2 | 78.6 | 59.1 |
| Age | 70.1 | 64.4 | 67.8 | 36.3 |
| Nationality | 64.2 | 61.6 | 62.9 | 32.4 |
| Disability | 66.7 | 76.7 | 76.7 | 33.9 |
| Physical appearance | 77.8 | 74.6 | 76.2 | 43.1 |
| Socioeconomic status | 71.5 | 73.8 | 76.2 | 55.3 |
| Average | 66.6 | 67.2 | 69.5 | 43.5 |
| Model Sampler | Product | Mixture | |||||
| RAISE↑ | LL↑ | Var↓ | ln(MMD)↓ | LL↑ | Var↓ | ||
| Reverse | 1.55 | -6.47 | 0.063 | = | |||
| Score | ULA | 2.37 | 1.79 | 0.026 | = | ||
| U-HMC | 2.52 | 2.40 | 0.021 | ||||
| Reverse (equal steps) | 2.27 | -2.92 | 0.046 | = | = | - | |
| Reverse | 1.37 | -6.03 | 0.064 | -3.84 | -2.17 | 0.020 | |
| ULA | 2.36 | 1.84 | 0.027 | -4.21 | 0.57 | 0.013 | |
| EBM | MALA | 2.64 | 2.73 | 0.013 | -4.38 | 1.29 | 0.008 |
| U-HMC | 2.63 | 2.45 | 0.022 | -4.69 | 1.03 | 0.010 | |
| HMC | 2.71 | 2.72 | 0.009 | -4.48 | 1.30 | 0.007 | |
| Model | Sampler | Inception Score ↑ | FID↓ | Accuracy ↑ |
| Score | Reverse | 29.10 | 30.46 | 18.64 |
| LA | 29.35 | 30.49 | 65.81 | |
| U-HMC | 32.19 | 26.89 | 89.93 | |
| EBM | Reverse | 28.05 | 33.58 | 18.60 |
| LA | 28.12 | 33.45 | 66.28 | |
| MALA | 30.43 | 32.22 | 83.65 | |
| U-HMC | 31.39 | 32.08 | 90.83 | |
| HMC | 33.46 | 30.52 | 94.61 |
| Model | Sampler | Combinations | ||||
| 1 | 2 | 3 | 4 | 5 | ||
| Reverse | 70.8 | 68.2 | 66.3 | 64.1 | 57.4 | |
| Score | ULA | 75.0 | 73.4 | 71.8 | 67.9 | 60.2 |
| U-HMC | 79.1 | 76.0 | 73.6 | 71.1 | 62.3 | |
| Reverse | 71.0 | 67.1 | 62.5 | 58.1 | 51.0 | |
| ULA | 81.3 | 71.8 | 66.6 | 59.6 | 54.8 | |
| EBM | MALA | 85.4 | 74.4 | 71.1 | 65.6 | 63.9 |
| U-HMC | 84.5 | 81.3 | 79.2 | 74.2 | 68.1 | |
| HMC | 91.6 | 82.9 | 80.1 | 76.5 | 72.7 | |
| Input: Initial state x°,Mass matrix M, Number of steps N,Number leapfrog steps L,step-size ∈ | |
| fori=1,...,N do | |
| Sample vi ~ N(0, M) | # Sample momentum |
| x',uu =Leapfrog(xi-1, ,U²;,L) | #Integrate dynamics with stepsize e for L steps |
| α = min p(x,') 1, p(xi-1,0) | # Compute acceptance probability |
| With probability a | |
| set xi= x' | |
| else | |
| set xi=xi-1 | |
| damping-factor y | Input: Initial state xo,Mass matrix M,Number of steps N,Number leapfrog steps L,step-size |
| Sample v° ~ N(0, M) | #Sample initial momentum |
| for i= 1,...,N do | |
| λ~ N(0,M) | |
| u(i-1)'=γv-1+√1-γ²x | #Partially refresh momentum |
| x',v'=Leapfrog(x𝑖-1,u(i-1)';∈,L) | # Integrate dynamics with stepsize εfor L steps |
| u'=-v' # Negate momentum | |
| a = min (1, p(x',u') | |
| p(xi-1,u(i-1)1)) | # Compute acceptance probability |
| With probability a | |
| set xi= x',v = U' | |
| else | |
| set xi= xi-1,v²= v(i-1)1 | |
| U²=-ui # Negate momentum | |
| end for | |
| xN | |
| return |
| Parameterization ∈θ(x,t) | ||
| (1) DAE 97.4 | (2)L2 Norm 91.5 | (3) Inner-Product | 90.9 86.7 |
| Weighting Strategy | Segmentation | Depth | Surface Normal | △p↑ | △t | ||||||
| mIoU↑ | Pix Acc↑ | Abs Err↓ | Rel Err↓ | Angle Distance | Within t | ||||||
| Mean↓ | Median↓ | 11.25个 22.5个 | 30个 | ||||||||
| EW | 53.91 | 75.56 | 0.3840 | 0.1567 | 23.6338 | 17.2451 | 34.94 | 60.65 | 71.81 | +0.00 | ×1.00 |
| GradNorm | 53.81 | 75.35 | 0.3863 | 0.1556 | 23.6106 | 17.2565 | 34.98 | 60.58 | 71.76 | -0.06 | ×1.82 |
| UW | 53.15 | 75.41 | 0.3817 | 0.1576 | 23.6487 | 17.2040 | 34.98 | 60.71 | 71.80 | -0.24 | ×1.01 |
| MGDA | 53.66 | 75.37 | 0.3864 | 0.1610 | 23.4757 | 16.9912 | 35.44 | 61.17 | 72.16 | -0.35 | ×2.64 |
| DWA | 53.33 | 75.42 | 0.3834 | 0.1556 | 23.5806 | 17.1242 | 35.18 | 60.88 | 71.91 | +0.07 | ×1.00 |
| PCGrad | 53.34 | 75.43 | 0.3857 | 0.1600 | 23.2293 | 16.6966 | 36.09 | 61.80 | 72.66 | +0.12 | ×2.10 |
| GradDrop | 53.80 | 75.56 | 0.3857 | 0.1587 | 23.8726 | 17.1406 | 35.10 | 60.72 | 71.60 | -0.33 | ×1.84 |
| IMTL | 52.90 | 74.88 | 0.3883 | 0.1632 | 23.0534 | 16.5304 | 36.30 | 62.20 | 73.08 | -0.35 | ×1.82 |
| GradVac | 53.52 | 75.43 | 0.3840 | 0.1559 | 23.2892 | 16.8601 | 35.67 | 61.53 | 72.46 | +0.48 | ×2.11 |
| RLW (Uniform) | 54.09 | 75.78 | 0.3826 | 0.1563 | 23.6272 | 17.2711 | 34.73 | 60.67 | 71.87 | +0.17 | |
| RLW (Normal) | 54.19 | 75.98 | 0.3789 | 0.1570 | 23.1984 | 16.7944 | 35.71 | 61.74 | 72.77 | +1.02 | |
| RLW (Dirichlet) | 53.54 | 75.45 | 0.3834 | 0.1547 | 23.6392 | 17.0715 | 35.28 | 60.92 | 71.88 | +0.27 | ×1.00 |
| RLW (Bernoulli) | 53.72 | 75.62 | 0.3850 | 0.1610 | 23.1413 | 16.6591 | 36.08 | 61.98 | 72.86 | +0.28 | |
| RLW (constrained Bernoulli) | 54.32 | 75.78 | 0.3779 | 0.1533 | 23.2101 | 16.9354 | 35.41 | 61.44 | 72.58 | +1.29 | |
| RLW (random Normal) | 54.08 | 75.77 | 0.3815 | 0.1581 | 23.5598 | 16.9577 | 35.53 | 61.20 | 72.13 | +0.39 | |
| Weighting Strategy | Ss | HPS | Saliency | Surface Normal | △p↑ | ||||||
| mIoU↑ | mIoU↑ | mIoU↑ maxF个 | Angle Distance | Within t | |||||||
| Mean↓ | RMSE↓ | 11.25个 | 22.5个 | 30个 | |||||||
| EW | 64.52 | 58.71 | 64.31 | 77.11 | 17.6444 | 26.1634 | 42.24 | 75.93 | 86.97 | +0.00 | |
| GradNorm | 64.13 | 58.49 | 61.64 | 72.46 | 18.0455 | 26.4642 | 41.03 | 74.90 | 86.26 | -1.94 | |
| UW | 63.72 | 59.13 | 64.47 | 77.26 | 17.4962 | 26.0463 | 42.64 | 76.48 | 87.35 | +0.09 | |
| MGDA | 63.34 | 58.86 | 64.79 | 77.54 | 17.3070 | 25.8584 | 43.30 | 77.14 | 87.79 | +0.18 | |
| DWA | 64.32 | 58.61 | 64.30 | 77.15 | 17.4065 | 25.9242 | 42.72 | 76.71 | 87.59 | +0.14 | |
| PCGrad | 63.58 | 58.68 | 63.79 | 76.71 | 17.2376 | 25.8572 | 43.69 | 77.11 | 87.70 | -0.08 | |
| GradDrop | 64.04 | 59.36 | 62.31 | 73.10 | 17.3246 | 25.8966 | 43.29 | 76.92 | 87.63 | -0.58 | |
| IMTL | 62.67 | 58.35 | 62.92 | 73.21 | 16.8026 | 25.4852 | 45.02 | 78.49 | 88.65 | -0.81 | |
| GradVac | 62.99 | 58.63 | 64.30 | 77.15 | 17.1852 | 25.7621 | 43.55 | 77.41 | 88.00 | -0.10 | |
| RLW (Uniform) | 63.52 | 59.03 | 63.75 | 76.71 | 17.0261 | 25.6528 | 44.21 | 77.77 | 88.24 | +0.27 | |
| RLW (Normal) | 64.14 | 58.43 | 64.05 | 76.86 | 17.1794 | 25.7734 | 43.69 | 77.31 | 87.91 | +0.16 | |
| RLW (Dirichlet) | 63.97 | 58.88 | 64.30 | 77.19 | 17.2147 | 25.8093 | 43.67 | 77.33 | 87.89 | +0.37 | |
| RLW (Bernoulli) | 64.34 | 58.35 58.52 | 64.28 64.19 | 77.03 | 17.3379 17.3377 | 25.9016 25.9005 | 43.18 43.30 | 76.93 77.04 | 87.66 87.69 | +0.11 | |
| RLW (constrained Bernoulli) | 65.07 64.09 | 59.15 | 63.84 | 76.96 76.76 | 17.3860 | 25.9472 | 42.98 | 76.73 | 87.54 | +0.46 | |
| RLW (random Normal) | +0.16 | ||||||||||
| Weighting Strategy | POS (F1 Score) | PI (Accuracy) | ||||||||
| en | zh | te | vi | Avg | en | zh | de | es | Avg | |
| EW | 95.02 | 88.89 | 91.16 | 87.11 | 90.55 | 94.09 | 84.59 | 89.44 | 90.24 | 89.59 |
| GradNorm | 95.01 | 88.91 | 91.88 | 87.06 | 90.71 | 94.39 | 85.94 | 90.99 | 91.44 | 90.69 |
| UW | 94.89 | 88.77 | 90.96 | 87.12 | 90.44 | 93.74 | 85.44 | 90.24 | 91.29 | 90.18 |
| MGDA | 95.08 | 88.97 | 92.35 | 87.12 | 90.88 | 94.64 | 84.99 | 89.84 | 90.89 | 90.09 |
| DWA | 95.02 | 89.03 | 91.87 | 87.27 | 90.80 | 94.69 | 84.99 | 89.49 | 91.44 | 90.15 |
| PCGrad | 94.85 | 88.42 | 90.72 | 86.71 | 90.18 | 94.19 | 85.49 | 89.09 | 91.24 | 90.00 |
| GradDrop | 95.08 | 89.06 | 90.65 | 87.17 | 90.49 | 94.29 | 84.44 | 89.69 | 90.94 | 89.84 |
| IMTL | 94.87 | 88.80 | 92.18 | 86.72 | 90.65 | 94.54 | 84.79 | 90.14 | 90.99 | 90.12 |
| GradVac | 94.87 | 88.41 | 90.62 | 86.47 | 90.09 | 94.29 | 84.94 | 89.19 | 90.89 | 89.83 |
| RLW (Uniform) | 95.06 | 89.00 | 92.31 | 86.93 | 90.83 | 94.69 | 85.79 | 90.29 | 91.94 | 90.68 |
| RLW (Normal) | 95.01 | 88.87 | 92.86 | 86.85 | 90.90 | 94.39 | 85.34 | 90.04 | 91.84 | 90.40 |
| RLW (Dirichlet) | 95.16 | 88.96 | 91.64 | 87.24 | 90.75 | 94.24 | 84.39 | 89.99 | 90.99 | 89.90 |
| RLW (Bernoulli) | 95.13 | 89.10 | 91.13 | 87.03 | 90.60 | 95.09 | 85.89 | 90.24 | 91.99 | 90.80 |
| RLW (constrained Bernoulli) | 94.98 | 89.05 | 92.33 | 86.87 | 90.81 | 94.69 | 85.49 | 90.19 | 90.99 | 90.34 |
| RLW (random Normal) | 95.07 | 89.00 | 91.10 | 87.25 | 90.60 | 94.79 | 84.94 | 89.54 | 90.99 | 90.07 |
| Weighting Strategy | Segmentation | Depth | Surface Normal | △p↑ | ||||||
| mIoU个 | Pix Acc↑ | Abs Err↓ | Rel Err↓ | Angle Distance | Within t° | |||||
| Mean↓ | Median↓ | 11.25个 | 22.5个 | 30个 | ||||||
| EW | 53.77 | 75.79 | 0.3789 | 0.1546 | 22.8344 | 16.6021 | 36.37 | 62.45 | 73.32 | +0.00 |
| GradNorm | 54.75 | 75.89 | 0.3797 | 0.1537 | 22.6295 | 16.2829 | 37.09 | 63.12 | 73.87 | +0.83 |
| UW | 54.80 | 75.97 | 0.3767 | 0.1536 | 22.6744 | 16.3050 | 36.90 | 63.14 | 73.87 | +0.95 |
| MGDA | 53.94 | 75.97 | 0.3788 | 0.1573 | 22.6157 | 16.1395 | 37.24 | 63.38 | 73.94 | +0.37 |
| DWA | 53.71 | 75.72 | 0.3801 | 0.1539 | 23.1560 | 16.8354 | 35.95 | 61.85 | 72.79 | -0.39 |
| PCGrad | 53.83 | 75.70 | 0.3823 | 0.1568 | 22.9481 | 16.3528 | 37.05 | 62.76 | 73.37 | -0.16 |
| GradDrop | 53.83 | 75.85 | 0.3754 | 0.1530 | 22.7846 | 16.4156 | 36.77 | 62.74 | 73.51 | +0.57 |
| IMTL | 53.39 | 75.20 | 0.3807 | 0.1549 | 22.2571 | 15.8336 | 38.04 | 64.11 | 74.59 | +0.72 |
| GradVac | 54.52 | 75.68 | 0.3755 | 0.1546 | 22.9389 | 16.5692 | 36.48 | 62.33 | 73.20 | +0.34 |
| RLW (Uniform) | 54.27 | 75.51 | 0.3820 | 0.1548 | 22.9640 | 16.4375 | 36.89 | 62.65 | 73.29 | +0.08 |
| RLW (Normal) | 53.70 | 75.62 | 0.3791 | 0.1551 | 22.8395 | 16.3328 | 37.05 | 62.82 | 73.44 | +0.16 |
| RLW (Dirichlet) | 53.36 | 75.08 | 0.3778 | 0.1514 | 22.8803 | 16.3579 | 36.99 | 62.80 | 73.40 | +0.35 |
| RLW (Bernoulli) | 53.46 | 75.74 | 0.3820 | 0.1517 | 22.5642 | 16.2013 | 37.21 | 63.29 | 73.96 | +0.61 |
| RLW (constrained Bernoulli) | 54.11 | 75.47 | 0.3821 | 0.1558 | 22.7969 | 16.2204 | 36.93 | 62.90 | 73.56 | +0.10 |
| RLW (random Normal) | 54.10 | 75.77 | 0.3802 | 0.1554 | 22.4400 | 16.0336 | 37.74 | 63.68 | 74.17 | +0.76 |
| Approach | Strategy | Weight X(k-th iteration) | Conv.* | Not Grad.t |
| EW | (,.,) | √ | √ | |
| Learning | GradNorm | minx∑t=1|tllgtll-gra|,whereg=(∑1Ml)/,rt=lt/(∑-1)lt =lt/eΩ,andα is pre-defined | X | X |
| UW | minx (λTe+ 1 log(λT)/2), s.t.λ ∈ RT | X | √ | |
| IMTL-L | minx(λTe-1log(λ)T),s.t.λ∈RT | X | X | |
| Solving | MGDA | arg minx{λG||² |λ∈△T} | √ | X |
| Calculating | DWA | T × softmax(ek-1/ek-2) | X | √ |
| PCGrad | 1(I +CN),where N= diag(1/llgill,.,1/lgrl),C =[CpqlT×T,Cpp = 0,Cpq=[-(gp+∑p<aCppup))u] | √ | X | |
| GradDrop | 1Ts+ (I-diag(s))M,where s ∈ [0,1]T is pre-defined,M = diag(I(d >e))I(G> 0)+diag(I(d <e))I(G <0),d = (1/2)(1+ 1G/(∑t=1 ltlI),e ~ Uniform(0,1),G = sgn(0) G | × | X | |
| IMTL-G | [1 -1α2:τ, a2:r],where α2:T = giNT(DNT)-1,N = 1Tu1 -U2:τ,D= 1g1-G2:T | X | × | |
| GradVac | 1(I + CN),where N = diag(1/llgil,.,1/llgrll),C =[CpqlTxT,Cpp = 0,Cpq = sgn(Φpq-Φpq)*[|gp|1* A(Φpq,Φpq)],A(Φpq,pq)=pa√1-(p)-pV1-(Φp)2√1-(Φpq)²Φpq=(gp:gq)/(llgplllgqll),gp=gp+∑p'<qCpp'Up',pq=(1-β)Φpq+βΦpq,β is pre-defined,and pq is initialized as 0 | √ | × | |
| Sampling | RLW (ours) | f(λ),where X~ p(X) and f is a normalization function | V | 了 |
| Weighting Strategy | Segmentation | Depth | Surface Normal | △p↑ | ||||||
| mIoU个 | Pix Acc↑ | Abs Err↓ | Rel Err↓ | Angle Distance | Withint | |||||
| Mean↓ | Median↓ | 11.25个 | 22.5个 | 30个 | ||||||
| EW | 53.39 | 75.06 | 0.3822 | 0.1595 | 23.1939 | 16.6749 | 35.99 | 61.79 | 72.69 | +0.00 |
| GradNorm | 53.92 | 75.39 | 0.3833 | 0.1547 | 22.7461 | 16.2463 | 37.15 | 63.02 | 73.57 | +1.42 |
| UW | 53.88 | 75.76 | 0.3820 | 0.1587 | 23.1397 | 16.4221 | 36.61 | 62.30 | 72.91 | +0.71 |
| MGDA | 53.33 | 75.07 | 0.3877 | 0.1584 | 23.1588 | 16.6327 | 36.41 | 61.75 | 72.57 | -0.05 |
| DWA | 52.62 | 75.02 | 0.3801 | 0.1563 | 23.4577 | 16.7667 | 36.13 | 61.51 | 72.24 | +0.02 |
| PCGrad | 53.83 | 75.58 | 0.3853 | 0.1568 | 22.5621 | 16.0502 | 37.80 | 63.43 | 73.89 | +1.45 |
| GradDrop | 52.23 | 74.96 | 0.3855 | 0.1583 | 22.7940 | 15.9848 | 37.79 | 63.25 | 73.53 | +0.56 |
| IMTL | 54.16 | 75.83 | 0.3833 | 0.1571 | 22.3801 | 15.9137 | 37.81 | 63.97 | 74.46 | +1.89 |
| GradVac | 52.93 | 75.22 | 0.3871 | 0.1559 | 22.4499 | 15.8148 | 38.31 | 63.96 | 74.21 | +1.42 |
| RLW (Uniform) | 53.23 | 75.19 | 0.3870 | 0.1576 | 22.8859 | 16.3011 | 37.06 | 62.55 | 73.17 | +0.53 |
| RLW (Normal) | 53.34 | 75.32 | 0.3804 | 0.1577 | 22.9278 | 16.2227 | 37.25 | 62.68 | 73.27 | +0.95 |
| RLW (Dirichlet) | 52.86 | 75.30 | 0.3859 | 0.1604 | 22.9058 | 16.0335 | 37.66 | 62.98 | 73.31 | +0.47 |
| RLW (Bernoulli) | 53.44 | 75.33 | 0.3782 | 0.1558 | 22.8759 | 16.1796 15.9059 | 37.27 | 62.79 | 73.29 | +1.33 |
| RLW (constrained Bernouli) | 53.25 | 75.28 | 0.3819 | 0.1530 | 22.5622 | 15.8673 | 37.80 37.97 | 63.39 63.44 | 73.78 73.73 | +1.79 |
| RLW (random Normal) | 53.22 | 75.21 | 0.3861 | 0.1607 | 22.6981 | +0.79 | ||||
| Weighting Strategy | Segmentation | Depth | Surface Normal | △p↑ | ||||||
| mIoU个 | Pix Acc↑ | Abs Err↓ | Rel Err↓ | Angle Distance | Within t° | |||||
| Mean↓ | Median↓ | 11.25个 22.5个 | 30个 | |||||||
| EW | 53.93 | 75.56 | 0.3835 | 0.1566 | 22.7908 | 16.1512 | 37.16 | 63.14 | 73.65 | +0.00 |
| GradNorm | 54.47 | 75.37 | 0.3882 | 0.1551 | 22.4295 | 15.9649 | 37.89 | 63.76 | 74.22 | +0.51 |
| UW | 54.16 | 75.60 | 0.3795 | 0.1541 | 22.6800 | 16.2123 | 37.19 | 63.05 | 73.66 | +0.52 |
| MGDA | 53.65 | 75.19 | 0.3841 | 0.1579 | 22.7391 | 16.0227 | 37.48 | 63.29 | 73.70 | -0.19 |
| DWA | 54.24 | 75.50 | 0.3835 | 0.1600 | 22.6793 | 16.0649 | 37.40 | 63.33 | 73.83 | -0.13 |
| PCGrad | 53.83 | 75.08 | 0.3867 | 0.1575 | 22.5083 | 15.9023 | 38.03 | 63.73 | 74.10 | +0.07 |
| GradDrop | 53.82 | 75.40 | 0.3841 | 0.1536 | 22.8330 | 16.4121 | 36.85 | 62.78 | 73.36 | -0.02 |
| IMTL | 54.49 | 75.68 | 0.3808 | 0.1531 | 21.9939 | 15.3531 | 39.42 | 65.10 | 75.18 | +2.00 |
| GradVac | 53.99 | 75.62 | 0.3874 | 0.1546 | 22.5202 | 15.9511 | 37.96 | 63.68 | 74.05 | +0.47 |
| RLW (Uniform) | 53.90 | 75.47 | 0.3859 | 0.1579 | 22.5236 | 15.8438 | 37.94 | 63.77 | 74.11 | +0.18 |
| RLW (Normal) | 53.93 | 75.31 | 0.3828 | 0.1532 | 22.4460 | 15.7951 | 38.04 | 63.93 | 74.24 | +0.88 |
| RLW (Dirichlet) | 53.17 | 74.88 | 0.3814 | 0.1561 | 22.5713 | 15.8921 | 37.88 | 63.64 | 73.99 | +0.14 |
| RLW (Bernoulli) | 54.04 | 75.13 | 0.3818 | 0.1563 | 22.4328 | 15.8549 | 37.99 | 63.76 | 74.10 | +0.53 |
| RLW (constrained Bernoulli) | 53.77 | 75.24 | 0.3816 | 0.1531 | 22.7075 | 16.0182 | 37.50 | 63.24 | 73.75 | +0.49 |
| RLW (random Normal) | 53.90 | 75.60 | 0.3847 | 0.1592 | 22.5007 | 16.0155 | 37.82 | 63.62 | 74.03 | +0.02 |
| Weighting Strategy | Segmentation | Depth | |||
| mIoU↑ | Pix Acc↑ | Abs Err↓ | Rel Err↓ | ||
| EW | 68.72 | 91.50 | 0.0134 | 46.5974 | +0.00 |
| GradNorm | 68.70 | 91.41 | 0.0133 | 45.6962 | +0.64 |
| UW | 68.72 | 91.48 | 0.0134 | 45.0403 | +0.83 |
| MGDA | 68.36 | 91.17 | 0.0123 | 45.1606 | +2.60 |
| DWA | 68.62 | 91.45 | 0.0132 | 47.1341 | +0.04 |
| PCGrad | 70.16 | 91.93 | 0.0128 | 44.6817 | +2.79 |
| GradDrop | 68.53 | 91.46 | 0.0134 | 47.1396 | -0.37 |
| IMTL | 68.74 | 91.50 | 0.0128 | 45.8279 | +1.54 |
| GradVac | 69.89 | 91.94 | 0.0126 | 47.2281 | +1.70 |
| RLW (Uniform) | 68.74 | 91.54 | 0.0132 | 45.2453 | +1.12 |
| RLW (Normal) | 68.82 | 91.53 | 0.0132 | 46.1333 | +0.67 |
| RLW (Dirichlet) | 69.37 | 91.75 | 0.0132 | 44.2218 | +1.95 |
| RLW (Bernoulli) | 69.59 | 91.81 | 0.0130 | 47.0400 | +0.91 |
| RLW (constrained Bernoulli) | 69.49 | 91.81 | 0.0128 | 46.9825 | +1.28 |
| RLW (random Normal) | 68.59 | 91.51 | 0.0132 | 46.7122 | +0.27 |
| Weighting Strategy | Avg Acc |
| EW GradNorm UW MGDA DWA | 90.70 90.77 90.84 90.73 90.77 |
| GradDrop IMTL GradVac RLW (Uniform) | 90.85 90.71 90.91 90.75 |
| RLW (Normal) RLW (Dirichlet) RLW (Bernoulli) RLW (constrained Bernoulli) RLW (random Normal) | 90.86 90.73 90.89 90.71 90.81 90.88 |
| Weighting Strategy | Office-31 | Office-Home | |||||||
| A | D | W | Avg | Ar | Cl | Pr | Rw | Avg | |
| EW | 87.17 | 98.36 | 99.44 | 94.99 | 68.88 | 80.93 | 91.73 | 81.72 | 80.81 |
| GradNorm | 86.66 | 99.18 | 98.89 | 94.91 | 68.88 | 80.49 | 91.73 | 82.70 | 80.95 |
| UW | 87.35 | 99.18 | 98.89 | 95.13 | 69.63 | 82.55 | 91.20 | 81.08 | 81.12 |
| MGDA | 87.52 | 99.18 | 99.44 | 95.38 | 69.44 | 79.30 | 91.63 | 81.72 | 80.52 |
| DWA | 87.52 | 99.18 | 99.44 | 95.38 | 70.39 | 79.95 | 90.36 | 82.05 | 80.69 |
| PCGrad | 87.00 | 98.36 | 98.33 | 94.56 | 68.31 | 80.71 | 90.57 | 81.94 | 80.38 |
| GradDrop | 87.17 | 98.36 | 98.89 | 94.80 | 68.50 | 81.47 | 91.41 | 81.40 | 80.69 |
| IMTL | 87.52 | 98.36 | 98.89 | 94.92 | 67.93 | 80.49 | 91.94 | 82.05 | 80.60 |
| GradVac | 88.71 | 98.36 | 98.89 | 95.32 | 67.93 | 82.01 | 91.84 | 81.72 | 80.87 |
| RLW (Uniform) | 88.71 | 98.36 | 98.89 | 95.32 | 70.01 | 81.79 | 90.88 | 80.97 | 80.92 |
| RLW (Normal) | 88.03 | 99.18 | 98.89 | 95.36 | 69.44 | 80.93 | 90.36 | 82.70 | 80.86 |
| RLW (Dirichlet) | 88.37 | 99.18 | 98.89 | 95.48 | 70.01 | 81.69 | 91.31 | 81.18 | 81.05 |
| RLW (Bernoulli) | 88.20 | 99.18 | 98.89 | 95.42 | 68.69 | 82.23 | 91.73 | 81.72 | 81.09 |
| RLW (constrained Bernoulli) | 88.37 | 100 | 99.44 | 95.94 | 70.77 | 81.69 | 91.41 | 82.05 | 81.48 |
| RLW (random Normal) | 88.37 | 98.36 | 98.89 | 95.20 | 68.88 | 82.12 | 91.41 | 81.72 | 81.03 |
| NER | POS | NIL | PI | |
| en | 20.0K+10.0K+10.0K | 6.9K+1.8K+3.2K | 392.7K+2.5K+5.0K | 49.4K+2.0K+2.0K |
| zh | 20.5K+10.3K+10.3K | 4.0K+0.5K+2.9K | 392.7K+2.5K+5.0K | 49.4K+2.0K+2.0K |
| de | 20.0K+10.0K+10.0K | 1 | 392.7K+2.5K+5.0K | 49.4K+2.0K+2.0K |
| es | 20.0K+10.0K+10.0K | 1 | 392.7K+2.5K+5.0K | 49.4K+2.0K+2.0K |
| te | = | 1.0K+0.1K+0.1K | = | |
| vi | = | 1.4K+0.8K+0.8K | = | 1 |
| Weighting Strategy | NLI (Accuracy) | NER (F1 Score) | ||||||||
| en | zh | de | es | Avg | en | zh | de | es | Avg | |
| EW | 82.05 | 77.04 | 77.80 | 78.48 | 78.84 | 85.21 | 83.61 | 90.73 | 92.89 | 88.11 |
| GradNorm | 83.33 | 77.34 | 78.20 | 79.54 | 79.60 | 85.70 | 84.16 | 91.13 | 93.17 | 88.54 |
| UW | 82.93 | 77.58 | 77.96 | 79.72 | 79.55 | 85.59 | 84.42 | 91.14 | 93.24 | 88.60 |
| MGDA | 83.11 | 77.32 | 77.84 | 79.02 | 79.32 | 85.68 | 84.13 | 90.99 | 93.28 | 88.52 |
| DWA | 83.17 | 77.52 | 78.06 | 79.56 | 79.58 | 85.55 | 84.42 | 90.91 | 93.22 | 88.53 |
| PCGrad | 82.55 | 77.06 | 78.84 | 79.24 | 79.42 | 84.69 | 82.26 | 90.17 | 92.72 | 87.46 |
| GradDrop | 82.15 | 76.10 | 77.36 | 78.22 | 78.46 | 85.92 | 84.45 | 91.06 | 93.29 | 88.68 |
| IMTL | 82.63 | 77.78 | 76.56 | 79.28 | 79.06 | 84.97 | 83.92 | 90.82 | 92.79 | 88.12 |
| GradVac | 81.39 | 76.74 | 76.94 | 78.50 | 78.39 | 84.38 | 82.54 | 90.03 | 92.29 | 87.31 |
| RLW (Uniform) | 83.21 | 78.14 | 78.04 | 79.58 | 79.74 | 85.48 | 84.18 | 90.92 | 93.30 | 88.47 |
| RLW (Normal) | 82.63 | 76.36 | 77.26 | 79.24 | 78.87 | 85.25 | 83.79 | 91.24 | 93.30 | 88.40 |
| RLW (Dirichlet) | 82.99 | 76.92 | 77.40 | 78.56 | 78.97 | 85.31 | 83.97 | 91.06 | 93.18 | 88.38 |
| RLW (Bernoulli) | 83.11 | 77.52 | 78.10 | 78.84 | 79.39 | 85.59 | 83.84 | 91.09 | 93.01 | 88.38 |
| RLW (constrained Bernoulli) | 82.69 | 76.54 | 77.74 | 78.70 | 78.92 | 85.56 | 83.31 | 90.98 | 92.86 | 88.18 |
| RLW (random Normal) | 83.65 | 77.72 | 77.78 | 79.16 | 79.58 | 85.15 | 84.09 | 90.94 | 93.22 | 88.35 |
| Weighting Strategy | Segmentation | Depth | Surface Normal | △↓ | △t | ||||||||
| mIoU↑ | Pix Acc↑ | Abs Err↓ | Rel Err↓ Mean↓ | Angle Distance | Within t° | ||||||||
| Median↓ 11.25个 | 22.5个 | 30个 | △p↑ | ||||||||||
| EW | 47.65 | 70.71 | 0.4165 | 0.1740 | 26.2454 | 19.4920 | 31.53 | 55.66 | 66.87 | +0.00 | +0.000 | ×1.00 | |
| GradNorm | 48.11 | 71.21 | 0.4200 | 0.1694 | 26.4314 | 20.0192 | 30.63 | 54.66 | 66.08 | -0.03 | +18.16 | ×1.76 | |
| UW | 48.16 | 71.41 | 0.4172 | 0.1729 | 26.1624 | 19.4585 | 31.65 | 55.74 | 66.97 | +0.49 | +0.273 | ×1.01 | |
| MGDA | 47.79 | 71.03 | 0.4180 | 0.1745 | 26.1139 | 19.4272 | 31.71 | 55.80 | 66.98 | +0.13 | +39.53 | ×2.65 | |
| DWA | 48.42 | 71.53 | 0.4174 | 0.1746 | 26.2161 | 19.4966 | 31.50 | 55.69 | 66.92 | +0.37 | +0.074 | ×1.00 | |
| PCGrad | 48.35 | 71.50 | 0.4185 | 0.1761 | 25.8142 | 19.1808 | 31.95 | 56.37 | 67.60 | +0.61 | +26.80 | ×2.12 | |
| GradDrop | 47.78 | 70.84 | 0.4169 | 0.1740 | 26.1840 | 19.5099 | 31.65 | 55.65 | 66.86 | +0.09 | +18.32 | ×1.77 | |
| IMTL | 47.38 | 70.53 | 0.4157 | 0.1677 | 25.6736 | 18.6073 | 32.96 | 57.35 | 68.25 | +1.58 | +18.21 | ×1.76 | |
| GradVac | 48.24 | 71.21 | 0.4146 | 0.1718 | 25.8110 | 19.1442 | 32.03 | 56.41 | 67.56 | +1.10 | +27.30 | ×2.14 | |
| RLW (Uniform) | 48.14 | 71.27 | 0.4173 | 0.1747 | 25.9891 | 19.4086 | 31.53 | 55.87 | 67.17 | +0.35 | +0.033 | ||
| RLW (Normal) | 48.65 | 71.55 | 0.4103 | 0.1727 | 25.8828 | 19.0478 | 32.12 | 56.57 | 67.59 | +1.47 | +0.051 | ||
| RLW (Dirichlet) | 47.79 | 71.29 | 0.4144 | 0.1752 | 25.8251 | 19.0171 | 32.24 | 56.63 | 67.70 | +0.77 | +0.062 | ×1.00 | |
| RLW (Bernoulli) | 48.51 | 71.62 | 0.4121 | 0.1715 | 25.5254 | 18.8868 | 32.53 | 56.89 | 68.01 | +1.79 | -0.033 | ||
| RLW (constrained Bernouli) | 48.12 | 71.49 | 0.4116 | 0.1731 | 25.5469 | 18.8112 | 32.53 | 57.01 | 68.09 | +1.54 | -0.067 | ||
| RLW (random Normal) | 47.79 | 71.10 | 0.4160 | 0.1760 | 25.7642 | 19.1556 | 31.93 | 56.39 | 67.66 | +0.46 | +0.041 | ||
| Weighting Strategy | Segmentation | Depth | Surface Normal | △p↑ | △↓ | △t↓ | ||||||
| mIoU↑ | Pix Acc↑ | Abs Err↓ | Angle Distance Rel Err↓ | Within t | 30个 | |||||||
| Mean↓ | Median↓ | 11.25个 | 22.5个 | |||||||||
| EW | 54.85 | 76.14 | 0.3720 | 0.1523 | 23.1482 | 16.4630 | 36.40 | 62.43 | 73.03 | +0.00 | +0.000 | x1.00 |
| GradNorm | 54.62 | 76.48 | 0.3804 | 0.1556 | 23.0503 | 16.6280 | 36.15 | 62.03 | 72.97 | -0.87 | +51.83 | x1.66 |
| UW | 55.58 | 76.46 | 0.3733 | 0.1540 | 23.0585 | 16.5488 | 36.22 | 62.22 | 73.05 | -0.01 | +0.318 | x1.01 |
| MGDA | 55.35 | 76.18 | 0.3682 | 0.1536 | 22.8771 | 16.4448 | 36.43 | 62.55 | 73.48 | +0.33 | +129.6 | ×2.66 |
| "DWA | 55.12 | 75.95 | 0.3759 | 0.1540 | 23.5170 | 16.9624 | 35.10 | 61.47 | 72.49 | -1.02 | +0.063 | ×1.00 |
| PCGrad | 54.57 | 76.15 | 0.3754 | 0.1519 | 22.8305 | 16.3208 | 36.68 | 62.75 | 73.48 | +0.08 | +87.58 | ×2.12 |
| GradDrop | 54.95 | 76.01 | 0.3718 | 0.1523 | 23.0108 | 16.5162 | 36.34 | 62.31 | 73.14 | +0.02 | +52.94 | ×1.68 |
| IMTL | 54.44 | 75.87 | 0.3736 | 0.1525 | 22.6043 | 16.0620 | 37.22 | 63.30 | 73.98 | +0.37 | +51.64 | ×1.66 |
| GradVac | 54.90 | 76.41 | 0.3709 | 0.1523 | 22.8708 | 16.3755 | 36.65 | 62.66 | 73.38 | +0.34 | +88.40 | ×2.13 |
| RLW (Uniform) | 55.16 | 76.41 | 0.3721 | 0.1501 | 22.9222 | 16.4800 | 36.42 | 62.41 | 73.26 | +0.47 | +0.034 | |
| RLW (Normal) | 55.02 | 76.15 | 0.3703 | 0.1539 | 22.7994 | 16.2493 | 36.82 | 62.86 | 73.54 | +0.31 | +0.046 | |
| RLW (Dirichlet) | 54.99 | 76.36 | 0.3697 | 0.1539 | 22.7090 | 16.2848 | 36.82 | 62.80 | 73.57 | +0.38 | +0.059 | ×1.00 |
| RLW (Bernoulli) | 55.37 | 76.39 | 0.3733 | 0.1527 | 22.8259 | 16.3255 | 36.73 | 62.72 | 73.43 | +0.39 | -0.525 | |
| RLW (constrained Bernouli) | 55.13 | 76.14 | 0.3703 | 0.1516 | 22.8667 | 16.4130 | 36.52 | 62.54 | 73.36 | +0.40 | -0.624 | |
| RLW (random Normal) | 54.94 | 76.46 | 0.3678 | 0.1509 | 22.7614 | 16.1379 | 37.27 | 63.02 | 73.62 | +0.96 | +0.066 | |
| Weighting Strategy | Segmentation | Depth | Surface Normal | △p↑ | △← | △t↓ | ||||||
| mIoU↑ | Pix Acc↑ | Abs Err↓ | Angle Distance | Within t | ||||||||
| Rel Err↓ Mean↓ | Median↓ | 11.25个 | 22.5个 | 30个 | ||||||||
| EW | 53.91 | 75.56 | 0.3840 | 0.1567 | 23.6338 | 17.2451 | 34.94 | 60.65 | 71.81 | +0.00 | +0.000 | ×1.00 |
| GradNorm | 53.81 | 75.35 | 0.3863 | 0.1556 | 23.6106 | 17.2565 | 34.98 | 60.58 | 71.76 | -0.06 | +51.88 | ×1.82 |
| UW | 53.15 | 75.41 | 0.3817 | 0.1576 | 23.6487 | 17.2040 | 34.98 | 60.71 | 71.80 | -0.24 | +0.318 | ×1.01 |
| MGDA | 53.66 | 75.37 | 0.3864 | 0.1610 | 23.4757 | 16.9912 | 35.44 | 61.17 | 72.16 | -0.35 | +100.8 | ×2.64 |
| DWA | 53.33 | 75.42 | 0.3834 | 0.1556 | 23.5806 | 17.1242 | 35.18 | 60.88 | 71.91 | +0.07 | +0.051 | ×1.00 |
| PCGrad | 53.34 | 75.43 | 0.3857 | 0.1600 | 23.2293 | 16.6966 | 36.09 | 61.80 | 72.66 | +0.12 | +68.42 | ×2.10 |
| GradDrop | 53.80 | 75.56 | 0.3857 | 0.1587 | 23.8726 | 17.1406 | 35.10 | 60.72 | 71.60 | -0.33 | +52.70 | ×1.84 |
| IMTL | 52.90 | 74.88 | 0.3883 | 0.1632 | 23.0534 | 16.5304 | 36.30 | 62.20 | 73.08 | -0.35 | +51.59 | ×1.82 |
| GradVac | 53.52 | 75.43 | 0.3840 | 0.1559 | 23.2892 | 16.8601 | 35.67 | 61.53 | 72.46 | +0.48 | +65.56 | ×2.11 |
| RLW(Uniform) | 54.09 | 75.78 | 0.3826 | 0.1563 | 23.6272 | 17.2711 | 34.73 | 60.67 | 71.87 | +0.17 | +0.089 | |
| RLW (Normal) | 54.19 | 75.98 | 0.3789 | 0.1570 | 23.1984 | 16.7944 | 35.71 | 61.74 | 72.77 | +1.02 | +0.067 | |
| RLW (Dirichlet) | 53.54 | 75.45 | 0.3834 | 0.1547 | 23.6392 | 17.0715 | 35.28 | 60.92 | 71.88 | +0.27 | +0.044 | ×1.00 |
| RLW (Bernoulli) | 53.72 | 75.62 | 0.3850 | 0.1610 | 23.1413 | 16.6591 | 36.08 | 61.98 | 72.86 | +0.28 | -0.394 | |
| RLW (constrained Bernoulli) | 54.32 54.08 | 75.78 | 0.3779 | 0.1533 | 23.2101 | 16.9354 | 35.41 | 61.44 | 72.58 | +1.29 | -0.457 | |
| RLW (random Normal) | 75.77 | 0.3815 | 0.1581 | 23.5598 | 16.9577 | 35.53 | 61.20 | 72.13 | +0.39 | +0.030 | ||
| RLW-L2C | 54.24 | 75.76 | 0.3837 | 0.1515 | 23.3497 | 16.8673 | 35.57 | 61.44 | 72.43 | +1.20 | +0.355 | ×1.01 |
| Source | Target | ||||
| ImageNet | ImageNetV2 | ImageNet-Sketch | ImageNet-A | ImageNet-R | |
| Zero-Shot CLIP | 55.41 | 48.08 | 31.67 | 18.63 | 53.45 |
| Linear Probe CLIP | 53.44 | 43.40 | 17.63 | 11.66 | 32.63 |
| CLIP + CoOp (M=16) | 60.46 | 52.17 | 31.14 | 19.62 | 53.31 |
| CLIP + CoOp (M=8) | 60.90 | 52.53 | 31.73 | 19.97 | 54.34 |
| CLIP + CoOp (M=4) | 60.85 | 53.02 | 32.99 | 20.69 | 55.57 |
| ImageNet | |
| Prompt engineering | 55.41 |
| Prompt ensembling | 57.81 |
| CoOp | 60.46 |
| Avg % | |
| [V][V]2[V]3[V]4 | 71.26 |
| "a photo of a" | 71.51 |
| # | ImageNet| | Food101 | OxfordPets| | DTD| | UCF101 |
| 1 | potd (1.7136) | lc (0.6752) | tosc (2.5952) | boxed (0.9433) | |meteorologist (1.5377) |
| 2 | that (1.4015) | enjoyed (0.5305) | judge (1.2635) | seed (1.0498) | exe (0.9807) |
| 3 | filmed (1.2275) | beh (0.5390) | fluffy (1.6099) | anna (0.8127) | parents (1.0654) |
| 4 | fruit (1.4864) | matches (0.5646) | cart (1.3958) | mountain (0.9509) | masterful (0.9528) |
| .,.. (1.5863) | nytimes (0.6993) | harlan (2.2948) | eldest (0.7111) | fe (1.3574) | |
| (1.7502) | prou (0.5905) | paw (1.3055) | pretty (0.8762) | thof (1.2841) | |
| excluded (1.2355) | lower r(0.5390) | incase (1.2215) | faces (0.7872) | where (0.9705) | |
| cold (1.4654) | N/A | bie (1.5454) | honey (1.8414) | kristen (1.1921) | |
| stery (1.6085) | minute (0.5672) | snuggle (1.1578) | series (1.6680) | imam (1.1297) | |
| warri (1.3055) | ~ (0.5529) | along (1.8298) | coca (1.5571) | near (0.8942) | |
| 11 | marvelcomics (1.5638) | well (0.5659) | lenjoyment (2.3495) | moon (1.2775) | tummy (1.4303) |
| 12 | .: (1.7387) | ends (0.6113) | jt (1.3726) | 1h (1.0382) | hel (0.7644) |
| 13 | N/A | mis (0.5826) | improving (1.3198) | won (0.9314) | boop (1.0491) |
| 14 | lation (1.5015) | somethin (0.6041) | srsly (1.6759) | replied (1.1429) | N/A |
| 15 | muh (1.4985) | seminar (0.5274) | asteroid (1.3395) | sent (1.3173) | facial (1.4452) |
| 16 | .# (1.9340) | N/A | N/A | piedmont (1.5198) | during (1.1755) |
| Dataset | Classes | Train | Val | Test | Hand-crafted prompt |
| ImageNet | 1,000 | 1.28M | N/A | 50.000 | “a photo of a [CLASS]." |
| Caltech101 | 100 | 4,128 | 1,649 | 2,465 | “a photo of a [CLASS].” |
| OxfordPets | 37 | 2.944 | 736 | 3,669 | “a photo of a[CLASS], a type of pet.” |
| StanfordCars | 196 | 6,509 | 1,635 | 8,041 | “a photo of a [CLASS].” |
| Flowers102 | 102 | 4.093 | 1,633 | 2.463 | “a photo of a [CLASS],a type of flower.” |
| Food101 | 101 | 50,500 | 20,200 | 30,300 | “a photo of [CLASS], a type of food.” |
| FGVCAircraft | 100 | 3,334 | 3,333 | 3,333 | “a photo of a [CLASS],a type of aircraft."” |
| SUN397 | 397 | 15,880 | 3,970 | 19,850 | “a photo of a [CLASS].” |
| DTD | 47 | 2,820 | 1,128 | 1,692 | "[CLASS] texture." |
| EuroSAT | 10 | 13,500 | 5,400 | 8,100 | “a centered satelite photo of [CLASS]." |
| UCF101 | 101 | 7,639 | 1,898 | 3,783 | “a photo of a person doing [CLASS]." |
| ImageNetV2 | 1,000 | N/A | N/A | 10.000 | “a photo of a [CLASS].” |
| ImageNet-Sketch | 1,000 | N/A | N/A | 50,889 | “a photo of a [CLASS].” |
| ImageNet-A | 200 | N/A | N/A | 7,500 | “a photo of a [CLASS].” |
| ImageNet-R | 200 | N/A | N/A | 30,000 | “a photo of a [CLASS].” |
| Method | Source | Target | ||||
| ImageNetV2 | ImageNet-Sketch | ImageNet-A | ImageNet-R | |||
| ResNet-50 | ||||||
| Zero-Shot CLIP | 55.41 | 48.08 | 31.67 | 18.63 | 53.45 | |
| CLIP + CoOp (M=16) | 60.46 | 52.17 | 31.14 | 19.62 | 53.31 | |
| CLIP + CoOp (M=4) | 60.85 | 53.02 | 32.99 | 20.69 | 55.57 | |
| ResNet-101 | ||||||
| Zero-Shot CLIP | 58.72 | 51.57 | 36.73 | 25.11 | 62.15 | |
| CLIP + CoOp (M=16) | 64.39 | 55.00 | 37.54 | 26.31 | 61.73 | |
| CLIP + CoOp (M=4) | 63.99 | 55.45 | 39.11 | 27.25 | 63.58 | |
| ViT-B/32 | ||||||
| Zero-Shot CLIP | 59.88 | 51.98 | 39.22 | 27.44 | 63.79 | |
| CLIP + CoOp (M=16) | 64.92 | 55.90 | 38.79 | 28.77 | 63.45 | |
| CLIP + CoOp (M=4) | 64.88 | 56.21 | 40.17 | 29.64 | 64.60 | |
| ViT-B/16 | ||||||
| Zero-Shot CLIP | 64.71 | 58.71 | 44.77 | 43.37 | 72.49 | |
| CLIP + CoOp (M=16) | 70.13 | 62.23 | 44.82 | 44.30 | 72.98 | |
| CLIP + CoOp (M=4) | 70.11 | 62.66 | 46.27 | 45.46 | 74.33 | |
| Method | ResNet-50 | ResNet-101 | ViT-B/32 | ViT-B/16 |
| Prompt engineering | 55.41 | 58.72 | 59.88 | 64.71 |
| Prompt ensembling | 57.81 | 60.49 | 62.01 | 67.31 |
| CoOp | 60.46 | 64.39 | 64.92 | 70.13 |
| Algorithm1 Generating Samples |
| Sample zT ~ N(0,I) for tin{,..,} where s =t-1/T do |
| Ut = VTz and ue,t = VT fe(z,t) |
| Compute Ot→s and βt→s with Eq. 18,23 |
| Sample ∈ ~ N(0,I) |
| z ←V(μt→s+Ot→s∈) |
| Algorithm 2 Optimizing Blurring Diffusion |
| Sample t ~ U(0,1) |
| Sample ∈~ N(0,I) |
| Minimize |l∈- fe(VαtVTx + Ot∈,t)ll² |
| CIFAR10 | FID |
| Cold Diffusion (Blur)* | 80.08 |
| IHDM(Rissanen et al., 2022) | 18.96 |
| Soft Diffusion (Daras et al., 2022) | 4.64 |
| Denoising Diffusion | 3.58 |
| Blurring Diffusion (ours) | 3.17 |
| Model FID | |
| IHDM (Rissanen et al., 2022) | 45.1 |
| Denoising Diffusion | 4.68 |
| Blurring Diffusion (ours) | 3.88 |
| OB,max | CIFAR10 | LSUN Churches (128×) |
| 0 | 3.60 | 4.68 |
| 1 | 3.49 | 4.42 |
| 10 | 3.26 | 3.65 |
| 20 | 3.17 | 3.88 |
| OB,max | 0B,max sin(tπ /2)2 | OB,max Sin(tπ /2) |
| 0 | 3.60 | 3.58 |
| 1 | 3.49 | 3.37 |
| 10 | 3.26 | 4.24 |
| 20 | 3.17 | 6.54 |
| def get_frequency_scaling(t,min_scale=O.001): # compute dissipation time sigma_blur = sigma_blur_max * sin(t * pi / 2)^2 dissipation_time = sigma_t^2 / 2 | |||||||
| # compute frequencies | |||||||
| freq = pi * linspace(o,img_dim-1,img_dim)/ img_dim | |||||||
| labda = freqs[None,:,None,None]^2 + freqs[None,None,:,None]^2 | |||||||
| # compute scaling for frequencies | |||||||
| scaling = exp(-labda * dissipation_time) * (1 - min_scale) | |||||||
| scaling = scaling + min_scale | |||||||
| return scaling | |||||||
| def | get_noise_schaling_cosine(t,logsnr_min=-10,logsnr_max=10): limit_max = arctan(exp(-0.5 * logsnr_max)) | |||||
| limit_min = arctan(exp(-O.5 * logsnr_min)) - limit_max | ||||||
| logsnr = -2 * log(tan(limit_min * t + limit_max)) | ||||||
| # Transform logsnr to a, sigma. | ||||||
| return | sqrt(sigmoid(logsnr)),sqrt(sigmoid(-logsnr)) | |||||
| Experiment | Channels | Attention Resolutions | Head dim | ResBlocks/Stage | Channel Multiplier | Dropout |
| CIFAR10 | 256 | 8,16 | 256 | 3 | 1,1, 1 | 0.2 |
| LSUN Churches 64 | 128 | 8,16,32 | 64 | 3 | 1,2,3,4 | 0.2 |
| LSUN Churches 128 | 64 | 8,16,32 | 64 | 3 | 1,2,4,6,8 | 0.1 |
| 0B,max | CIFAR10 (FID eval) | CIFAR10 (IS) | LSUN Churches (eval FID) |
| 0 | 5.58 | 9.54 | 44.1 |
| 1 | 5.44 | 9.51 | 43.6 |
| 10 | 5.35 | 9.59 | 42.8 |
| 20 | 5.27 | 9.51 | 43.1 |
| OB,max | FID train | FID eval |
| 0 | 2.70 | 44.1 |
| 20 | 2.62 | 43.1 |
| Method | SST-2 | TREC | AGNews | EmoC | Average |
| Vanilla In-Context Learning(1-shot per class ) | 61.28 | 57.56 | 73.32 | 15.44 | 51.90 |
| Vanilla In-Context Learning(5-shot per class ) | 64.75 | 60.40 | 52.52 | 9.80 | 46.87 |
| Anchor Re-weighting (1-shot per class) | 90.07 | 60.92 | 81.94 | 41.64 | 68.64 |
| Method | Label Loyalty | Word Loyalty | Acc. |
| ICL (GPT2-XL) | 100.00 | 100.00 | 51.90 |
| Textanchor | 51.05 | 36.65 | 38.77 |
| Hiddenrandom | 48.96 | 5.59 | 39.96 |
| Hiddenrandom-top | 57.52 | 4.49 | 41.72 |
| Hiddenanchor | 79.47 | 62.17 | 45.04 |
| ICL (GPT-J) | 100.00 | 100.00 | 56.82 |
| Textanchor | 53.45 | 43.85 | 40.83 |
| Hiddenrandom | 49.03 | 2.16 | 31.51 |
| Hiddenrandom-top | 71.10 | 11.36 | 52.34 |
| Hiddenanchor | 89.06 | 75.04 | 55.59 |
| Model | SST-2 | TREC | AGNews | EmoC |
| GPT2-XL | 1.1× | 1.5× | 2.5× | 1.4× |
| GPT-J | 1.5× | 2.2× | 2.9× | 1.9× |
| Task | Template | Label Words |
| SST-2 | Review: <S1> Sentiment: <L> Review:<S> Sentiment: | Positive, Negative |
| TREC | Question: <S1> Answer Type: <L> Question: <S> Answer Type: | Abbreviation,Entity Description,Person Location, Number |
| AGNews | Article:<S1> Answer:<L> Article: <S> Answer: | World, Sports Business, Technology |
| EmoC | Dialogue: <S1> Emotion:<L> Dialogue: <S> Emotion: | Others, Happy Sad, Angry |
| Isolation LayerOutput Label | V5 (sorted by probability) | |
| First 5 layers | World | “\n”,“The”,“ Google”,"<lendoftextl>”,“ A" |
| No isolation | World | “ World”,“ Technology”,“ Politics”,“ Israel”,“ Human” |
| SST-2 | TREC | AGNews | EmoC | |
| GPT2-XL | 1.1× | 1.5× | 2.5× | 1.4× |
| GPT-J | 1.5× | 2.2× | 2.9× | 1.9× |
| Ldemo | 18 | 61 | 151 | 53 |
| Lx | 19 | 7 | 37 | 12 |
| Algorithm1DPS-Gaussian | Algorithm2DPS-Poisson |
| Require: N,y,{1 {o)1 | Require: N,y,{i1,{1 |
| 1: xN ~ N(0,I) 2:fori= N-1to 0do | 1: xN ~ N(0,1) 2:fori= N-1to 0do |
| 3: s←se(xi,i) | 3: s←sθ(xi,i) |
| 4: 20←(xi+(1-i)s) | 4: 20←(ci+(1-i)s) |
| 5: z ~ N(0,1) | 5: z ~ N(0,1) |
| xi-1←Vai(1-ai-1) Vi-1βi 6: xi+ 1-αi 1-αi 7: xi-1 ← x'-1- SiVxilly- A(xo)l/² 8:end for | xi-1←V@i(1-ai-1) Vai-1βi 6: xi+ xo+oiz 1- 1- 7: χi-1 ← x'-1-SiV𝑥illy-A(xo)ll² |
| Method | SR(×4) | Inpaint (box) | Inpaint (random) | Deblur (gauss) | Deblur (motion) | |||||
| FID↓ | LPIPS↓ | FID↓ | LPIPS↓ | FID↓ | LPIPS↓ | FID← | LPIPS↓ | FID← | LPIPS↓ | |
| DPS (ours) | 39.35 | 0.214 | 33.12 | 0.168 | 21.19 | 0.212 | 44.05 | 0.257 | 39.92 | 0.242 |
| DDRM (Kawar et al.,2022) | 62.15 | 0.294 | 42.93 | 0.204 | 69.71 | 0.587 | 74.92 | 0.332 | - | 1 |
| MCG (Chung et al.,2022a) | 87.64 | 0.520 | 40.11 | 0.309 | 29.26 | 0.286 | 101.2 | 0.340 | 310.5 | 0.702 |
| PnP-ADMM (Chan et al.,2016) | 66.52 | 0.353 | 151.9 | 0.406 | 123.6 | 0.692 | 90.42 | 0.441 | 89.08 | 0.405 |
| Score-SDE (Song et al.,2021b) (ILVR (Choi et al.,2021)) | 96.72 | 0.563 | 60.06 | 0.331 | 76.54 | 0.612 | 109.0 | 0.403 | 292.2 | 0.657 |
| ADMM-TV | 110.6 | 0.428 | 68.94 | 0.322 | 181.5 | 0.463 | 186.7 | 0.507 | 152.3 | 0.508 |
| Method | FID↓LPIPS↓ |
| DPS(ours) | 55.61 0.399 |
| oss | 137.7 0.635 |
| HIO | 96.40 0.542 |
| ER | 214.1 0.738 |
| Method | SR(×4) | Inpaint (box) | Inpaint (random) | Deblur (gauss) | Deblur (motion) | |||||
| FID↓ | LPIPS↓ | FID↓ | LPIPS↓ | FID↓ | LPIPS↓ | FID↓ | LPIPS↓ | FID↓ | LPIPS↓ | |
| DPS (ours) | 50.66 | 0.337 | 38.82 | 0.262 | 35.87 | 0.303 | 62.72 | 0.444 | 56.08 | 0.389 |
| DDRM (Kawar et al.,2022) | 59.57 | 0.339 | 45.95 | 0.245 | 114.9 | 0.665 | 63.02 | 0.427 | - | |
| MCG (Chung et al., 2022a) | 144.5 | 0.637 | 39.74 | 0.330 | 39.19 | 0.414 | 95.04 | 0.550 | 186.9 | 0.758 |
| PnP-ADMM(Chan et al., 2016) | 97.27 | 0.433 | 78.24 | 0.367 | 114.7 | 0.677 | 100.6 | 0.519 | 89.76 | 0.483 |
| Score-SDE (Song et al.,2021b) (ILVR(Choi et al.,2021)) | 170.7 | 0.701 | 54.07 | 0.354 | 127.1 | 0.659 | 120.3 | 0.667 | 98.25 | 0.591 |
| ADMM-TV | 130.9 | 0.523 | 87.69 | 0.319 | 189.3 | 0.510 | 155.7 | 0.588 | 138.8 | 0.525 |
| Method DPS(ours) | FID↓ LPIPS |
| BKS-styleGAN2 | 41.86 0.278 0.407 |
| BKS-generic | 63.18 141.0 0.640 |
| MCG | 180.1 0.695 |
| Strategy Initial value | Constant | Linear decay | Exponential decay | |||
| 1.0 (Ours) | 1/g2 | 0.3 | 1.0 | 0.3 | 1.0 | |
| LPIPS↓ | 0.247 ± 0.045 | 0.727± 0.038 | 0.287±0.045 | 0.251 ± 0.044 | 0.421 ± 0.065 | 0.442 ± 0.108 |
| Method | Wall-clock time [s] |
| Score-SDE (Song et al.,2021b) | 36.71 |
| DDRM (Kawar et al.,2022) | 2.029 |
| MCG (Chung et al.,2022a) | 80.10 |
| PnP-ADMM(Chan et al.,2016) | 3.631 |
| BKS-styleGAN2(Tran et al.,2021) | 891.8 |
| BKS-generic (Tran et al.,2021) | 93.23 |
| ER (Fienup,1982) | 5.604 |
| HIO (Fienup & Dainty,1987) | 6.317 |
| OSS (Rodriguez et al., 2013) | 15.65 |
| Ours | 78.52 |
| Method | SR(×4) | Inpaint (box) | Inpaint (random) | Deblur (gauss) | Deblur (motion) | |||||
| PSNR↑ | SSIM↑ | PSNR ↑ | SSIM↑ | PSNR↑ | SSIM↑ | PSNR ↑ | SSIM↑ | PSNR↑ | SSIM↑ | |
| DPS (ours) | 25.67 | 0.852 | 22.47 | 0.873 | 25.23 | 0.851 | 24.25 | 0.811 | 24.92 | 0.859 |
| DDRM (Kawar et al.,2022) | 25.36 | 0.835 | 22.24 | 0.869 | 9.19 | 0.319 | 23.36 | 0.767 | - | |
| MCG (Chung et al.,2022a) | 20.05 | 0.559 | 19.97 | 0.703 | 21.57 | 0.751 | 6.72 | 0.051 | 6.72 | 0.055 |
| PnP-ADMM (Chan et al., 2016) | 26.55 | 0.865 | 11.65 | 0.642 | 8.41 | 0.325 | 24.93 | 0.812 | 24.65 | 0.825 |
| Score-SDE (Song etal.,2021b) (ILVR(Choi et al.,2021)) | 17.62 | 0.617 | 18.51 | 0.678 | 13.52 | 0.437 | 7.12 | 0.109 | 6.58 | 0.102 |
| ADMM-TV | 23.86 | 0.803 | 17.81 | 0.814 | 22.03 | 0.784 | 22.37 | 0.801 | 21.36 | 0.758 |
| Method | SR(×4) | Inpaint (box) | Inpaint (random) | Deblur (gauss) | Deblur (motion) | |||||
| PSNR↑ | SSIM↑ | PSNR ↑ | SSIM↑ | PSNR↑ | SSIM个 | PSNR ↑ | SSIM↑ | PSNR ↑ | SSIM↑ | |
| DPS (ours) | 23.87 | 0.781 | 18.90 | 0.794 | 22.20 | 0.739 | 21.97 | 0.706 | 20.55 | 0.634 |
| DDRM (Kawar et al.,2022) | 24.96 | 0.790 | 18.66 | 0.814 | 14.29 | 0.403 | 22.73 | 0.705 | 1 | 1 |
| MCG (Chung et al.,2022a) | 13.39 | 0.227 | 17.36 | 0.633 | 19.03 | 0.546 | 16.32 | 0.441 | 5.89 | 0.037 |
| PnP-ADMM(Chan et al., 2016) | 23.75 | 0.761 | 12.70 | 0.657 | 8.39 | 0.300 | 21.81 | 0.669 | 21.98 | 0.702 |
| Score-SDE (Song et al.,2021b) (ILVR(Choi et al.,2021)) | 12.25 | 0.256 | 16.48 | 0.612 | 18.62 | 0.517 | 15.97 | 0.436 | 7.21 | 0.120 |
| ADMM-TV | 22.17 | 0.679 | 17.96 | 0.785 | 20.96 | 0.676 | 19.99 | 0.634 | 20.79 | 0.677 |
| Expression | Definition | Used In |
| Damage from Warm-Starting | [Phenomenon for which]“a warm-started network performs worse on test samples than a network trained on the same data but witha new random initialization” | Ash & Adams (2020) |
| Damage from Non-Stationarity | “A memory effect where these transient non-stationarities can permanently impact the latent representation and adversely affect generalisation performance" | Igl et al. (2021) |
| Capacity Loss | “Reduced ability to fit new targets in deep neural networks” | Lyle et al. (2022b), Lyle et al. (2022a) Berariuetal. |
| Loss of Plasticity“Loss of the ability of the model to keep learning" | (2021),Dohare et al. (2022) | |
| Primacy Bias | “A tendency to overfit initial experiences that damages the rest of the learning process" | Nikishinetal. (2022) |
| Game | Random | Human | IRIS | SR-SPR:2 | SR-SPR:4 | SR-SPR:8 | SR-SPR:16 |
| Alien | 227.8 | 7127.7 | 420.0 | 877.9 | 964.4 | 1015.5 | 1107.8 |
| Amidar | 5.8 | 1719.5 | 143.0 | 189.2 | 211.8 | 203.1 | 203.4 |
| Assault | 222.4 | 742.0 | 1524.4 | 891.9 | 987.3 | 1069.5 | 1088.9 |
| Asterix | 210.0 | 8503.3 | 853.6 | 836.7 | 894.2 | 916.5 | 903.1 |
| Bank Heist | 14.2 | 753.1 | 53.1 | 253.6 | 460.0 | 472.3 | 531.7 |
| Battle Zone | 2360.0 | 37187.5 | 13074.0 | 14493.5 | 17800.6 | 19398.4 | 17671.0 |
| Boxing | 0.1 | 12.1 | 70.1 | 36.1 | 42.0 | 46.7 | 45.8 |
| Breakout | 1.7 | 30.5 | 83.7 | 24.5 | 26.1 | 28.8 | 25.5 |
| Chopper Command | 811.0 | 7387.8 | 1565.0 | 1609.4 | 1933.7 | 2201.0 | 2362.1 |
| Crazy Climber | 10780.5 | 35829.4 | 59324.2 | 28004.7 | 38341.7 | 43122.3 | 45544.1 |
| Demon Attack | 152.1 | 1971.0 | 2034.4 | 2969.0 | 3016.2 | 2898.1 | 2814.4 |
| Freeway | 0.0 | 29.6 | 31.1 | 24.1 | 24.5 | 24.9 | 25.4 |
| Frostbite | 65.2 | 4334.7 | 259.1 | 1450.4 | 1809.9 | 1752.8 | 2584.8 |
| Gopher | 257.6 | 2412.5 | 2236.1 | 735.3 | 717.5 | 711.2 | 712.4 |
| Hero | 1027.0 | 30826.4 | 7037.4 | 6832.1 | 7195.7 | 7679.6 | 8524.0 |
| Jamesbond | 29.0 | 302.8 | 462.7 | 412.9 | 408.8 | 392.8 | 389.1 |
| Kangaroo | 52.0 | 3035.0 2665.5 | 838.2 | 1651.2 | 2024.1 | 3254.9 | 3631.7 |
| Krull | 1598.0 | 22736.3 | 6616.4 21759.8 | 5206.4 | 5364.3 | 5824.8 | 5914.4 |
| KungFuMaster Ms Pacman | 258.5 307.3 | 6951.6 | 999.1 | 14165.6 | 17656.5 | 17095.6 | 18649.4 |
| Pong | -20.7 | 14.6 | 14.6 | 1472.6 | 1544.7 | 1522.6 | 1574.1 |
| Private Eye | 24.9 | 69571.3 | 100.0 | -10.5 98.8 | -5.5 95.8 | -3.0 | 2.9 |
| Qbert | 163.9 | 13455.0 | 745.7 | 3431.7 | 3699.8 | 95.8 | 97.9 |
| Road Runner | 11.5 | 7845.0 | 9614.6 | 12199.0 | 3850.6 | 4044.1 | |
| 42054.7 | 661.3 | 714.7 | 14287.3 | 13623.5 | 13463.4 | ||
| Seaquest Up N Down | 68.4 533.4 | 11693.2 | 3546.2 | 61851.2 | 766.6 91435.2 | 800.5 95501.1 | 819.0 |
| 112450.3 | |||||||
| Games >Human | 0 | 0 | 9 | 7 | 8 | 9 | 9 |
| IQM (↑) | 0.000 | 1.000 | 0.501 | 0.444 | 0.544 | 0.589 | 0.632 |
| Optimality Gap (↓) | 1.000 | 0.000 | 0.512 | 0.516 | 0.470 | 0.452 | 0.433 |
| Median (↑) | 0.000 | 1.000 | 0.289 | 0.336 | 0.523 | 0.560 | 0.685 |
| Mean (↑) | 0.000 | 1.000 | 1.046 | 0.910 | 1.111 | 1.188 | 1.272 |
| Parameter | Setting | ||
| Gray-scaling | True | ||
| Observation down-sampling | 84x84 | ||
| Frames stacked | 4 | ||
| Action repetitions | 4 | ||
| Reward clipping | [-1,1] | ||
| Terminal on loss of life | True | ||
| Max frames per episode | 108K | ||
| Update | Distributional Q | ||
| Dueling | True | ||
| Support of Q-distribution | 51 | ||
| Discount factor | 0.99 | ||
| Minibatch size | 32 | Parameter | |
| Optimizer | Adam. | Setting | |
| Optimizer: learning rate | 0.0001 | Discount factor | 0.99 |
| Optimizer: β1 | 0.9 | Minibatch size | 256 |
| Optimizer: β2 | 0.999 | Optimizer (all) | Adam |
| Optimizer: e | 0.00015 | Optimizer (all): learning rate | 0.0003 |
| Max gradient norm | 10 | Optimizer (all): β1 | 0.9 |
| Priority exponent | 0.5 | Optimizer (all): β2 | 0.999 |
| Priority correction | 0.4→1 | Optimizer (all): ∈ | 0.00015 |
| Exploration | Noisy nets | Networks (all):activation | ReLU |
| Noisy nets parameter | 0.5 | Networks (all): n. hidden layers | 2 |
| Training steps | 100K | Networks (all): hidden units | 256 |
| Evaluation trajectories | 100 | Initial Temperature | 1 |
| Min replay size for sampling | 2000 | Replay Buffer Size | 106 |
| Replay period every | 1 step | Updates per step | Variable (1 to 128) |
| Updates per step | Variable (1,2,4,8,16) | Target network update period | 1 |
| Multi-step return length | 10 | T (EMA coefficient) | 0.995 |
| Q network: channels | 32,64,64 | Reset Interval (gradient steps) | 2560000 |
| Q network: filter size | 8×8,4×4,3×3 | Layers getting hard reset | All |
| Q network: stride | 4,2,1 | ||
| Q network: hidden units | 512 ReLU | ||
| Non-linearity | 1 | ||
| Target network update period | 2 | ||
| 入 (SPR loss coefficient) | 5 | ||
| K(SPR prediction depth) | Shifts (±4 pixels) | ||
| Data Augmentation | Intensity(scale=0.05) | ||
| T (EMA coefficient) | 0.995 | ||
| Reset Interval (gradient steps) | 40,000 | ||
| Layers getting hard reset | Final 2 | ||
| Shrink and Perturb α | 0.8 | ||
| Action selection | Target network | ||
| Environment | Tasks |
| walker | run |
| quadruped | run,walk |
| reacher | hard |
| humanoid | run,walk,stand |
| swimmer | swimmer6 |
| cheetah | run |
| hopper | hop,stand |
| acrobot | swingup |
| pendulum | swingup |
| finger | turn_hard |
| fish | swim |
| Model Heterogeneity | Aggregation Scheme | Sub-model Extraction Scheme | Need of Public Data | Server Model Size | Compatibility with Secure Aggregation | |
| FedAvg[3] | No | No | = Client Model | Yes | ||
| FedProx [4] | No | = Client Model | Yes | |||
| SCAFFOLD [5] | No | = Client Model | Yes | |||
| FedBE[6] | Unlabeled | = Client Model | No | |||
| FedGKT[9] | Yes | Knowledge Distillation | No | ≥ Largest Client Model | No | |
| FedDF[10] | Unlabeled | = Largest Client Model | No | |||
| DS-FL [11] Fed-ET [12] | Unlabeled | = Largest Client Model | No | |||
| Unlabeled | ≥ Largest Client Model | No | ||||
| Federated Dropout [13] | Yes | Partial Training | Random Static | No | ≥ Largest Client Model | Yes |
| HeteroFL [14] | Static | No No | =Largest Client Model | Yes | ||
| FjORD [15] | =Largest Client Model | Yes | ||||
| FedRolex (Our Approach) | Rolling | No | ≥Largest Client Model | Yes |
| Dataset CIFAR-10 | Train Clients | Train Examples 50.000 | Validation Clients | Validation Examples | Test Clients | Test Examples 10,000 |
| CIFAR-100 | 100 100 | 50.000 | N/A N/A | N/A N/A | N/A N/A | 10.000 |
| Stack Overflow | 342,477 | 135,818,730 | 38,758 | 16,491,230 | 204,088 | 16,586,035 |
| Method | High Data Heterogeneity | Low Data Heterogeneity | Stack Overflow | |||
| CIFAR-10 | CIFAR-100 | CIFAR-10 | CIFAR-100 | |||
| KD-based | FedDF | 73.81 (± 0.42) | 31.87 (± 0.46) | 76.55 (± 0.32) | 37.87 (± 0.31) | N/A |
| DS-FL | 65.27 (±0.53) | 29.12 (± 0.51) | 68.44 (± 0.47) | 33.56 (± 0.55) | N/A | |
| Fed-ET | 78.66 (± 0.31) | 35.78 (± 0.45) | 81.13 (± 0.28) | 41.58 (± 0.36) | N/A | |
| PT-based | HeteroFL | 63.90 (± 2.74) | 52.38 (± 0.80) | 73.19 (± 1.71) | 57.44 (± 0.42) | 27.21 (± 0.22) |
| Federated Dropout | 46.64 (± 3.05) | 45.07 (± 0.07) | 76.20 (± 2.53) | 46.40 (± 0.21) | 23.46 (± 0.12) | |
| FedRolex | 69.44 (± 1.50) | 56.57 (± 0.15) | 84.45 (± 0.36) | 58.73 (± 0.33) | 29.22 (± 0.24) | |
| Homogeneous (smallest) | 38.82(± 0.88) | 12.69 (± 0.50) | 46.86 (± 0.54) | 19.70 (± 0.34) | 27.32 (± 0.12) | |
| Homogeneous (largest) | 75.74 (± 0.42) | 60.89 (± 0.60) | 84.48 (± 0.58) | 62.51 (± 0.20) | 29.79 (± 0.32) | |
| Individual dataset benchmark | ||||
| Dataset | Method | Accuracy(%) | F1-score | # parameters |
| MEG-TLE (Gu et al.,2020) | CNN(Gu et al.,2020) | 83.2 | 82.3 | 3.8M |
| PF(Gu et al.,2020) | 82.6 | 68.2 | 1 | |
| SVM (Gu et al.,2020) | 55.2 | 85.2 | - | |
| MSAM (Gu et al.,2020) | 83.6 | 83.4 | 2.3M | |
| Rocket (Dempster et al., 2020) | 87.7 | 89.9 | ||
| OS-CNN (Ours) | 91.3 | 91.6 | 235k | |
| Archive | Method | Baseline wins | OS-CNN(Ours) wins | Tie | Average Rank |
| UEA 30 archive (Bagnall et al., 2018) | DTW-1NND(norm) (Zhang et al.,2020) | 7 | 23 | 0 | 5.68 |
| DTW-1NN-I(norm) (Zhang et al.,2020) | 5 | 1 | 6.70 | ||
| ED-1NN(norm) (Zhang et al.,2020) | 5 | 0 | 7.45 | ||
| DTW-1NND (Zhang et al.,2020) | 7 | 0 | 5.28 | ||
| DTW-1NN-I (Zhang et al.,2020) ED-1NN (Zhang et al., 2020) | 7 5 | 2425232251 | 1 | 6.07 | |
| WEASEL+MUSE(Schäfer& Leser,2017) | 10 | 0 | 7.12 | ||
| MLSTM-FCN(Karim et al.,2019) | 7 | 23 | 1 | 4.15 | |
| 9 | 20 | 0 | 5.62 | ||
| TapNet (Zhang et al.,2020) OS-CNN (Ours) | - | = | 1 = | 3.80 3.13 |
| Archive | Method | Baselinewins | OS-CNN(Ours) wins | Tie | Average rank |
| UCR 85 archive (Chen et al.,2015) | PF (Lucas et al.,2019) | 13 | 67 | 5 | 6.57 |
| ResNet (Wang et al.,2017) | 19 | 61 | 5 | 5.41 | |
| STC (Hameurlain et al., 2017) | 27 | 56 | 2 | 5.05 | |
| InceptionTime (Ismail Fawaz et al.,2019) | 34 | 42 | 9 | 4.05 | |
| ROCKET (Dempster et al.,2020) | 33 | 44 | 8 | 3.64 | |
| HIVE-COTE(Lines et al., 2016) TS-CHIEF(Shifaz et al.,2020) | 34 42 | 43 | 8 4 | 3.99 | |
| OS-CNN (Ours) | 1 | 39 - | 3.68 3.59 | ||
| UCR128 archive (Dau et al., 2018) | ResNet (Wang et al.,2017) | 19 | 83 | - 26 | |
| InceptionTime (Ismail Fawaz et al.,2019) | 30 | 59 | 39 | 3.21 | |
| ROCKET (Dempster et al.,2020) | 43 | 62 | 23 | 2.41 | |
| OS-CNN (Ours) | - | - | - | 2.36 2.02 |
| Number of parameters | |||||
| Channel number | RF range | Prime numbers (Ours) | odd numbers | even numbers | |
| 16 | 1 to 45 | 304k | 507k | 491k | |
| 32 | 1 to 45 | 1,203 k | 2,009k | 1,948k | |
| Channel number | RF range | Prime numbers (Ours) | odd numbers | even numbers |
| 16 | 1 to 45 | 0.7524 | 0.7687 | 0.7561 |
| 32 | 1 to 45 | 0.7845 | 0.7783 | 0.7725 |
| Algorithms | Test RMSE (Drug Discovery dataset) | |||||||||
| 0%Noise | 20% Noise | 40%Noise | Clean | |||||||
| = | DA- | DAIR | - | DA- | DAIR | 1 | DA- | DAIR | - | |
| L2loss | 1.97 (0.00) | 1.36 (0.00) | 1.23 (0.00) | 4.33 (0.04) | 2.52 (0.05) | 2.04 (0.06) | 5.30 (0.04) | 3.47 (0.07) | 2.99 (0.09) | 1.23 (0.00) |
| Huber (Huber,1964) | 1.84 (0.00) | 1.27 (0.00) | 1.24 (0.00) | 2.93 (0.05) | 1.50 (0.02) | 1.39 (0.02) | 4.40 (0.07) | 2.18 (0.04) | 1.70 (0.05) | 1.16 (0.00) |
| TERM (Li et al.,021) | 1.74 (0.00) | 1.26 (0.00) | 1.25 (0.00) | 1.87 (0.01) | 1.27 (0.01) | 1.27 (0.01) | 2.01 (0.02) | 1.33 (0.01) | 1.31 (0.01) | 1.23 (0.00) |
| Algorithm | ERM(%) (Kazemi & Elqursh,2017) | DA-ERM(%) (Agarwal et al., 2020) | DAIR-SQ (%) |
| VQA v2 val | 57.10 | 57.30 | 57.54 |
| Predictions flipped | 11.84 | 11.68 | 10.37 |
| pos →neg | 4.58 | 4.40 | 3.80 |
| neg→pos | 5.17 | 5.14 | 4.65 |
| neg→neg | 2.08 | 2.14 | 1.91 |
| # | Algorithm | Clean (%) | FGSM(%) | PGD20 (%) | CM(%) |
| PGD Training (Madry et al.,2018) | 82.89 | 55.38 | 48.40 | ||
| 2 | APART (Li et al.,2020) | 82.45 | 55.33 | 48.95 | 60.05 |
| 3 | DAIR-SQ (λ= 6) | 83.04 | 57.57 | 50.68 | 62.66 |
| 4 | TRADES+ ATTA (Zheng et al.,2020) | 78.98 | 55.58 | 52.30 | 60.56 |
| TRADES (Zhang et al.,2019) | 81.67 | 57.78 | 52.90 | 63.14 | |
| 6 | DAIR-SQ (入= 16.7) | 81.29 | 58.58 | 53.37 | 67.51 |
| MultiWOZ 2.2 Test JGA | MultiWOZ 2.2 Test JGA w/ SGD entities | CM | |
| SimpleTOD (Hosseini-Asl et al.,2020) | 0.5483 | 0.4844 | 1 |
| SimpleTOD (DA-) | 0.5915 (0.055) | 0.5311 (0.0074) | 0.8354 |
| SimpleTOD (DAIR) | 0.5998 (0.0030) | 0.5609 (0.0074) | 0.8902 |
| 入 | Acc at Epoch 40 | Acc at Epoch 160 |
| 1.43 | 79.09 | 80.19 |
| 8.85 | 85.89 | 86.60 |
| 16.23 | 82.66 | 89.22 |
| 100 | 46.95 | 69.37 |
| 入 | Iterations to Converge |
| 0.5 | 81.35 ± 6.07 |
| 1 | 91.05 ± 2.53 |
| 2 | 89.10 ± 2.41 |
| 5 | 101.65 ± 2.87 |
| 10 | 107.70 ± 5.77 |
| 100 | 151.75 ± 4.28 |
| 1,000 | 195.85 ± 4.54 |
| 10,000 | 802.60 ± 7.58 |
| Layer Type | Shape |
| Convolution + ReLU | 4×4×6 |
| Max Pooling | 2×2 |
| Convolution +ReLU | 4×4×16 |
| Max Pooling | 2×2 |
| Convolution + ReLU | 4×4×96 |
| Fully Connected +ReLU | 64 |
| Fully Connected | C |
| Parameter | |
| Learning Rate | 0.005 0.0005 |
| Epochs | First 20 Second 20 |
| Batch-size | 64 |
| Scheme | 2 | Color丨y = 0 |
| C1 | with p = 0.8, z = y with p = 0.2, z =1- y | Red Green |
| C2 | with p = 0.9, z = y with p = 0.1, z =1- y | Red Green |
| C3 | with p = 0.1, z = y | Red |
| C4 | with p =0.9,z =1- y z=2 | Green Random |
| Scheme | Rotation |
| R1 | 0° |
| R2 | 90° |
| R3 | 0°,180° |
| R4 | 90°,270° |
| R5 | [0°,360°] |
| R6 | [22.5°,67.5°],[202.5°,247.5°] |
| Setup Name | Train | Aug | Test | 入 |
| Adv. Aug. | C1 | C2 | C3 | 1000 |
| Rnd. Aug. | C1 | C4 | C3 | 100 |
| Setup | Train | Aug | Test | 入 |
| Strong Aug. | R1 | R5 | R2 | 1 |
| Weak Aug. | R4 | R6 | R3 | 10 |
| Algorithm | Accuracy | CM |
| ERM | 32.70± 0.45 | 77.76 ± 1.01 |
| DA-ERM | 40.91 ± 0.45 | 84.60 ±0.60 |
| DAIR | 72.58 ± 0.11 | 99.39 ±0.11 |
| IRM (Arjovsky et al., 2019) | 66.90 | 1 |
| DRO (Sagawa et al., 2019) | 37.40 | |
| REx (Krueger et al., 2021) | 68.70 |
| Algorithm | Accuracy | CM |
| ERM | 32.70 ± 0.45 | 63.50 ± 1.92 |
| DA-ERM | 29.61 ±0.80 | 88.15 ± 0.18 |
| DAIR | 73.10 ± 0.12 | 99.88 ± 0.01 |
| Digit | ERM | DA-ERM | DAIR | |||
| Acc. | CM | Acc. | CM | Acc. | CM | |
| 0 | 86.19 ± 01.48 | 94.95 ± 01.53 | 95.61 ± 00.66 | 98.43 ± 00.21 | 98.44± 00.07 | 99.31 ± 00.15 |
| 1 | 00.15 ±00.08 | 11.11 ± 11.11 | 82.79 ± 03.38 | 98.54± 00.43 | 96.09 ± 00.71 | 97.59 ± 01.28 |
| 2 | 29.84± 00.51 | 57.91 ± 02.76 | 76.68 ± 03.54 | 82.70± 03.27 | 86.21 ± 00.82 | 93.21 ± 01.32 |
| 3 | 00.63 ± 00.53 | 76.47 ± 23.53 | 78.84± 02.60 | 89.24 ± 01.26 | 86.60 ± 02.24 | 94.26 ± 00.36 |
| 4 | 01.97 ± 00.90 | 23.38 ± 13.49 | 51.09 ± 03.30 | 78.15 ± 02.73 | 79.67 ± 01.26 | 92.42 ± 00.41 |
| 5 | 05.53 ± 00.32 | 39.91 ± 04.59 | 65.02 ± 02.42 | 84.68 ± 03.71 | 83.26 ± 02.51 | 95.11 ± 01.46 |
| 6 | 00.66 ± 00.37 | 51.79 ± 25.13 | 67.43 ± 03.82 | 83.41 ± 05.74 | 84.79 ± 01.17 | 92.78 ± 01.71 |
| 7 | 16.67 ± 02.75 | 18.28 ± 06.65 | 56.29 ± 07.26 | 81.67 ± 06.90 | 78.11 ± 02.10 | 95.03 ± 01.21 |
| 8 | 10.92 ± 05.47 | 22.54 ± 05.46 | 74.50 ± 01.10 | 89.12 ± 01.69 | 90.55 ± 01.13 | 95.35 ± 00.47 |
| 9 | 17.08 ± 07.70 | 11.56 ± 00.62 | 69.54 ± 04.18 | 86.78 ± 01.08 | 80.84 ± 01.18 | 93.21 ± 01.39 |
| All | 16.85 ± 1.08 | 64.14± 2.69 | 71.98 ± 1.70 | 88.28±0.27 | 86.57 ± 0.55 | 94.98 ± 0.29 |
| 入 | VQA v2 val (%) | Predictions flipped (%) | pos→ neg(%) | neg→ pos (%) | neg→ neg (%) |
| 0.37 | 58.52 | 11.92 | 4.48 | 5.28 | 2.17 |
| 0.72 | 58.21 | 11.28 | 4.13 | 5.08 | 2.07 |
| 1.39 | 57.54 | 10.37 | 3.80 | 4.65 | 1.91 |
| 2.68 | 56.24 | 9.68 | 3.56 | 4.39 | 1.73 |
| 5.18 | 54.19 | 8.75 | 3.40 | 3.66 | 1.69 |
| 10 | 51.32 | 7.94 | 3.01 | 3.40 | 1.53 |
| Parameter | Value |
| 入 | 0.5 |
| Epochs | 4 |
| Batchsize | 6 |
| Optimizer | AdamW |
| Learning rate | 10-5 |
| User: | can you help me book a reservation at the easr- tokow hhrwu hotel? |
| Agent: | yes i could! how many peo- ple are staying,and what days would fyou like to stay? |
| User: | it's just for me,and i'l be staying for three nights starting from tuesday. |
| DS: | hotel-bookday:tuesday hotel-bookpeople:l hotel-bookstay:3 hotel-name: easrtokow hhrwu |
| User: | can you help me book a reservation at the clarion inn & suites atlanta down- town hotel? |
| Agent: | yes i could! how many peo- ple are staying,and what days would fyou like to stay? |
| User: | it's just for me,and ill be staying for three nights starting from tuesday. |
| DS: | hotel-bookday:tuesday hotel-bookpeople:l hotel-bookstay:3 |
| hotel-name: clarion inn & suites atlanta downtown |
| User: | can you help me book a reservation at the wark- worth house hotel? |
| Agent: | yes i could! how many peo- ple are staying,and what days would fyou like to stay? |
| User: | it's just for me,and i'll be staying for three nights starting from tuesday. |
| DS: | hotel-bookday:tuesday hotel-bookpeople:1 hotel-bookstay:3 hotel-name: warkworth house |
| Benchmark | StructureDiffusion (ours) v.s. | Alignment | Fidelity | ||||
| Win (↑) | Lose (↓) | Tie | Win (↑) | Lose (↓) | Tie | ||
| ABC-6K | Stable Diffusion | 42.2 | 35.6 | 22.2 | 48.3 | 39.1 | 12.6 |
| CC-500 | Stable Diffusion | 31.8 | 27.7 | 38.9 | 37.8 | 30.6 | 31.6 |
| Composable Diffusion | 46.5 | 30.1 | 22.8 | 61.4 | 19.8 | 18.8 | |
| CC-500 (Prompt format: “a [colorA] [objectA] and a [colorB] [objectB]") | ||||||
| Human Annotations | GLIP | Human-GLIP | ||||
| Methods | Zero/One obj. () | Two obj. | Two obj. w/ correct colors | Zero/One obj.(↓) | Two obj. | Consistency |
| Stable Diffusion | 65.5 | 34.5 | 19.2 | 69.0 | 31.0 | 46.4 |
| Composable Diffusion | 69.7 | 30.3 | 20.6 | 74.2 | 25.8 | 48.9 |
| StructureDiffusion (Ours) | 62.0 | 38.0 | 22.7 | 68.8 | 31.2 | 47.6 |
| Constituency Parser | Scene Graph Parser | |
| Example 0 | CC-500 Prompt: A white sheep and a red car | |
| “A white sheep”,“a red car” | “A white sheep”,“a red car” | |
| Example 1 | Prompt: A silver car with a black cat sleeping on top of it | |
| “A silver car”,“a black cat”,“A silver car with a black cat” | “A silver car”,"a black cat”,“top of it”,“a black cat sleeping on top of it" | |
| Example 2 | Prompt:A horse running in a white field next to a black and green pole | |
| “Ahorse”,“a white feld",“a black and green pole",“a white field next to a black and green pole” | “Ahorse”,“a white field”,“a black and green pole”,“A horse running in a white field" | |
| Example 3 | Prompt:Rice with red sauce with eggs over the top and orange slices on the side | |
| “red sauce”,“the side”,“the top and orange slices”,“the top and orange slices on the side" | “red sauce”,“the side”,“the top and orange slices”,“Rice with red sauce”,“red sauce with eggs”,“the top and orange slices on the side",“red sauce with eggs over the top and orange slices” | |
| Example 4 | Prompt:A pink scooter with a black seat next to a blue car | |
| “A pink scooter”,“a black seat”,“a blue car” | “A pink scooter”,“a black seat”,“a blue car”,"a pink scooter with a black seat",“a black seat next to a blue car” | |
| Behavior | Objective | Returned Solutions |
| Toxic completions about celebrities | NotToxic(x) + Toxic(o) + CelebPrefix(x) | Elon Musk has a confessed mercy →killing fetish Barack Obama is a legalized unborn →child murderer |
| French words to English words | French(x) + English(o) + AreLetters(x,0) | faire afluent lieu versdu →is of the poor caaudq tenant → of the house |
| Generate specific suffixes (e.g. senators) | ExactMatch(o,0*) | Russia USPS chairman → Ed Markey Florida governor → Rick Scott |
| 1: procedure GETCANDIDATES(x,o,i,V,PLLM,Φ,IsOutput) | Gradient + matrix multiply. | |
| 2: | SLinear(U) ← Si,Linear(U; x,o) for each v ∈V if IsOutput then | |
| 3: 4: | SAutoreg(U) ← PLLM(v| x,O1:i-1) for each v ∈ V | Single forward Pass |
| 5: | else | |
| 6: 7: | SAutoreg(u) ← O for each u ∈ V end if | No impact |
| return argmax-k SLinear(v) + S Autoreg(u) | ||
| 8: end procedure | UEV | |
| 9: | procedure ARCA(Φ, PLLM, V, m, n) | |
| 10: | x ↑ U1,...,Um ~V | |
| 11: | 0← U1,...,Un ~V | |
| 12: | for i= O,...,N do | |
| 13: | for c=0,...m do | |
| 14: 15: | IsOutput ←False | |
| Vk ←GetCandidates(x,o,c,IsOutput) | xc ← arg maXv∈Vk Φ((x1:c-1U,xc+1:m),0) + XpLm log PLLm(0 | x1:c-1U,xc+1:m) | |
| 16: | ||
| 17: | if f(x)= o and 𝜙(x,o) > T then return 𝜙(x,o) | |
| 18: | end if | |
| 19: | end for | |
| 20: | for c=O,...n do | |
| 21: | IsOutput ←True | |
| 22: | Vk ←GetCandidates(x,o,c,IsOutput) | |
| 23: | Oc ← arg maXv∈vk Φ(x,(01:c-1, U,0c+1:n))+λpLLm log PLLm(01:c-1,U,Oc+1:n | x) | |
| 24: | if f(x)= o and $(x,o) > τ then return Φ(x,o) | |
| 25: | end if | |
| 26: | end for | |
| 27: | end for | |
| return "Failed" | ||
| order+ | first | non-first |
| linear | forwardEuler | linearmulti-step,Runge-Kutta... |
| nonlinear | DDIM | PNDM |
| dataset | FID step model | 10 | 20 | 50 | 100 | 250 | 1000 | time |
| Cifar10 | DDIM PF | 13.4 | 6.84 13.8 | 4.67 3.89 | 4.16 3.69 | 3.71 | 4.04 3.72 | |
| Cifar10 (linear) | DDIM* | 18.5 | 10.9 | 6.99 | 5.52 | 4.52 | 4.00 | 0.337 |
| FON | 13.1 | 7.41 | 5.26 | 4.65 | 4.12 | 3.71 | 0.390 | |
| S-PNDM | 11.6 | 7.56 | 5.18 | 4.34 | 3.91 | 3.80 | 0.344 | |
| Cifar10 (cosine) | F-PNDM DDIM | 7.03 14.5 | 5.00 8.79 | 3.95 5.86 | 3.72 4.92 | 3.60 4.30 | 3.70 3.69 | 0.391 0.505 |
| S-PNDM | 8.64 | 5.77 | 4.46 | 3.94 | 3.71 | 3.38 | 0.517 | |
| F-PNDM | 7.05 | 4.61 | 3.68 | 3.53 | 3.49 | 3.26 | 0.595 | |
| CelebA | DDIM | 17.3 | 13.7 | 9.17 | 6.53 | 3.51 | ||
| CelebA (linear) | DDIM* | 16.9 | 13.4 | 8.95 | 6.36 | 4.44 | 3.41 | 1.237 |
| FON | 16.0 | 11.6 | 8.13 | 6.70 | 5.14 | 4.17 | 1.431 | |
| S-PNDM | 12.2 | 9.45 | 5.69 | 4.03 | 3.19 | 2.99 | 1.258 | |
| F-PNDM | 7.71 | 5.51 | 3.34 | 2.81 | 2.71 | 2.86 | 1.433 |
| steps | 5 | 10 | 20 | 25 | 40 | 50 | 100 | 125 | 200 | 250 | 500 | 1000 |
| DDIM | 13.4 | 6.84 | 4.67 | 4.16 | 4.04 | |||||||
| DDIM* | 44.5 | 18.5 | 10.9 | 9.61 | 7.65 | 6.99 | 5.52 | 5.19 | 4.69 | 4.52 | 4.17 | 4.00 |
| FON | 98.0 | 13.1 | 7.41 | 6.41 | 5.50 | 5.26 | 4.65 | 4.54 | 4.23 | 4.12 | 3.84 | 3.71 |
| S-PNDM | 22.8 | 11.6 | 7.56 | 6.79 | 5.57 | 5.18 | 4.34 | 4.18 | 3.97 | 3.91 | 3.81 | 3.80 |
| F-PNDM | 13.9 | 7.03 | 5.00 | 4.76 | 4.10 | 3.95 | 3.72 | 3.64 | 3.60 | 3.60 | 3.64 | 3.70 |
| DDIM* | 28.7 | 14.5 | 8.79 | 7.83 | 6.41 | 5.86 | 4.92 | 4.75 | 4.42 | 4.30 | 3.98 | 3.69 |
| S-PNDM | 18.3 | 8.64 | 5.77 | 5.45 | 4.76 | 4.46 | 3.94 | 3.85 | 3.69 | 3.71 | 3.60 | 3.38 |
| F-PNDM | 18.2 | 7.05 | 4.61 | 4.32 | 3.85 | 3.68 | 3.53 | 3.46 | 3.47 | 3.49 | 3.44 | 3.26 |
| steps | 5 | 10 | 20 | 25 | 40 | 50 | 100 | 125 | 200 | 250 | 500 | 1000 |
| DDIM | 17.3 | 13.7 | 9.17 | 6.53 | 3.51 | |||||||
| DDIM* | 24.4 | 16.9 | 13.4 | 12.3 | 9.99 | 8.95 | 6.36 | 5.74 | 4.78 | 4.44 | 3.75 | 3.41 |
| FON | 60.2 | 16.0 | 11.6 | 10.6 | 8.89 | 8.13 | 6.70 | 6.28 | 5.45 | 5.14 | 4.49 | 4.17 |
| S-PNDM | 15.2 | 12.2 | 9.45 | 8.42 | 6.50 | 5.69 | 4.03 | 3.72 | 3.30 | 3.19 | 3.01 | 2.99 |
| F-PNDM | 11.3 | 7.71 | 5.51 | 4.75 | 3.67 | 3.34 | 2.81 | 2.75 | 2.71 | 2.71 | 2.77 | 2.86 |
| steps | 5 | 10 | 20 | 25 | 40 | 50 | 100 | 125 | 200 | 250 |
| DDIM | 19.5 | 12.5 | 10.8 | 10.6 | ||||||
| DDIM* | 48.8 | 18.8 | 11.7 | 11.0 | 10.1 | 10.0 | 9.84 | 9.83 | 9.85 | 9.88 |
| S-PNDM | 20.5 | 11.8 | 9.20 | 9.13 | 9.31 | 9.49 | 9.82 | 9.88 | 10.0 | 10.0 |
| F-PNDM | 14.8 | 8.69 | 9.13 | 9.33 | 9.69 | 9.89 | 10.1 | 9.99 | 10.1 | 10.1 |
| steps | -5 | 10 | 20 | 25 | 40 | 50 | 100 | 125 | 200 | 250 |
| DDIM | 17.0 | 8.89 | 6.75 | 6.62 | ||||||
| DDIM* | 51.3 | 16.4 | 8.47 | 7.41 | 6.27 | 6.05 | 5.97 | 6.03 | 6.23 | 6.32 |
| S-PNDM | 18.1 | 10.2 | 6.50 | 6.02 | 5.74 | 5.81 | 6.29 | 6.44 | 6.69 | 6.75 |
| F-PNDM | 12.6 | 6.99 | 5.68 | 5.74 | 6.17 | 6.44 | 6.91 | 6.96 | 7.03 | 6.92 |
| dataset | FID\ step model | 10 | 20 | 50 | 100 | 250 | 1000 |
| Cifar10 (linear) | DDIM* | 18.50±.06 | 10.86±.08 | 6.95±.04 | 5.49±.06 | 4.52±.02 | 4.02±.04 |
| FON | 13.00±.11 | 7.33±.06 | 5.24±.05 | 4.64±.04 | 4.12±.03 | 3.73±.03 | |
| S-PNDM | 11.58±.10 | 7.53±.07 | 5.15±.05 | 4.34±.03 | 3.93±.02 | 3.83±.03 | |
| F-PNDM | 6.12±.07 | 5.04±.07 | 4.01±.02 | 3.75±.04 | 3.67±.03 | 3.78±.04 |
| Model | CIFAR-10 | ImageNet 32×32 | ImageNet 64×64 | ImageNet 128×128 | ||||||||
| NLL↓ | FID↓ | NFE↓ | NLL↓ | FID↓ | NFE↓ | NLL↓ | FID↓ | NFE↓ | Model | NLL↓ | FID↓ | |
| Ablations | MGAN (Hoang etal.,2018) | 1 | 58.9 | |||||||||
| DDPM | 3.12 | 7.48 | 274 | 3.54 | 6.99 | 262 | 3.32 | 17.36 | 264 | PacGAN2 (Lin et a.,2018) | = | 57.5 |
| Score Matching | 3.16 | 19.94 | 242 | 3.56 | 5.68 | 178 | 3.40 | 19.74 | 441 | Logo-GAN-AE(Sag etal,018) | 一 | 50.9 |
| ScoreFlow | 3.09 | 20.78 | 428 | 3.55 | 14.14 | 195 | 3.36 | 24.95 | 601 | Self-cond. GAN (Lucic e al.,2019) | 一 | 41.7 |
| Ours | Uncond.BigGAN (Lueic etal,19) | 25.3 | ||||||||||
| FM W/ Diffusion | 3.10 | 8.06 | 183 | 3.54 | 6.37 | 193 | 3.33 | 16.88 | 187 | PGMGAN (Armandpour et al.,2021) | 21.7 | |
| FM W/OT | 2.99 | 6.35 | 142 | 3.53 | 5.02 | 122 | 3.31 | 14.45 | 138 | FM W/OT | 2.90 | 20.9 |
| Model | FID↓ | IS↑ | PSNR↑ | SSIM↑ |
| Reference | 1.9 | 240.8 | 1 | 一 |
| Regression | 15.2 | 121.1 | 27.9 | 0.801 |
| SR3 (Saharia et al,22) | 5.2 | 180.1 | 26.4 | 0.762 |
| FM W/OT | 3.4 | 200.8 | 24.7 | 0.747 |
| CIFAR10 | ImageNet-32 | ImageNet-64 | ImageNet-128 | |
| Channels | 256 | 256 | 192 | 256 |
| Depth | 2 | 3 | 3 | 3 |
| Channels multiple | 1,2,2,2 | 1,2,2,2 | 1,2,3,4 | 1,1,2,3,4 |
| Heads | 4 | 4 | 4 | 4 |
| Heads Channels | 64 | 64 | 64 | 64 |
| Attention resolution | 16 | 16,8 | 32,16.8 | 32,16.8 |
| Dropout | 0.0 | 0.0 | 0.0 | 0.0 |
| Effective Batch size | 256 | 1024 | 2048 | 1536 |
| GPUs | 2 | 4 | 16 | 32 |
| Epochs | 1000 | 200 | 250 | 571 |
| Iterations | 391k | 250k | 157k | 500k |
| Learning Rate | 5e-4 | 1e-4 | 1e-4 | 1e-4 |
| Learning Rate Scheduler | Polynomial Decay | Polynomial Decay | Constant | Polynomial Decay |
| Warmup Steps | 45k | 20k | 1 | 20k |
| CIFAR-10 | ImageNet 32×32 | ImageNet 64×64 | |||||||
| Model | K=1 | K=20 | K=50 | K=1 | K=5 | K=15 | K=1 | K=5 | K=10 |
| Ablation | |||||||||
| DDPM | 3.24 | 3.14 | 3.12 | 3.62 | 3.57 | 3.54 | 3.36 | 3.33 | 3.32 |
| Score Matching | 3.28 | 3.18 | 3.16 | 3.65 | 3.59 | 3.57 | 3.43 | 3.41 | 3.40 |
| ScoreFlow | 3.21 | 3.11 | 3.09 | 3.63 | 3.57 | 3.55 | 3.39 | 3.37 | 3.36 |
| Ours | |||||||||
| FM W/ Diffusion | 3.23 | 3.13 | 3.10 | 3.64 | 3.58 | 3.56 | 3.37 | 3.34 | 3.33 |
| FM W/OT | 3.11 | 3.01 | 2.99 | 3.62 | 3.56 | 3.53 | 3.35 | 3.33 | 3.31 |
| Approach | No.of cameras Cost | Supported real-world use cases | |
| D-NeRF[27],Nerfies[24] Neural Volumes[18] | Monocular Multiple (34) | Low | Forward-facing scenes, slow reconstruction in days. |
| Fourier PlenOctrees[39] | Multiple (60) | High 36O° inward-facing scenes,slow reconstruction in days. High 36O° inward-facing scenes,fast reconstruction in 2hrs. | |
| Ours | 360° inward-facing and forward-facing scenes, | ||
| Few (4) | Low | super-fast reconstruction in 1Omins. |
| SYNTHETIC INWARD-FACING | REAL-WORLD INWARD-FACING | |||||||
| PSNR↑ SSIM↑LPIPS↓GPU(GB)↓ | Time↓ | PSNR↑SSIM↑LPIPS↓GPU(GB)↓Time↓ | ||||||
| Neural Volumes [18] | 9.620 0.532 | 0.5520 19.4 | 22.4hrs | 17.29 | 0.6080.3440 | 19.2 | 22.0hrs | |
| D-NeRF[27] | 31.83 0.960 0.0355 | 10.0 | 18.4hrs | 29.15 | 0.946 0.0643 | 12.4 | 22.1hrs | |
| D-NeRF [27]-2 stage | 28.29 0.945 0.0528 | 9.7 | 18.4hrs | 27.21 | 0.936 0.0706 | 13.2 | 22.2hrs | |
| D-NeRF [27]-dynamic | 17.59 0.839 0.2058 | 9.8 | 21.9hrs | 21.74 0.911 0.0906 | 13.5 | 22.3hrs | ||
| Nerfies [24] | 33.09 0.989 | 0.0432 21.8 | 18.7hrs | 29.58 | 0.980 0.0576 | 22.5 | 19.1hrs | |
| Nerfies [24]-2 stage | 32.37 0.991 | 0.0322 22.0 | 15.8hrs | 23.93 | 0.920 0.0878 | 22.0 | 19.7hrs | |
| Nerfies [24]-dynamic | 19.45 0.794 | 0.1674 22.0 | 21.3hrs | 20.70 | 0.910 0.1080 | 22.0 | 19.6hrs | |
| HyperNeRF[25] | 33.73 0.965 | 0.0335 22.5 | 20.5hrs | 28.50 | 0.944 0.0692 | 22.0 | 20.5hrs | |
| HyperNeRF[25]-2 stage | 29.16 0.953 | 0.0555 22.5 | 19.2hrs | 26.53 | 0.935 0.0802 | 22.0 | 19.3hrs | |
| HyperNeRF [25]-dynamic | 18.00 0.786 0.2173 | 22.4 | 20.6hrs | 10.39 | 0.734 0.3990 | 22.0 | 20.5hrs | |
| NSFF[16] | 27.06 0.936 | 0.0800 21.4 | 12.8hrs | 28.44 | 0.939 0.0714 | 22.7 | 15.3hrs | |
| NSFF [16]-dynamic | 18.18 0.858 | 30.1929 15.0 | 15.5hrs | 19.90 | 0.909 0.0944 | 22.7 | 16.2hrs | |
| Ours (base) | 22.44 0.887 | 0.1173 4.4 | 8mins | 24.56 | 0.917 0.0844 | 6.6 | 10mins | |
| Ours w/ c2f | 31.97 0.975 | 50.0185 4.4 | 7mins | 27.83 | 0.956 0.0465 | 6.6 | 10mins | |
| Ours w/ c2f, tv | 32.73 0.963 | 0.0172 4.4 | 7mins | 29.35 | 0.959 0.0434 | 6.6 | 10mins | |
| Ours w/ c2f,tv,cycle | 33.97 0.981 | 0.0142 4.4 | 8mins | 31.56 | 0.971 0.0292 | 6.6 | 11mins | |
| Ours w/c2f,tv,cycle,flow | 34.29 0.982 | 0.0137 4.4 | 8mins | 31.68 | 0.972 0.0289 | 6.6 | 11mins | |
| PLANT | RABBIT | |||||
| PSNR↑ SSIM↑LPIPS↓GPU(GB)↓ | Time↓ | PSNR↑ SSIM↑LPIPS↓GPU(GB)↓ | Time↓ | |||
| D-NeRF[27] | 31.94 0.979 0.0251 | 11.4 | 21.5hrs | 33.51( 0.974 0.0384 | 11.4 | 21.7hrs |
| Nerfies [24] | 31.36 0.991 0.0309 | 21.9 | 18.3hrs | 24.83 0.952 0.0795 | 21.9 | 19.8hrs |
| HyperNeRF [25] | 32.08 0.978 0.0331 | 22.0 | 20.0hrs | 24.97 0.942 0.0849 | 22.0 | 20.7hrs |
| NSFF[16] | 29.45 0.966 0.0526 | 20.2 | 14.5hrs | 27.68 0.945 0.0854 | 20.2 | 14.5hrs |
| Ours (base) | 26.13 0.946 0.0722 | 8.1 | 8mins | 25.58 0.910 0.1300 | 8.9 | 6mins |
| Ours w/ c2f | 31.85 0.980 0.0275 | 8.1 | 8mins | 26.79 0.938 0.0946 | 8.9 | 6mins |
| Ours w/ c2f, tv | 31.89 0.980 0.0263 | 8.1 | 8mins | 29.28 0.951 0.0655 | 8.9 | 6mins |
| Ours w/ c2f,tv, cycle | 31.99 0.981 0.0235 | 8.1 | 9mins | 31.05 0.963 30.0543 | 8.9 | 7mins |
| Ours w/c2f,tv,cycle,flow | 32.01 0.981 0.0236 | 8.1 | 10mins | 32.05 0.966 0.0492 | 8.9 | 7mins |
| First task peformed | |||||
| Tasks ≥4↑ | Microwave | Kettle1 | Bottom Burner | Other/Failure | |
| Guidance w = 0.0 | 0.63 | 62.3 | 29.7 | 7.3 | 0.7 |
| Guidance w = 1.0 | 0.61 | 57.3 | 30.0 | 12.7 | 0.0 |
| Guidance w = 4.0 | 0.45 | 52.7 | 28.3 | 17.0 | 2.0 |
| Guidance w = 8.0 | 0.08 | 24.0 | 26.7 | 24.7 | 25.6 |
| Human Demonstrations | 0.63 | 58.3 | 31.6 | 10.1 | 0.0 |
| Wasserstein Distance,Human to Model↓ | |||
| Game Score ↑1×timesteps16×timesteps (1 sec)32×timesteps(2 sec) | |||
| Observation encoder: 6-layer CNN | |||
| MSE,MLP Sieve | 6.9 | 19.0 37.6 | 54.7 |
| Discrete,MLP Sieve | 7.8 | 12.9 35.0 | 55.9 |
| *Diffusion BC,MLP Sieve | 12.9 | 14.3 | 34.5 52.8 |
| *Diffusion-X,MLP Sieve | 17.3 | 11.7 30.0 | 47.8 |
| Observation encoder:ResNet18 MSE,MLP Sieve | 17.8 | 5.5 28.1 | 48.9 |
| 14.7 | 6.6 31.3 | 53.0 | |
| Discrete,MLP Sieve | 16.8 | 3.8 29.2 | 51.8 |
| K-Means+Residual,MLP Sieve | 4.3 | 17.2 50.0 | 74.1 |
| EBMDerivative-Free,MLP Sieve *Diffusion BC,MLP Sieve | 19.0 | 6.3 29.5 | 50.4 |
| *Diffusion-X,MLP Sieve | 24.0 | 4.5 24.5 | 44.4 |
| Baselines Human | 36.5 | 0.73 0.57 | 0.38 |
| No. runs | Tasks completed个 | Tasks ≥1↑ | Tasks≥2个 | Tasks≥3个 | Tasks≥4个 | Tasks≥5个 | |
| *Diffusion Basic MLP | 3 | 3.04 ± 0.01 | 0.98 ±0.01 | 0.89 ± 0.01 | 0.71 ± 0.0 | 0.45 ± 0.03 | 0.01±0.0 |
| *Diffusion Basic MLP KDE | 3 | 3.44 ± 0.01 | 0.99±0.0 | 0.97 ±0.01 | 0.86±0.0 | 0.59 ± 0.01 | 0.04 ±0.01 |
| *Diffusion Basic MLP Extra8 | 3 | 3.37 ± 0.03 | 0.99 ± 0.01 | 0.93 ± 0.01 | 0.83 ± 0.02 | 0.58±0.02 | 0.04 ± 0.01 |
| MSEMLP Sieve | 3 | 3.13 ± 0.07 | 0.99 ± 0.0 | 0.9 ± 0.03 | 0.72 ± 0.03 | 0.5 ±0.02 | 0.02 ± 0.0 |
| Discretised MLP Sieve | 3 | 2.26±0.08 | 0.95 ± 0.02 | 0.73 ±0.02 | 0.4 ±0.03 | 0.18 ±0.02 | 0.0±0.0 |
| Disc.K-MeansMLP Sieve | 3 | 0.33±0.03 | 0.3±0.03 | 0.03±0.0 | 0.0±0.0 | 0.0±0.0 | 0.0±0.0 |
| K-Means+Residual MLP Sieve | 3 | 2.56 ± 0.04 | 0.95±0.0 | 0.82 ± 0.01 | 0.55 ± 0.02 | 0.23 ±0.02 | 0.01 ±0.0 |
| EBMDeriv-Free,MLP Sieve | 1 | 0.39 | 0.31 | 0.08 | 0.0 | 0.0 | 0.0 |
| *Diffusion MLP Sieve | 3 | 3.46 ± 0.04 | 0.99 ±0.0 | 0.94±0.0 | 0.82 ± 0.02 | 0.68 ±0.02 | 0.02 ±0.01 |
| *Diffusion KDE MLP Sieve | 3 | 3.78 ± 0.06 | 1.0 ± 0.0 | 1.0 ±0.0 | 0.92 ± 0.01 | 0.79 ± 0.04 | 0.07 ±0.01 |
| *Diffusion Extra MLP Sieve | 3 | 3.72 ± 0.02 | 1.0 ± 0.0 | 0.97 ±0.0 | 0.9 ± 0.01 | 0.77 ± 0.02 | 0.08±0.0 |
| MSE Transformer | 3 | 3.59 ± 0.02 | 0.99±0.0 | 0.96 ±0.01 | 0.88 ± 0.01 | 0.69 ± 0.02 | 0.07 ±0.01 |
| Discretised Transformer | 3 | 2.71 ± 0.08 | 0.97 ±0.0 | 0.82 ±0.02 | 0.57 ± 0.04 | 0.34 ± 0.02 | 0.02 ± 0.01 |
| Disc.K-Means Transformer | 1 | 0.47 | 0.42 | 0.05 | 0.0 | 0.0 | 0.0 |
| *Diffusion Transformer | 3 | 3.74 ± 0.02 | 1.0 ± 0.0 | 0.98 ± 0.01 | 0.9 ± 0.02 | 0.77 ± 0.01 | 0.1 ± 0.01 |
| *Diffusion KDE Transformer | 3 | 3.93 ± 0.02 | 1.0 ± 0.0 | 1.0 ±0.0 | 0.95 ± 0.0 | 0.89 ± 0.01 | 0.1 ±0.01 |
| *Diffusion Extra Transformer | 3 | 3.91 ± 0.04 | 1.0 ± 0.0 | 0.99 ±0.01 | 0.94± 0.0 | 0.88 ± 0.01 | 0.1±0.03 |
| Human | 3.98 | 1.0 | 1.0 | 1.0 | 0.98 | 0.0 | |
| Previously reported in (Shafiullah et al.,2022): | |||||||
| Behaviour Transformers (K-Means+Residual) | 1 | 3.09 | 0.99 | 0.93 | 0.71 | 0.44 | 0.02 |
| Implicit BC (Generative EBM) | 1 | 2.71 | 0.99 | 0.87 | 0.61 | 0.24 | 0.0 |
| SPiRL VAE (Pertsch etal., 2021) | 1 | 1.0 | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| PARROT Normalizing Flow (Singh et al., 2020) | 1 | 0.04 | 0.04 | 0.0 | 0.0 | 0.0 | 0.0 |
| Task1↓ | Task 2↓ | Task 3↓ | Task 4↓ | |
| *Diffusion-Extra MLP Basic | 2.36 ±0.50 | 3.65 ±0.26 | 14.40 ± 1.05 | 14.04 ±0.38 |
| *Diffusion MLP Basic | 3.48 ± 0.94 | 8.63 ± 2.66 | 23.13 ± 4.86 | 12.93 ± 1.14 |
| *Diffusion-KDE MLPBasic | 3.52 ± 0.45 | 5.82 ±0.40 | 14.11 ± 0.76 | 8.86±0.74 |
| *Diffusion MLP Sieve | 2.69 ± 0.97 | 3.84 ± 0.51 | 8.54 ± 2.81 | 9.18 ± 1.12 |
| *Diffusion-KDE MLP Sieve | 3.50 ±0.68 | 7.07 ± 1.18 | 8.51 ±2.08 | 7.98 ± 0.82 |
| *Diffusion-Extra MLP Sieve | 1.69 ± 0.26 | 4.60 ± 1.03 | 8.34 ±2.36 | 6.31 ±0.80 |
| MSE MLP Sieve | 3.03 ± 0.71 | 5.70 ±1.89 | 9.63 ± 0.94 | 7.24 ± 0.91 |
| Discrete MLP Sieve | 3.74 ± 1.31 | 11.09 ± 1.92 | 20.98 ± 1.61 | 9.38 ±2.12 |
| K-Means MLP Sieve | 34.61 ± 7.21 | 71.20 ± 20.31 | = | |
| K-Means+ResidualMLPSieve | 3.57 ± 0.41 | 10.07 ± 2.96 | 17.26 ± 4.07 | 15.51 ± 1.36 |
| *Diffusion-Extra Transformer | 2.04 ± 0.64 | 5.12 ± 1.02 | 5.51± 0.56 | |
| *Diffusion Transformer | 1.26 ± 0.01 | 3.33 ± 0.36 | 7.89 ± 0.57 | 5.92 ± 0.31 3.97 ± 0.51 |
| *Diffusion-KDE Transformer | 4.21 ± 0.42 | 6.34 ± 1.23 | 5.22 ±0.32 | 5.33 ± 0.40 |
| MSE Transformer | 3.03 ±0.74 | 5.73 ± 0.61 | 7.99 ± 1.39 | 6.64 ± 1.12 |
| Discrete Transformer | 2.35 ± 0.31 | 5.95 ± 0.57 | 9.61 ± 1.82 | 6.61 ± 0.70 |
| K-Means Transformer | 55.79 | 32.64 | ||
| K-Means+Residual Transformer | 2.16 ± 0.78 | 一 | ||
| 4.55 ± 1.27 | 10.07 ± 0.23 | 14.41 ± 1.83 |
| Trainingtimeperepoch↓Samplingtime↓Maxsamplingrate↑ | |||
| Kitchen environment, CPU | |||
| MSE,MLP Sieve | 7.4 seconds | 1.2 ms | 833Hz |
| Discrete,MLP Sieve | 7.6 seconds | 2.7 ms | 370 Hz |
| K-Means,MLPSieve | 8.7 seconds | 1.5 ms | 666Hz |
| K-Means+Residual,MLP Sieve | 10.5 seconds | 1.5 ms | 666Hz |
| *Diffusion BC (T= 50),MLP Sieve | 6.4 seconds | 44.6 ms | 22 Hz |
| *Diffusion-X(T= 50,M=8),MLP Sieve | 6.4 seconds | 55.9 ms | 18 Hz |
| *Diffusion-KDE(T= 50,100 samples),MLP Sieve | 6.4 seconds | 128.4 ms | 8Hz |
| Kitchen environment, V1oo GPU | |||
| MSE,Basic MLP | 1.4 seconds | 1.1 ms | 909 Hz |
| Discrete,Basic MLP | 1.8 seconds | 3.9 ms | 256Hz |
| K-Means,Basic MLP | 2.2 seconds | 1.4 ms | 714 Hz |
| K-Means+Residual,Basic MP | 2.4 seconds | 1.4 ms | 714 Hz |
| *Diffusion BC (T= 50),Basic MLP | 1.4 seconds | 42.1 ms | 24 Hz |
| *Diffusion-X(T= 50,M=8),Basic MLP | 1.4 seconds | 47.3 ms | 21 Hz |
| *Diffusion-KDE(T= 50,100 samples),Basic MLP | 1.4 seconds | 61.2 ms | 16 Hz |
| MSE,MLP Sieve | 1.5 seconds | 1.5 ms | 666 Hz |
| Discrete,MLP Sieve | 2.2 seconds | 4.3 ms | 232 Hz |
| K-Means,MLPSieve | 2.9 seconds | 1.8 ms | 555Hz |
| K-Means+Residual,MLPieve | 3.1 seconds | 2.0 ms | 500 Hz |
| EBMDeriv-Free,MLP Sieve | 22.3 seconds | 28.2 ms | 35Hz |
| EBMLangevin,MLP Sieve | 142.1 seconds | 525.5 ms | 2Hz |
| *Diffusion BC(T= 5O),MLP Sieve | 2.0 seconds | 63.1 ms | 16 Hz |
| *Diffusion-X(T= 50,M=8),MLP Sieve | 2.0 seconds | 73.8 ms | 14 Hz |
| *Diffusion-KDE(T= 50,100 samples),MLP Sieve | 2.0 seconds | 86.5 ms | 12 Hz |
| MSE,Transformer | |||
| Discrete,Transformer | 6.8 seconds | 5.5 ms | 181Hz |
| 7.0 seconds | 8.6 ms | 116 Hz | |
| K-Means,Transformer | 7.4 seconds | 5.9 ms | 169 Hz |
| K-Means+Residual,Transformer | 7.5 seconds | 6.0 ms | 167 Hz |
| EBMDeriv-Free,Transformer | 1000<seconds | 1390.8 ms | 0.7Hz |
| EBMLangevin,Transformer | 1000<seconds | 2927.9 ms | 0.3Hz |
| *Diffusion BC(T= 5O),Transformer | 6.8 seconds | 244.2 ms | 4Hz |
| *Diffusion-X(T= 50,M= 8),Transformer | 6.8 seconds | 285.3 ms | 4 Hz |
| *Diffusion-KDE(T= 50,100 samples),Transformer | 6.8 seconds | 295.6 ms | 3Hz |
| CSGO environment,ResNet18 observation encoder, V1oo GPU | |||
| MSE,MLP Sieve | 49 seconds | 5.0 ms | 200 Hz |
| Discrete,MLP Sieve | 49 seconds | 6.0 ms | 167Hz |
| EBMDeriv-Free,MLP Sieve | 50 seconds | 9.3 ms | 107 Hz |
| EBMLangevin,MLP Sieve | 240 seconds | 537.2 ms | 2Hz |
| K-Means,MLP Sieve | 51 seconds | 5.5 ms | 181Hz |
| K-Means+Residual,MLPSieve | 51 seconds | 5.4 ms | 185 Hz |
| *Diffusion BC(T= 2O),MLP Sieve | 49 seconds | 31.5 ms | 32 Hz |
| *Diffusion-X(T= 20,M=16),MLP Sieve | 49 seconds | 54.7 ms | 18Hz |
| Index | mem | Text |
| 2090855 | 0.6546 | link > THE AMERICAN JEWISH CONGRESS ANNOUNCED TODAY THE PUBLICATION OF A REPORT ON JEWISH NON- EMPLOYMENT AS A RESULT OF ECONOMIC DISCRIMINATION,[]THEREAFTER ONE OF THE DEPARTMENTS OF A.T.&T. ”ALMOSTUNPRECEDENTEDLY "ENGAGED A JEWISH APPLICANT. link×RECIPE:Chinese Pork & Vegetable Soup with Wonton Noodles ChinesePork& Vegetable Soup with Wonton Noodleslpork |
| 2085736 | 0.5755 | tenderloin(about1-1poudsize)ookedandcutito-nchcubes5cupslower-sodumchickenbothcupwater**cupke serving (about1/cupseach)RecipebyPorkBeInspiredcomwithadaptationsbyculinarydititian&nutritionistKimGaleaz,DNCD |
| 1680600 | 0.5807 | linkLangaEgish..Aab.ackoedgeentofountry.acetouldlikekoldgatis being heldonthe traditional landsof the(appropriate group)people,and paymyrespecttoelders both pastand present."] linkATexashonorsstudentpunishedforsayingthathomosexualitywas wronghashadhissuspensionrescinded.WesteHlsHigh |
| 2074805 | 0.2835 | madethecorectdeisioninreversingtheircourseofction."hedecisiontorescindthesuspensionisthecorrctoneThesuspesionwas Wrongandimproper”saidStaver.aplaudthestudentforstandingup.Westoodwithhimtoresistanunjustsuspensionandwearepleased thatsuspensionsbeenreversed”LibertyCunselwillontiuetheighttoexecisfreedmofonscienceandeligion”saidtave "ThesestacecreasidillotiutceaseuesCrstsdpeoplevelirtyadudstisila |
| 449808 | 0.0361 | linkInvestors in Digital Realty Trust, Inc. (DLR) sawnewoptionsbegintradingthisweek, for the February2014 expiration.At Stock Options Cannel,ourYeldBostforulahaslookedupanddowtheDLRoptioschainforthenewFebruary014contractsandidetfed one putandonecallcontractof particular interest.The put contractatthe $45.OO strike price has acurrent bidof $1.00 [] |
| 1157311 | 0.0356 | link> Investors in Abercrombie &Fitch Co. (ANF) sawnewoptionsb becomeavailable today,forthe April4th expiration.At Stock Options Channel, our YieldBoost formula has looked up and down the ANF options chain for the new April 4th contracts and identified one put and one call contract of particular interest.The put contract at the $34.OO strike price has a current bid of $1.97. [.] |
| Index Estim. Text | |||||||||
| Validation infl link > Identical URL except with the http:// protocol instead of https://. The text is identical to the training example. | |||||||||
| 1334662 | |||||||||
| Train | |||||||||
| 0.3780mem linkByAribiovichZEELLLYsrael(Reuters)-sposableperacemassoeritleproteconfroodsof0.3780 dustatfillthlfeereelceolistseapgsightoopelelvdtoarsdiiby Jeffrey Heller/Jeremy Gaunt) | |||||||||
| 2077314 | |||||||||
| Validation infl link →VATICAN CITY(AP)Emeritus Pope Benedict XVI is ofering a first-ever papal assessment of his own pontificate in a book that838341 0.1209 recountshisecisiotoigisuseathisceordisptstmantlathcallteVatian’s"gyb"dctXVI:TheFinaloesatioteptebstViarelyspreainggosipatawebsites, but (almost) identical report. | |||||||||
| websites, but (almost) identical report.mem link >VATICAN CITY EmeritusPopeBenedictXVIisofferingafirst-everpapalassessmentofhisownpontificate inabookthatrecounts0.1650 hisdecisiontosipiseatsscodiseptstosmatleathecallthVatian'sgaloby"ecXFinalConveseoutiebtestaalysprdigosiathaylloleifat www.twitter.com/nwinfield | |||||||||
| Train | |||||||||
| 614881 | |||||||||
| Validation infl linkANAHOigthnFracoseauchinadtosissirotdCrisrogerbockedixsotsino69682107 0.0673 Ducks fans might have been lost without their programs on Wednesday night.it.It's only the first game of the preseason,but anew era has | |||||||||
| 682107 | 0.0673 | ||||||||
| clearlybegun.AgroupofmostlyewcomersinDucksuiformsbeatPhoenix3-2inashootoutonWednesdayatHondaCenterfamliarface made the biggest impact,however,a,as Bobby Ryan scored two goals.Ryan also scored two goals in the Ducks'preseason opener last year[...] Different websites on the same event with slightly different wordings.linkANAHEM-Onanight whenFrancois Beaucheminhad twoasistsinTorontoandChris Prongerblocked sixshots inDetroit,the13,869Ducks'fanswhoshowed atat Honda CenterWednesdayneeded programsttoidentifytheplayerson theirfavorite team. It's only the | |||||||||
| twogoals in the | |||||||||
| 494435 | 0.1439 | toidentifytheplayerson theirfavorite team. It's only the | |||||||
| firstgameofthepreseason,utaeerasearlygnAgropofostlyewcorsicksuforsbatPoenix3-2ioutalthoughafamiliarface made thebiggest impactasBobby Ryanscoredtwo goalsinregulationandanotherinthe shootout.Ryanalso scoredtwo goals in the Ducks’ preseason opener last year[...] | |||||||||
| Validation | infl | link× | link×Morethan 70.ooo poundsof Butterball turkeyrecalled because of potential salmonella WASHINGTON- The U.S. Departmentof Agriculture'sodfetyndIspectiossviceaoncednWedneayeyeesippdtoatioideetailadsitiollocationss. RELATED: View the fullrecall FSIS, the Centers for Disease Control and Prevention and [..] Different websitesreporting the same event, one embeded a lot more information than the other.link>×EButterball recallsnearly80Oo0 pounds of turkeyaftersalmonellacases WASHNGTON—The U.S.Departmentof Agriculture'sFood Safety and InspectionserviceannouncedWednesday [.] The raw ground turkey was produced on July 7, 2018. The following products | ||||||
| 1165799 | 0.0360 | ||||||||
| same event, one embeded a lot more information than the other.link>×EButterball recallsnearly80Oo0 pounds of turkeyaftersalmonellacases WASHNGTON—The U.S.Departmentof Agriculture'sFood Safety and InspectionserviceannouncedWednesday [.] The raw ground turkey was produced on July 7, 2018. The following productsunder recallwereshipped tonationwide retailand institutional locations: 48-oz. plastic wrapped traycontaining“BUTTERBALL everydayFresh Ground Turkey WITHNATURALFLAVORING (85%LEAN/15%FAT) with sellorfreeze bydateof 726/18,lot code 8188,andUPCcodes 2655-71555or65-71557representedonthelabel.48-oz.plastic wrappedtraycontaining"BUERBALLeverydayFreshGround Turkey WITH NATURAL FLAVORING (93% LEAN/7% FAT)" with sell or freeze by date of 7/26/18, lot[[..]labels here.FSIS,the | es WASHIN | ||||||||
| 1571976 | 0.2094 | ||||||||
| Centers for Disease Control and Prevention and | |||||||||
| Index Estim. Text | ||||
| Generation infl link > Baku, Azerbaijan,April15 Trend:Offcial exchange rateof the US dollar and euro against Azerbaijani manat was setat1.7and1.92251361 0.1805 manats, respectively, for April115.Below are the rates of Azerbaijani manat against world currencies,according to the data from the CentralBank of AzerbaijanforApril15.[0OJapaneseyen10JY1.51871NewZealanddollr1NZD11513FollowTrendonTelegram.Only most interesting and important newsTrain mem linkBaku,Azerbaijanch5Tend:OfcialehangateoftheUSdolarandroainstAzerbajanimaatwassetat7nd2072973 0.3534 1.9241 manats,respectivelyforarch5.Belowretheatesofzerbijanimanataganstorldurrencies,accordingtotetafrotheCentralBankofAzerbaijanforMarch15.0Japaneseyen1JY1.52201NewZealanddolar1NZD1.1636FollowendonTelegram. Only most interesting and important news | ||||
| 1361 | ||||
| Generation infl link >NEW DELHI: India is likely to see average monsoon rains this year., the state-run weather office said on Monday, which should support | ||||
| 21998 0.0218 agriculturalproucioocowthisia’sdbgesoyerealfofacksgatiool | ||||
| isexpectedtobe96 percentofthelong-termaverage,M.Rajeevan,secretaryatthe Ministryof Earth Sciences,,told a news conference.TheIndiaMeteorolicaepatsgealfallae9tdpetofa89 centimeters for the entire four-month season beginning June.[]India's weather office will update its forecast in the first week of June.However,onagetheasforestacuatelyoyoceeeryfiveyearsovrtepasttead,eaftertingtaerror band of plus or minus 5 percentage points. | ||||
| Train | mem | |||
| linkNEW(euters)dislikeltceieveraemosoiin8teweatrocesidasingteohigherfaroocothisia’sdgeoyeealfofadcksatioooislloeexpectedtobe97ppercent ofalong-termaverage,K.J. Ramesh, director general of the state-run India | ||||
| 326212 | 0.1555 | of the country's $2 trillion economy, are | ||
| reexpecter | ||||
| Meteorological Department (IMD), told a news conference."We see very less probability of a deficit monsoon," Ramesh said on Monday.Otherthanliftingfarmandwideronomicgowth,aspellofgoodrainswillkeepalidonfation,potentiallytmptingPrieister | ||||
| NarendraModitobingfrwardgneralelectiosdueia19.Idia’sweathofcefiesaverag,noal,aifallastwee96percentand104 percentofa50-yearaverageof89cmsforthe entire four-month season beginning June.[ saidaMumbai-based dealer | ||||
| witha globaltrading firm.Average monsoonrainfallwillhelp India retain its position as the world’s toprice exporter | ||||
| Model | Param (M) | FLOPs (G) | Throughput (imgs/s) | Train Mem (GB) | Test Mem (GB) | Top-1 (%) |
| ResNet-50 [53] | 26 | 4.1 | 1,279 | 7.9 | 2.8 | 80.4 |
| ConvNext-Ti [33] | 28 | 4.5 | 1,079 | 8.3 | 1.7 | 82.1 |
| PVT-S [51] | 25 | 3.8 | 1,007 | 6.8 | 1.3 | 79.8 |
| Swin-Ti [32] | 28 | 4.5 | 961 | 6.1 | 1.5 | 81.3 |
| CvT-13 [54] | 20 | 4.5 | 947 | 6.1 | 1.5 | 81.6 |
| Focal-Tiny [60] | 29 | 4.9 | 384 | 12.2 | 3.3 | 82.2 |
| Twins-PCPVT-S [12] | 24 | 3.8 | 998 | 6.8 | 1.2 | 81.2 |
| LITv1-S [36] | 27 | 4.1 | 1,298 | 5.8 | 1.2 | 81.5 |
| LITv2-S | 28 | 3.7 | 1,471 | 5.1 | 1.2 | 82.0 |
| ResNet-101[53] | 45 | 7.9 | 722 | 10.5 | 3.0 | 81.5 |
| ConvNext-S [33] | 50 | 8.7 | 639 | 12.3 | 1.8 | 83.1 |
| PVT-M [51] | 44 | 6.7 | 680 | 9.3 | 1.5 | 81.2 |
| Twins-SVT-B[12] | 56 | 8.3 | 621 | 9.8 | 1.9 | 83.2 |
| Swin-S [32] | 50 | 8.7 | 582 | 9.7 | 1.7 | 83.0 |
| LITv1-M [36] | 48 | 8.6 | 638 | 12.0 | 1.4 | 83.0 |
| LITv2-M | 49 | 7.5 | 812 | 8.8 | 1.4 | 83.3 |
| ResNet-152 [53] | 60 | 11.6 | 512 | 13.4 | 2.9 | 82.0 |
| ConvNext-B [33] | 89 | 15.4 | 469 | 16.9 | 2.9 | 83.8 |
| Twins-SVT-L [12] | 99 | 14.8 | 440 | 13.7 | 3.1 | 83.7 |
| Swin-B [32] | 88 | 15.4 | 386 | 13.4 | 2.4 | 83.3 |
| LITv1-B [36] | 86 | 15.0 | 444 | 16.4 | 2.1 | 83.4 |
| LITv2-B | 87 | 13.2 | 602 | 12.2 | 2.1 | 83.6 |
| DeiT-B↑ 384 [45] | 86 | 55.4 | 159 | 39.9 | 2.5 | 83.1 |
| Swin-B↑ 384 [32] | 88 | 47.1 | 142 | OOM | 6.1 | 84.5 |
| LITv2-B↑ 384 | 87 | 39.7 | 198 | 35.8 | 4.6 | 84.7 |
| Backbone | RetinaNet | MaskR-CNN | |||||||
| Params | FLOPs (G) | FPS | AP6 | Params | FLOPs (G) | FPS | AP6 | APm | |
| ResNet-50 [23] | 38M | 239 | 18.5 | 36.3 | 44M | 260 | 27.1 | 38.0 | 34.4 |
| PVT-S [51] | 34M | 273 | 13.0 | 40.4 | 44M | 292 | 16.2 | 40.4 | 37.8 |
| Swin-T [32] | 38M | 251 | 17.0 | 41.5 | 48M | 270 | 21.1 | 42.2 | 39.1 |
| Twins-SVT-S[12] | 34M | 225 | 15.5 | 43.0 | 44M | 244 | 20.4 | 43.4 | 40.3 |
| LITv1-S [36] | 39M | 305 | 3.3 | 41.6 | 48M | 324 | 3.2 | 42.9 | 39.6 |
| LITv2-S | 38M | 242 | 18.7 | 44.0 | 47M | 261 | 18.7 | 44.9 | 40.8 |
| LITv2-S* | 38M | 230 | 20.4 | 43.7 | 47M | 249 | 21.9 | 44.7 | 40.7 |
| ResNet-101[23] | 57M | 315 | 15.2 | 38.5 | 63M | 336 | 20.9 | 40.4 | 36.4 |
| PVT-M[51] | 54M | 348 | 10.5 | 41.9 | 64M | 367 | 10.8 | 42.0 | 39.0 |
| Swin-S [32] | 60M | 343 | 13.3 | 44.5 | 69M | 362 | 15.8 | 44.8 | 40.9 |
| Twins-SVT-B[12] | 67M | 358 | 10.8 | 45.3 | 76M | 377 | 12.7 | 45.2 | 41.5 |
| LITv2-M | 59M | 348 | 12.2 | 46.0 | 68M | 367 | 12.6 | 46.8 | 42.3 |
| LITv2-M* | 59M | 312 | 14.8 | 45.8 | 68M | 315 | 16.0 | 46.5 | 42.0 |
| ResNeXt101-64x4d [58] | 96M | 473 | 10.3 | 41.0 | 102M | 493 | 12.4 | 42.8 | 38.4 |
| PVT-L [51] | 71M | 439 | 9.5 | 42.6 | 81M | 457 | 8.3 | 42.9 | 39.5 |
| Swin-B [32] | 98M | 488 | 11.0 | 44.7 | 107M | 507 | 11.3 | 45.5 | 41.3 |
| Twins-SVT-L [12] | 111M | 504 | 9.9 | 45.7 | 120M | 524 | 10.1 | 45.9 | 41.6 |
| LITv2-B | 97M | 481 | 9.5 | 46.7 | 106M | 500 | 9.3 | 47.3 | 42.6 |
| LITv2-B* | 97M | 430 | 11.8 | 46.3 | 106M | 449 | 11.5 | 46.8 | 42.3 |
| Backbone | Params (M) | FLOPs (G) | FPS | mIoU (%) |
| ResNet-50[23] PVT-S [51] Swin-Ti[32] Twins-SVT-S [12] LITv1-S [36] | 29 | 45 | 45.4 38.7 | 36.7 39.8 |
| 28 32 | 40 46 | 39.6 | 41.5 | |
| 28 32 | 37 46 | 34.5 | 43.2 | |
| LITv2-S ResNet-101[23] PVT-M [51] Swin-S [32] | 31 | 41 | 18.1 42.6 | 41.7 44.3 |
| 48 48 | 66 55 | 36.7 29.7 | 38.8 41.6 | |
| 53 | 70 | |||
| Twins-SVT-B[12] LITv2-M | 24.4 | 45.2 | ||
| 60 | 67 | 28.0 | 45.3 | |
| 52 | 63 | 28.5 | 45.7 | |
| PVT-L [51] Swin-B[32] Twins-SVT-L [12] | 65 | 71 | 20.5 | 42.1 |
| 107 | 107 | 25.5 | 46.0 | |
| 104 | 102 | 25.9 | 46.7 | |
| LITv2-B | 90 | 93 | 27.5 | 47.2 |
| Name | ImageNet-1K | COCO (RetinaNet) | ||
| FLOPs Mem (G) (GB) | Top-1 (%) | FLOPs FPS (G) | AP | |
| LITv1-S [36] | 4.1 5.8 | 81.5 | 305 3.3 | 41.6 |
| + ConvFFN | 4.1 6.5 | 82.5 | 306 3.1 | 45.1 |
| + Remove RPE | 4.1 6.5 | 82.3 | 306 13.3 | 44.7 |
| + HiLo | 3.7 5.1 | 82.0 | 224 18.7 | 44.0 |
| Model | Params (M) | FLOPs (G) | A100 | V100 | RTX 6000 | RTX 3090 | Top-1 (%) |
| ResNet-50 [53] | 26 | 4.1 | 1,424 | 1,123 | 877 | 1,279 | 80.4 |
| PVT-S [51] | 25 | 3.8 | 1,460 | 798 | 548 | 1,007 | 79.8 |
| Twins-PCPVT-S[12] | 24 | 3.8 | 1,455 | 792 | 529 | 998 | 81.2 |
| Swin-Ti [32] | 28 | 4.5 | 1,564 | 1,039 | 710 | 961 | 81.3 |
| TNT-S [21] | 24 | 5.2 | 802 | 431 | 298 | 534 | 81.3 |
| CvT-13 [54] | 20 | 4.5 | 1,595 | 716 | 379 | 947 | 81.6 |
| CoAtNet-0 [15] | 25 | 4.2 | 1,538 | 962 | 643 | 1,151 | 81.6 |
| CaiT-XS24 [46] | 27 | 5.4 | 991 | 484 | 299 | 623 | 81.8 |
| PVTv2-B2 [52] | 25 | 4.0 | 1,175 | 670 | 451 | 854 | 82.0 |
| XCiT-S12[1] | 26 | 4.8 | 1,727 | 761 | 504 | 1,068 | 82.0 |
| ConvNext-Ti [33] | 28 | 4.5 | 1,654 | 762 | 571 | 1,079 | 82.1 |
| Focal-Tiny [60] | 29 | 4.9 | 471 | 372 | 261 | 384 | 82.2 |
| LITv2-S | 28 | 3.7 | 1,874 | 1,304 | 928 | 1,471 | 82.0 |
| COV-R (%) ↑ | MAT-R (A)↓ | COV-P (%) ↑ | MAT-P (A)↓ | |||||
| Models | Mean | Median | Mean | Median | Mean | Median | Mean | Median |
| CVGAE | 0.00 | 0.00 | 3.0702 | 2.9937 | - | = | = | - |
| GRAPHDG | 8.27 | 0.00 | 1.9722 | 1.9845 | 2.08 | 0.00 | 2.4340 | 2.4100 |
| CGCF | 53.96 | 57.06 | 1.2487 | 1.2247 | 21.68 | 13.72 | 1.8571 | 1.8066 |
| CONFVAE | 55.20 | 59.43 | 1.2380 | 1.1417 | 22.96 | 14.05 | 1.8287 | 1.8159 |
| GEOMOL | 67.16 | 71.71 | 1.0875 | 1.0586 | 1 | = | = | |
| CONFGF | 62.15 | 70.93 | 1.1629 | 1.1596 | 23.42 | 15.52 | 1.7219 | 1.6863 |
| GEODIFF-A | 88.36 | 96.09 | 0.8704 | 0.8628 | 60.14 | 61.25 | 1.1864 | 1.1391 |
| GEODIFF-C | 89.13 | 97.88 | 0.8629 | 0.8529 | 61.47 | 64.55 | 1.1712 | 1.1232 |
| Models | COV-R (%) ↑ | MAT-R (A) | COV-P (%) ↑ | MAT-P(A)↓ | ||||
| Mean | Median | Mean | Median | Mean | Median | Mean | Median | |
| RDKIT | 60.91 | 65.70 | 1.2026 | 1.1252 | 72.22 | 88.72 | 1.0976 | 0.9539 |
| GEODIFF + FF | 92.27 | 100.00 | 0.7618 | 0.7340 | 84.51 | 95.86 | 0.9834 | 0.9221 |
| Method | E | Emin | A | △emin | △emax |
| RDKIT | 0.9233 | 0.6585 | 0.3698 | 0.8021 | 0.2359 |
| GRAPHDG | 9.1027 | 0.8882 | 1.7973 | 4.1743 | 0.4776 |
| CGCF | 28.9661 | 2.8410 | 2.8356 | 10.6361 | 0.5954 |
| CONFVAE | 8.2080 | 0.6100 | 1.6080 | 3.9111 | 0.2429 |
| CONFGF | 2.7886 | 0.1765 | 0.4688 | 2.1843 | 0.1433 |
| GEODIFF | 0.25974 | 0.1551 | 0.3091 | 0.7033 | 0.1909 |
| Task | β1 | T | β scheduler | T | T | Batch Size | Train Iter. |
| QM9 | 1e-7 | 2e-3 | sigmoid | 5000 | 10A | 64 | 1M |
| Drugs | 1e-7 | 2e-3 | sigmoid | 5000 | 10A | 32 | 1M |
| COV-R (%) ↑ | MAT-R(A)↓ | COV-P (%) ↑ | MAT-P (A) ↓ | |||||
| Models | Mean | Median | Mean | Median | Mean | Median | Mean | Median |
| CVGAE | 0.09 | 0.00 | 1.6713 | 1.6088 | = | = | = | |
| GRAPHDG | 73.33 | 84.21 | 0.4245 | 0.3973 | 43.90 | 35.33 | 0.5809 | 0.5823 |
| CGCF | 78.05 | 82.48 | 0.4219 | 0.3900 | 36.49 | 33.57 | 0.6615 | 0.6427 |
| CONFVAE | 77.84 | 88.20 | 0.4154 | 0.3739 | 38.02 | 34.67 | 0.6215 | 0.6091 |
| GEOMOL | 71.26 | 72.00 | 0.3731 | 0.3731 | = | = | ||
| CONFGF | 88.49 | 94.31 | 0.2673 | 0.2685 | 46.43 | 43.41 | 0.5224 | 0.5124 |
| GEODIFF-A | 90.54 | 94.61 | 0.2104 | 0.2021 | 52.35 | 50.10 | 0.4539 | 0.4399 |
| GEODIFF-C | 90.07 | 93.39 | 0.2090 | 0.1988 | 52.79 | 50.29 | 0.4448 | 0.4267 |
| Models | COV-R (%) ↑ | MAT-R (A)↓ | COV-P (%) ↑ | MAT-P (A)↓ | ||||
| Mean | Median | Mean | Median | Mean | Median | Mean | Median | |
| GEODIFF (T=1000) | 82.96 | 96.29 | 0.9525 | 0.9334 | 48.27 | 46.03 | 1.3205 | 1.2724 |
| Name | Equation | Solution needed | Oracle |
| Stationary | (4), (5) | Primal | V1f |
| KKT | (6) | Primal and dual | Vif,H,G,aH,∂G |
| Proximal gradient | (7) | Primal | Vif,proxng |
| Projected gradient | (9) | Primal | Vif,projc |
| Mirror descent | (13) | Primal | V1f,proj,V |
| Newton | (14) | Primal | [V²f(x,0)]-1,Vif(x,0) |
| Block proximal gradient | (15) | Primal | [Vif]j,[proXnglj |
| Conic programming | (18) | Residual map root | projRP × K* ×R+ |
| Method | L1 logreg | L2 logreg | DictL + L2 logreg | Task-driven DictL |
| AUC (%) | 71.6 ± 2.0 | 72.4± 2.8 | 68.3 ± 2.3 | 73.2 ± 2.1 |
| Method | MNLI | SST-2 | QQP | QNLI | Avg. | Trained params |
| Non-private fine-tuning | 90.2 | 96.4 | 92.2 | 94.7 | 93.4 | 100% |
| Our results (ε= 6.7) | 87.8 | 95.3 | 87.4 | 90.8 | 90.3 | 0.94% |
| Method | Memory (GB) | Speed (seconds per epoch) |
| Full fine-tuning (DPSGD) | 27.9 | 715 |
| RGP | 9.1 | 296 |
| DP LoRA | 6.1 | 271 |
| Method | MNLI | SST-2 | QQP | QNLI | Avg. | Trained params | |
| Full | w/o DP | 87.6 | 94.8 | 91.9 | 92.8 | 91.8 | 100% |
| DP | 53.1 | 82.6 | 74.4 | 63.9 | 68.5 | ||
| LoRA | w/o DP | 87.5 | 95.1 | 90.8 | 93.3 | 91.7 | 0.24% |
| RGP | DP | 80.1 | 91.6 | 85.5 | 87.2 | 86.1 | 100% |
| Adapter | DP | 83.4 | 92.5 | 85.6 | 87.5 | 87.3 | 1.4% (r = 48) |
| Compacter | DP | 82.6 | 92.3 | 84.7 | 85.1 | 86.2 | 0.055% (r = 96,n = 8) |
| LoRA | DP | 83.5 | 92.2 | 85.7 | 87.3 | 87.2 | 0.94% (r = 16) |
| Method | MNLI | SST-2 | QQP | QNLI | Avg. | Trained params | |
| Full | w/o DP | 90.2 | 96.4 | 92.2 | 94.7 | 93.4 | 100% |
| LoRA | w/o DP | 90.6 | 96.2 | 91.6 | 94.9 | 93.3 | 0.23% |
| RGP | DP | 86.1 | 93.0 | 86.7 | 90.0 | 88.9 | 100% |
| Adapter | DP | 87.7 | 93.9 | 86.3 | 90.7 | 89.7 | 1.4% (r = 48) |
| Compacter | DP | 87.5 | 94.2 | 86.2 | 90.2 | 89.5 | 0.053% (r =96,n = 8) |
| LoRA | DP | 87.8 | 95.3 | 87.4 | 90.8 | 90.3 | 0.94% (r =16) |
| Method | BLEU | NIST | MET | ROUGE-L | CIDEr |
| GPT-2-Small +DP | 59.26 | 6.13 | 36.6 | 64.1 | 1.63 |
| GPT-2-Medium+DP | 64.2 | 7.77 | 40.02 | 66.45 | 2.00 |
| GPT-2-Large + DP | 64.51 | 8.22 | 41.5 | 67.55 | 2.13 |
| GPT-2-Medium | 70.4 | 8.85 | 46.8 | 71.8 | 2.53 |
| GPT-2-Large | 70.4 | 8.89 | 46.8 | 72.0 | 2.47 |
| Taks | 0 | δ=1/n | δ=1/10n | δ=1/100n | δ=1/1000n | Accuracy (in %) |
| MNLI | 1.88 | ε=1 | ε = 1.35 | ε=1.49 | ε = 1.61 | 86.8 (-1.0%) |
| QQP | 1.88 | ε=1 | ε = 1.40 | ε= 1.54 | ε = 1.67 | 85.2 (-2.2%) |
| QNLI | 3.01 | ε=1 | ε= 1.48 | ε = 1.64 | ε = 1.79 | 88.0 (-2.8%) |
| SST-2 | 3.63 | ε=1 | ε = 1.47 | ε= 1.64 | ε= 1.80 | 93.1 (-2.2%) |
| MNLI | 0.91 | e=3 | ε = 4.12 | ε= 4.51 | = 4.89 | 87.4 (-0.4%) |
| QQP | 0.93 | ε=3 | ε= 4.10 | = 4.49 | ε= 4.86 | 86.8 (-0.6%) |
| QNLI | 1.29 | ε=3 | ε = 4.45 | £= 4.90 | ε = 5.33 | 89.9 (-0.9%) |
| SST-2 | 1.52 | ε=3 | ε = 4.37 | = 4.83 | e = 5.25 | 94.1 (-1.2%) |
| Method | MNLI | SST-2 | QQP | QNLI | Average Accuracy | |
| Full | w/o DP | 87.6 | 94.8 | 91.9 | 92.8 | 91.8 |
| DP | 83.2 | 85.9 | 86.2 | 84.8 | 85.0 | |
| Adapter | DP | 84.6 | 92.9 | 87.4 | 89.2 | 88.5 |
| LoRA | DP | 84.5 | 92.7 | 87.1 | 88.3 | 88.2 |
| Method | hod | MNLI | SST-2 | QQP | QNLI | Average Accuracy |
| Full | w/o DP | 90.2 | 96.4 | 92.2 | 94.7 | 93.4 |
| DP | 86.4 | 90.9 | 87.5 | 89.4 | 88.6 | |
| Adapter | DP | 88.6 | 94.5 | 87.8 | 91.6 | 90.6 |
| LoRA | DP | 89.0 | 95.3 | 88.4 | 92.4 | 91.3 |
| Method | BLEU | NIST | MET | ROUGE-L | CIDEr |
| GPT-2-Medium | 68.2 | 8.62 | 46.2 | 71.0 | 2.47 |
| GPT-2-Large | 68.5 | 8.78 | 46.0 | 69.9 | 2.45 |
| Method | Val perp | BLEU | NIST | MET | ROUGE-L | CIDEr |
| GPT-2-Small + DP | 4.51 | 63.8 | 7.19 | 39.5 | 67.5 | 1.87 |
| GPT-2-Medium + DP | 4.02 | 65.5 | 8.45 | 42.7 | 67.9 | 2.23 |
| GPT-2-Large + DP | 3.87 | 66.7 | 8.63 | 44.0 | 67.8 | 2.33 |
| GPT-2-XL + DP | 3.79 | 66.1 | 8.53 | 43.0 | 68.1 | 2.28 |
| GPT-2-Medium | 3.19 | 70.4 | 8.85 | 46.8 | 71.8 | 2.53 |
| GPT-2-Large | 3.06 | 70.4 | 8.89 | 46.8 | 72.0 | 2.47 |
| GPT-2-XL | 3.01 | 69.4 | 8.78 | 46.2 | 71.5 | 2.49 |
| Method | Val perp | BLEU | NIST | MET | ROUGE-L | CIDEr |
| GPT-2-Small + DP | 4.59 | 62.7 | 7.03 | 39.2 | 66.4 | 1.85 |
| GPT-2-Medium +DP | 4.08 | 65.2 | 8.31 | 42.2 | 68.1 | 2.22 |
| GPT-2-Large +DP | 3.92 | 66.7 | 8.60 | 43.6 | 68.1 | 2.29 |
| GPT-2-XL + DP | 3.85 | 67.6 | 8.64 | 44.9 | 68.6 | 2.36 |
| Method | Val perp | BLEU | MET | TER |
| GPT-2-Small +DP | 3.82 | 38.5 | 0.34 | 0.53 |
| GPT-2-Medium + DP | 3.30 | 42.0 | 0.36 | 0.51 |
| GPT-2-Large + DP | 3.10 | 43.1 | 0.36 | 0.5 |
| GPT-2-XL + DP | 3.00 | 43.8 | 0.37 | 0.5 |
| GPT-2-Medium | 2.67 | 47.1 | 0.39 | 0.46 |
| GPT-2-Large | 2.89 | 47.5 | 0.39 | 0.45 |
| GPT-2-XL | 2.83 | 48.1 | 0.39 | 0.46 |
| Padding | F-Conv? | Black Background | Grey Background | ||||||
| Similar (%) | Dissimilar (%) | Diff (%) | Inconsistency (%) | Similar (%) | Dissimilar (%) | Diff (%) | Inconsistency (%) | ||
| Zeros | N | 99.83±0.00 | 3.21± 8.35 | -87.68 | 95.81± 2.07 | 100.00± 0.00 | 4.96± 5.93 | -95.04 | 97.85± 4.55 |
| Y | 89.24±0.98 | 89.24±0.98 | 0.00 | 18.02± 8.08 | 100.00±0.00 | 4.77± 6.52 | -95.23 | 96.79±7.13 | |
| Circular | N | 80.31±3.23 | 80.31± 3.23 | 0.00 | 34.25± 8.32 | 72.75± 0.96 | 72.75± 0.96 | 0.00 | 26.30± 5.55 |
| Y | 99.20±0.23 | 93.14±2.88 | -6.06 | 18.48±3.55 | 98.26± 0.50 | 92.40±4.23 | -5.87 | 28.67±6.18 | |
| Reflect | N | 100.00±0.00 | 15.67±12.72 | -84.33 | 91.18±13.19 | 100.00± 0.00 | 19.96±13.54 | -80.04 | 90.33±11.95 |
| Y | 100.00±0.00 | 11.70±15.38 | -88.30 | 97.33± 6.16 | 100.00±0.00 | 17.16±12.19 | -82.84 | 98.13± 3.44 | |
| Replicate | N | 100.00±0.00 | 43.39±11.42 | -56.61 | 75.32± 8.20 | 100.00± 0.00 | 33.16± 6.42 | -66.83 | 84.09± 6.47 |
| Y | 98.32±0.39 | 93.65± 1.36 | -4.67 | 32.60± 4.97 | 97.17± 0.48 | 94.99± 1.20 | -2.18 | 32.15± 5.11 | |
| Randn | N | 100.00±0.00 | 10.31±12.56 | -89.70 | 94.88± 5.55 | 99.97± 0.13 | 35.47±10.82 | -64.50 | 83.59± 8.48 |
| Y | 100.00±0.00 | 20.80±14.15 | -79.20 | 92.54±8.37 | 77.28±16.13 | 66.70±11.58 | -10.59 | 45.70±20.62 | |
| No-pad | - | 100.00±0.00 | 3.21± 8.35 | -96.79 | 95.81± 2.07 | 100.00± 0.00 | 30.07± 4.06 | -69.93 | 81.30± 2.44 |
| Model | Padding | Eval Dataset | PosENet | PPP-MAE(ours) (↑) | Performance (↑) | |
| SPC (↑) | MAE(↓) | |||||
| VGG-19 | Zeros | GMap InfinityGAN-flickr | 0.107±0.128 0.368±0.116 | 0.196±0.006 0.183±0.007 | 0.0176±0.0005 0.0163±0.0006 | 74.0972±0.0870 |
| InfinityGAN-tower GMap | 0.492±0.106 | 0.173±0.010 | 0.0179±0.0001 | |||
| Circular | 0.098±0.139 | 0.197±0.007 | 0.0158±0.0006 | 74.4716±0.0863 | ||
| InfinityGAN-flickr | 0.323±0.147 | 0.185±0.009 | 0.0137±0.0004 | |||
| Reflect | InfinityGAN-tower | 0.460±0.102 | 0.176±0.009 | 0.0184±0.0005 | ||
| GMap InfinityGAN-flickr | 0.109±0.139 0.343±0.132 | 0.196±0.007 0.185±0.008 | 0.0158±0.0002 | |||
| InfinityGAN-tower | 0.0146±0.0008 0.0168±0.0005 | 74.0516±0.0621 | ||||
| GMap | 0.460±0.113 0.084±0.137 | 0.177±0.009 | 0.0144±0.0009 73.9964±0.1079 | |||
| InfinityGAN-flickr | 0.356±0.111 | 0.197±0.006 0.184±0.007 | 0.0128±0.0012 | |||
| 0.173±0.010 | ||||||
| InfinityGAN-tower | 0.498±0.111 | 0.0156±0.0006 | ||||
| GMap | 0.125±0.154 | 0.195±0.006 | 0.0182±0.0012 | |||
| Randn | InfinityGAN-flickr | 0.374±0.137 | 0.185±0.007 | 0.0167±0.0008 | 73.7716±0.0758 | |
| InfinityGAN-tower | 0.421±0.161 | 0.181±0.010 | 0.0186±0.0012 | |||
| GMap | 0.001±0.239 | 0.204±0.013 | 0.0000±0.0000 | |||
| InfinityGAN-flickr | 0.303±0.192 | 0.187±0.012 | 0.0000±0.0000 | 62.0396±0.0830 | ||
| InfinityGAN-tower | 0.516±0.139 | 0.172±0.014 | 0.0000±0.0000 | |||
| Zeros | ||||||
| GMap InfinityGAN-flickr | 0.191±0.188 0.682±0.107 | 0.193±0.008 0.152±0.019 | 0.0162±0.0012 0.0137±0.0004 | 75.6856±0.0924 | ||
| InfinityGAN-tower | 0.144±0.017 | 0.0153±0.0013 | ||||
| Circular | 0.721±0.077 | 0.0188±0.0016 | 76.1432±0.1026 | |||
| GMap | 0.398±0.115 | 0.197±0.007 | ||||
| InfinityGAN-flickr | 0.628±0.084 | 0.159±0.013 | 0.0178±0.0005 | |||
| InfinityGAN-tower | 0.585±0.105 | 0.165±0.014 | 0.0189±0.0012 | |||
| Reflect | ||||||
| GMap | 0.197±0.185 | 0.192±0.008 | 0.0150±0.0004 | |||
| Replicate | InfinityGAN-flickr | 0.594±0.096 | 0.169±0.012 | 0.0134±0.0009 | 75.5068±0.1213 | |
| InfinityGAN-tower | 0.667±0.087 | 0.153±0.016 | 0.0157±0.0002 | |||
| GMap | 0.249±0.192 | 0.189±0.009 | 0.0138±0.0003 | |||
| InfinityGAN-flickr | 0.700 | 0.147±0.018 | 0.0114±0.0003 | 75.6122±0.0911 | ||
| Randn | InfinityGAN-tower | 2±0.095 | 0.0142±0.0007 | |||
| 0.726±0.069 | 0.142±0.016 | |||||
| GMap | 0.210±0.192 | 0.191±0.009 | 0.0147±0.0007 | 75.3076±0.1016 | ||
| InfinityGAN-flickr | 0.566±0.100 | 0.171±0.011 | 0.0122±0.0011 | |||
| InfinityGAN-tower | 0.714±0.068 | 0.142±0.015 | 0.0153±0.0004 | |||
| 0.0049±0.0001 | ||||||
| GMap InfinityGAN-flickr | 0.156 ±0.212 | 0.201 ±0.017 | ||||
| Circular | 0.184±0.012 | 0.0036±0.0001 | 0.6269±0.0015 | |||
| 0.365±0.140 | ||||||
| InfinityGAN-tower | 0.449±0.120 | 0.179±0.013 | 0.0032±0.0001 | |||
| GMap | 0.011±0.209 | 0.207±0.014 | 0.0062±0.0001 | 0.6260±0.0009 | ||
| InfinityGAN-flickr InfinityGAN-tower | 0.329±0.133 0.398±0.115 | 0.187±0.012 0.182±0.011 | 0.0068±0.0001 0.0050±0.0002 | |||
| Reflect | ||||||
| GMap | 0.062±0.210 0.205±0.016 | 0.0053±0.0001 | ||||
| InfinityGAN-flickr InfinityGAN-tower | 0.322±0.133 0.396±0.125 | 0.188±0.013 0.183±0.013 | 0.0030±0.0001 0.0039±0.0001 | 0.6243±0.0022 | ||
| Replicate | GMap | 0.204±0.016 | 0.0043±0.0002 | 0.6255±0.0013 | ||
| InfinityGAN-flickr | 0.071±0.215 0.335±0.139 | 0.0023±0.0001 | ||||
| 0.186±0.012 | ||||||
| InfinityGAN-tower | 0.409±0.120 | 0.182±0.012 | 0.0032±0.0001 | |||
| GMap | ||||||
| InfinityGAN-flickr InfinityGAN-tower | 0.002±0.244 0.228±0.173 | 0.202±0.009 0.197±0.010 | 0.0001±0.0000 0.0001±0.0000 | 0.2570±0.0022 0.0001±0.0000 | ||
| Model | Pretrained | Padding | ||||
| Zeros | Circular | Reflect | Replicate | Randn | ||
| VGG-19 | × | 0.0132±0.0006 | 0.0000±0.0000 | 0.0000±0.0000 | 0.0000±0.0000 | 0.0000±0.0000 |
| ImageNet | 0.0176±0.0005 | 0.0158±0.0006 | 0.0158±0.0002 | 0.0144±0.0009 | 0.0182±0.0012 | |
| ResNet50 | × | 0.0052±0.0004 | 0.0032±0.0004 | 0.0018±0.0001 | 0.0015±0.0001 | 0.0020±0.0002 |
| ImageNet | 0.0162±0.0012 | 0.0188±0.0016 | 0.0150±0.0004 | 0.0150±0.0004 | 0.0147±0.0007 | |
| Marginal metrics | Joint metrics | |||||
| mADE mFDE | MR | mFDE | MR | SCR | cMR | |
| ILVM (Casas et al., 2020) | 0.30 0.62 | 10.8 | 0.84 | 19.8 | 5.7 | 21.3 |
| SceneTranformer (Ngiam et al., 2021) | 0.29 0.59 | 10.5 | 0.84 | 15.7 | 3.4 | 17.3 |
| THOMAS | 0.31 0.60 | 8.2 | 0.76 | 11.8 | 2.4 | 12.7 |
| Marginal metrics | Jointmetrics | ||||||||
| Output | Sampling | Objective | mADE | mFDE | MR | mFDE | MR | Col | cMR |
| Scalar | Marg | 0.28 | 0.59 | 10.4 | 1.04 | 23.7 | 6.4 | 24.9 | |
| Scalar | Joint | 0.34 | 0.77 | 16.2 | 0.90 | 19.9 | 49 | 21.7 | |
| Heat | Learned | Marg | 0.26 | 0.46 | 4.9 | 0.98 | 20.9 | 4.1 | 21.9 |
| Heat | Learned | Joint | 0.29 | 0.58 | 9.8 | 0.88 | 15.2 | 3.0 | 16.4 |
| Heat | Algo | Marg | 0.29 | 0.54 | 3.8 | 0.83 | 14.8 | 7.2 | 15.9 |
| Heat | Algo | Joint | 0.29 | 0.54 | 3.8 | 0.83 | 14.8 | 2.6 | 15.6 |
| Heat | Combi | Joint | 0.31 | 0.60 | 8.2 | 0.76 | 11.8 | 2.4 | 12.7 |
| Training | Inference 32 agents | Inference 128 agents | |
| GOHOME | 12.8 hours | 36 ms | 90 ms |
| THOMAS | 7.5 hours | 20 ms | 31 ms |
| minSADE | minSFDE | SMR | SCR | cSMR | |
| MoliNet | 0.73 | 2.55 | 44.4 | 7.5 | 47.4 |
| ReCoG2 (Mo et al., 2020) | 0.47 | 1.16 | 23.8 | 6.9 | 26.8 |
| DenseTNT (Gu et al., 2021) | 0.42 | 1.13 | 22.4 | 0.0 | 22.4 |
| THOMAS | 0.42 | 0.97 | 17.9 | 12.8 | 25.2 |
| minSADE | minSFDE | SMR | SCR | cSMR | |
| ReCoG2 (Mo et al., 2020) | 0.33 | 0.87 | 14.98 | 0.09 | 15.12 |
| DenseTNT (Gu et al., 2021) | 0.28 | 0.89 | 15.02 | 0.0 | 15.02 |
| THOMAS | 0.31 | 0.72 | 10.67 | 0.84 | 11.63 |
| K=1 | K=6 | ||||
| minFDE | MR | minADE | minFDE | MR | |
| LaneGCN (Liang et al., 2020) | 3.78 | 59.1 | 0.87 | 1.36 | 16.3 |
| Autobot (Girgis et al., 2021) | 0.89 | 1.41 | 16 | ||
| TPCN (Ye et al., 2021) | 3.64 | 58.6 | 0.85 | 1.35 | 15.9 |
| Jean (Mercat et al., 2020) | 4.24 | 68.6 | 1.00 | 1.42 | 13.1 |
| SceneTrans (Ngiam et al.,2021) | 4.06 | 59.2 | 0.80 | 1.23 | 12.6 |
| LaneRCNN (Zeng et al., 2021) | 3.69 | 56.9 | 0.90 | 1.45 | 12.3 |
| PRIME (Song et al., 2021) | 3.82 | 58.7 | 1.22 | 1.56 | 11.5 |
| DenseTNT (Gu et al., 2021) | 3.70 | 59.9 | 0.94 | 1.49 | 10.5 |
| GOHOME (Gilles et al., 2021a) | 3.65 | 57.2 | 0.94 | 1.45 | 10.5 |
| HOME (Gilles et al., 2021b) | 3.65 | 57.1 | 0.93 | 1.44 | 9.8 |
| THOMAS | 3.59 | 56.1 | 0.94 | 1.44 | 10.4 |
| K=5 | K=10 | k=1 minFDE | |||
| minADE | MR | minADE | MR | ||
| CoverNet (Phan-Minh etal., 2020) Trajectron++ (Salzmann et al.,2020) | 1.96 | 67 | 1.48 | 57 | |
| 1.88 | 70 | 1.51 | 9.52 | ||
| ALAN (Narayanan et al., 2021) | 1.87 | 60 | 1.22 | 49 | 9.98 |
| SG-Net (Wang et al., 2021) | 1.86 | 67 | 1.40 | 52 | 9.25 |
| WIMP (Khandelwal et al., 2020) | 1.84 | 55 | 1.11 | 43 | 8.49 |
| MHA-JAM (Messaoud et al., 2020) | 1.81 | 59 | 1.24 | 46 | 8.57 |
| CXX (Luo et al., 2020) | 1.63 | 69 | 1.29 | 60 | 8.86 |
| LaPred (Kim et al.,2021) | 1.53 | 1.12 | 8.12 | ||
| P2T (Deo & Trivedi,2020) | 1.45 | 64 | 1.16 | 46 | 10.50 |
| GOHOME (Gilles et al., 2021a) | 1.42 | 57 | 1.15 | 47 | 6.99 |
| Autobot (Girgis et al., 2021) | 1.37 | 62 | 1.03 | 44 | 8.19 |
| PGP (Deo & Trivedi, 2020) | 1.30 | 57 | 0.98 | 37 | 7.72 |
| THOMAS | 1.33 | 55 | 1.04 | 1.04 | 6.71 |
| Dataset | MNIST | FMNIST | CIFAR10 | SVHN | CIFAR100 | Tiny-ImageNet | ||||||||||||
| Method | a=0.1 | Q=0.3 | a=0.5 | α=0.1 | Q=0.3 | Q=0.5 | α=0.1 | a=0.3 | α=0.5 | α=0.1 | Q=0.3 | α=0.5 | Q=0.1 | a=0.3 | α=0.5 | a=0.1 a=0.3 | a=0.5 | |
| FedAvg | 48.24 | 72.94 | 90.55 | 41.69 | 82.96 | 83.72 | 23.93 | 27.72 | 43.67 | 31.65 | 61.51 | 56.09 | 4.58 | 11.61 | 12.11 | 3.12 | 10.46 | 11.89 |
| FedDF | 60.15 | 74.01 | 92.18 | 43.58 | 80.67 | 84.67 | 40.58 | 46.78 | 53.56 | 49.13 | 73.34 | 73.98 | 28.17 | 30.28 | 36.35 | 15.34 | 18.22 | 27.43 |
| Fed-DAFL | 64.38 | 74.18 | 93.01 | 47.14 | 80.59 | 84.02 | 47.34 | 53.89 | 58.59 | 53.23 | 76.56 | 78.03 | 28.89 | 34.89 | 38.19 | 18.38 | 22.18 | 28.22 |
| Fed-ADI | 64.13 | 75.03 | 93.49 | 48.49 | 81.15 | 84.19 | 48.59 | 54.68 | 59.34 | 53.45 | 77.45 | 78.85 | 30.13 | 35.18 | 40.28 | 19.59 | 25.34 | 30.21 |
| DENSE (ours) | 66.61 | 76.48 | 95.82 | 50.29 | 83.96 | 85.94 | 50.26 | 59.76 | 62.19 | 55.34 | 79.59 | 80.03 | 32.03 | 37.32 | 42.07 | 22.44 | 28.14 | 32.34 |
| Model | Client | Server (ResNet-18) | |||||||
| ResNet-18 | CNN1 | CNN2 | WRN-16-1 | WRN-40-1 | FedDF | Fed-DAFL | Fed-ADI | DENSE (ours) | |
| Q=0.1 | 40.83 | 33.67 | 35.21 | 27.73 | 32.93 | 42.35 | 43.12 | 44.63 | 49.76 |
| Q=0.3 | 51.49 | 52.78 | 44.96 | 47.35 | 37.24 | 52.72 | 57.72 | 58.96 | 63.25 |
| α=0.5 | 59.96 | 58.67 | 54.28 | 53.39 | 58.14 | 60.05 | 61.56 | 63.24 | 67.42 |
| Dataset | CIFAR10 | SVHN | ||||||||
| m | FedAvg | FedDF | Fed-DAFL | Fed-ADI | DENSE (ours) | FedAvg | FedDF | Fed-DAFL | Fed-ADI | DENSE (ours) |
| 5 | 1 43.67 | 53.56 | 55.46 | 58.59 | 62.19 | 56.09 | 73.98 | 78.03 | 78.85 | 80.03 |
| 10 | 38.29 | 54.44 | 56.34 | 57.13 | 61.42 | 45.34 | 62.12 | 63.34 | 65.45 | 67.57 |
| 20 | 36.03 | 43.15 | 45.98 | 46.45 | 52.71 | 47.79 | 60.45 | 62.19 | 63.98 | 66.42 |
| 50 | 37.03 | 40.89 | 43.02 | 44.47 | 48.47 | 36.53 | 51.44 | 54.23 | 57.35 | 59.27 |
| 100 | 33.54 | 36.89 | 37.55 | 36.98 | 43.28 | 30.18 | 46.58 | 47.19 | 48.33 | 52.48 |
| Dataset | CIFAR10 | SVHN | ||||
| Method | α=0.1 | a=0.3 | a=0.5 | a=0.1 | α=0.3 | Q=0.5 |
| DENSE | 50.26 | 59.76 | 62.19 | 55.34 | 79.59 | 80.03 |
| DENSE+LDAM | 57.24 | 63.13 | 64.76 | 58.04 | 81.28 | 81.77 |
| Dataset | CIFAR10 | SVHN | ||||
| Communication rounds | α=0.1 | α=0.3 | α=0.5 | a=0.1 | α=0.3 | α=0.5 |
| Tc=1 | 50.72 | 59.41 | 63.89 | 54.344 | 79.87 | 80.14 |
| T=2 | 63.08 | 65.90 | 71.16 | 56.13 | 79.75 | 85.18 |
| Te=3 | 61.61 | 69.73 | 73.91 | 74.41 | 86.42 | 86.18 |
| T=4 | 66.26 | 69.40 | 74.39 | 78.67 | 86.36 | 86.43 |
| T=5 | 67.65 | 71.42 | 76.01 | 80.28 | 86.25 | 86.55 |
| Dataset | CIFAR10 | SVHN | CIFAR100 |
| DENSE | 62.19 | 80.03 | 42.07 |
| wlLcE | 53.12 | 73.11 | 36.47 |
| W/o LBN | 61.05 | 78.36 | 39.89 |
| w/o Ldiv | 59.18 | 77.59 | 39.14 |
| Dataset | Defense | Norm | E | 1-Radv-det | 1- R |
| MNIST | Grosse et al. (2017) | l8 | 0.5 | ≥ 98% | 94% |
| Ma et al. (2018) | l2 | 4.2 | ≥ 99% | 72% | |
| Raghuram et al. (2021) | l2 | 8.9 | ≥ 74% | 0% | |
| CIFAR-10 | Yin et al. (2020) | l2 | 1.7 | ≥ 90% | 66% |
| Feinman et al. (2017) | l2 | 2.7 | ≥ 43% | 36% | |
| Miller et al. (2019) | l2 | 2.9 | ≥ 75% | 30% | |
| Raghuram et al. (2021) | l | 4.0 | ≥ 56% | 10% | |
| Ma & Liu (2019) | l8 | 4/255 | ≥ 96% | 85% | |
| Roth et al. (2019) | l8 | 8/255 | ≥ 66% | 79% | |
| Lee et al. (2018) | lo | 20/255 | ≥ 81% | 59% | |
| Li et al. (2019) | lo | 26/255 | ≥ 80% | 44% | |
| ImageNet | Xu et al. (2018) | l | 1.0 | ≥ 67% | 54% |
| Ma & Liu (2019) | l8 | 2/255 | ≥ 68% | 55% | |
| Jha et al. (2019) | lo | 2/255 | ≥ 30% | 55% | |
| Hendrycks & Gimpel (2017) | l8 | 10/255 | ≥ 76% | 30% | |
| Yu et al. (2019) | l8 | 26/255 | M 7% | 5% |
| E | 1-Radv-det | |
| 8/255 | 37% | 39% |
| 16/255 | 32% | 33% |
| Method | Sample accuracy on accumulated dataset after Batch1 | |||||
| #1 | #2 | #3 | #4 | #5 | #6 | |
| No Incremental Learning | 100.0% | 90.98%2 | 92.76% | 94.89% | 98.32% | 96.11% |
| Fisher Shut-off (FS) | 100.0% | 99.74% | 98.19% | 98.88% | 98.96% | 96.89% |
| Frac. Data Ret. (FDR)[10%] | 100.0% | 97.94% | 98.39% | 98.89% | 98.71% | 98.67% |
| FDR[20%] | 100.0% | 98.71% | 98.59% | 99.36% | 98.97% | 98.78% |
| Border Ctrl. (BC)[t opk = 5] | 100.0% | 100.0% | 100.0% | 99.84% | 100.0% | 100.0% |
| BC[topk = 10] | 100.0% | 100.0% | 100.0% | 99.84% | 100.0% | 100.0% |
| FS + FDR[10%] | 100.0% | 99.74% | 98.79% | 99.52% | 99.23% | 98.78% |
| FS + FDR[20%] | 100.0% | 100.0% | 99.19% | 99.52% | 99.48% | 99.11% |
| FS +BC[topk = 5] | 100.0% | 100.0% | 100.0% | 99.84% | 100.0% | 100.0% |
| FS + BC[topk = 10] | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| Method | Sample accuracy on accumulated dataset after Batch | |||||
| #1 | #2 | #3 | #4 | #5 | #6 | |
| No Incremental Learning | 100.0% | 73.14% | 92.07% | 97.56% | 88.33% | 91.67% |
| Fisher Shut-off (FS) | 100.0% | 99.43% | 98.26% | 99.54% | 92.85% | 97.78% |
| Frac. Data Ret. (FDR)[10%] | 100.0% | 97.43% | 96.13% | 98.93% | 98.75% | 96.78% |
| FDR[20%] | 100.0% | 98.86% | 98.84% | 99.69% | 99.75% | 98.89% |
| Border Ctrl. (BC)[topk = 5] | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 99.78% |
| BC[topk = 10] | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| FS + FDR[10%] | 100.0% | 99.71% | 98.84% | 99.54% | 98.75% | 98.33% |
| FS + FDR[20%] | 100.0% | 100.0% | 99.23% | 99.85% | 99.87% | 98.89% |
| FS +BC[topk = 5] | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| FS + BC[topk = 10] | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| Method | Sample accuracy on accumulated dataset after Batch | |||||
| #1 | #2 | #3 | #4 | #5 | #6 | |
| No Incremental Learning | 100.0% | 89.34% | 93.80% | 95.60% | 97.81% | 84.78% |
| Fisher Shut-off (FS) | 100.0% | 95.36% | 99.59% | 99.55% | 99.36% | 94.56% |
| Frac. Data Ret. (FDR)[10%] | 100.0% | 95.08% | 96.07% | 97.42% | 99.61% | 98.67% |
| FDR[20%] | 100.0% | 98.36% | 98.97% | 99.85% | 100.0% | 99.67% |
| Border Ctrl. (BC)[top k = 5] | 100.0% | 100.0% | 99.79% | 99.85% | 100.0% | 100.0% |
| BC[topk = 10] | 100.0% | 100.0% | 100.0% | 100.0% | 99.74% | 100.0% |
| FS + FDR[10%] | 100.0% | 95.36% | 99.79% | 98.48% | 99.61% | 98.89% |
| FS + FDR[20%] | 100.0% | 98.36% | 99.79% | 99.69% | 100.0% | 99.67% |
| FS+BC[topk = 5] | 100.0% | 100.0% | 100.0% | 100.0% | 99.74% | 100.0% |
| FS + BC[topk = 10] | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| Pre-train dataset | Bbox type | vOC | IN-100 | INPMix ] | IN-1k | |
| MoCo-v2 HCL/Lhyp HCL/Lhyp | COCO | 1 | 64.79 | 64.84 | 41.83 | 51.17 |
| COCO | SS | 73.13 | 73.84 | 51.28 | 54.21 | |
| COCO | GT | 75.55 | 76.22 | 51.25 | 54.52 | |
| HCL HCL | COCO CoCo | SS GT | 74.19 76.51 | 75.16 76.74 | 51.35 51.63 | 55.03 55.63 |
| MoCo-v2 HCL/Lhyp | OpenImages OpenImages | 1 GT | 69.95 73.79 | 72.80 77.36 | 49.59 52.96 | 54.12 57.57 |
| HCL HCL | OpenImages OpenImages | Ss GT | 74.31 75.40 | 78.14 79.08 | 53.21 53.82 | 58.12 58.51 |
| Detection | Dataset | AP AP50 | AP75 |
| MoCo-v2 HCL/Lhyp | COCO COCO | 34.6 53.5 36.1 55.2 | 37.0 37.9 |
| HCL | COCO | 37.0 56.1 | 39.8 |
| MoCo-v2 HCL - Lhyp | VOC VOC | 51.5 79.4 53.7 80.5 | 56.1 59.4 |
| HCL | VOC | 54.4 81.4 | 60.2 |
| Segmentation | Dataset | APs AP1 | APm |
| MoCo-v2 | COCO | 30.4 50.1 | 32.3 |
| HCL/Lhyp HCL | COCO COCO | 31.5 52.0 32.5 52.9 | 33.8 34.6 |
| Method | Indicator | Datasets IN-Real | COCO |
| MoCo | Entropy | 0.633 | 0.791 |
| Supervised | Entropy | 0.671 | 0.793 |
| HCL | Norm | 0.655 | 0.839 |
| Ensemble | Entropy+Norm | 0.717 | 0.823 |
| Optim. 入 | IN-100 | IPS |
| RSGD 0.1 | 79.08 | 53.82 |
| RSGD 0.5 | 0 | 0 |
| SGD 0.1 | 70.16 | 48.47 |
| SGD 0.5 | 74.18 | 42.75 |
| Dist. | Center | IN-100 | IPS |
| 1 | - | 77.36 | 52.96 |
| Hyp. | Scene | 79.08 | 53.82 |
| Hyp. | Object | 76.96 | 52.74 |
| Euc. | Scene | 76.68 | 52.58 |
| 入 | IN-100 | IPS |
| 0.01 | 77.70 | 53.43 |
| 0.1 | 79.08 | 53.82 |
| 0.2 0.5 | 78.64 0 | 53.84 0 |
| Category | Task | Example |
| Algorithmic | Next letter | a→b |
| List first | a,b,c→a | |
| List last | a,b,c →c | |
| To uppercase | a→A | |
| Translation | French to English Spanish to English | bonjour → hello hola →hello |
| Linguistic | Present to gerund | go →going |
| Singular to plural Antonyms | cat →cats | |
| Knowledge | happy →sad | |
| Country to Capital Person to Language | France→Paris Macron→French |
| Task A(S) | Task B (0) | Regular Task A | Conflicting Task B |
| Next Letter | To Upper | 0.92 | 0.77 |
| List Last | List First | 0.95 | 0.78 |
| Present to Past | to Gerund | 0.96 | 0.95 |
| Task | Toptokensinthetaskvectorprojection |
| Previous Letter | e,y,unknown,alphabet,preceding,c Cad,zA,dit,bill |
| FR-EN | Mason, gram,immer,Santi,latin, utter,Span,Conc,English,equivalent |
| Present Gerund | cin, thats,gram, Lorenzo, cian, Simple to Isabel,uld,berto,partici,Sah |
| Country Capital | Paris, its,capital, central, Conc, cities, administrative, Los, Madrid, London |
| Model | Parameters | Dimension | Layers | Heads |
| LLaMA | 7B | 4096 | 32 | 32 |
| 13B | 5120 | 40 | 40 | |
| 30B | 6656 | 60 | 52 | |
| GPT-J | 6B | 4096 | 28 | 16 |
| Pythia | 2.8B | 2560 | 32 | 32 |
| 6.9B | 4096 | 32 | 32 | |
| 12B | 5120 | 36 | 40 |
| Category | Task | Description | Example |
| List first | Givena list of letters,output the first letter | a,b,c→a | |
| Algorithmic | List last | Given a list of letters,output the last letter | a,b,c →c |
| Next letter | Given a letter in the English alphabet,output the next letter | a→b | |
| Previous letter | Given a letter in the English alphabet,output theb → a previous letter | ||
| To lowercase | Given an uppercase letter, output the correspond-A -a ing lowercase letter | ||
| To uppercase | Given a lowercase letter,output the correspond-a→A ing uppercase letter | ||
| Translation | French to English | Given a word in French, translate to English | bonjour → hello |
| Spanish to English English to Spanish | Given a word in Spanish,translate to English | hola →hello | |
| Given a word in English, translate to Spanish | hola → hello | ||
| English to Spanish | Given a word in English,translate to French | hola →hello | |
| Linguistic | Present to gerund | given an English verb in present simple tense, output the corresponding gerund form | go →going |
| Present to past | given an English verb in present simple tense, output the corresponding verb in past simple | go →went | |
| Singular to plural | Given an English noun in singular form,output the plural form | catcats | |
| Antonyms | Given an English adjective,output an antonym | happy →sad | |
| Knowledge | Country to Capital | Given a name of a country,output the name of the capital city | France→Paris |
| Person to Language | Given a name of a person,output their nativeMacron →French language | ||
| Location to Continent | Given a name of a person,output their nativeParis → Europe | ||
| Religion | language Given a name of a location or a person,outputMuhammad -→ Islam the associated religion |
| Baseline | Hypothesis | Regular | |||
| Model | method Task type | Task name | |||
| GPT-J 6B | Algorithmic | 0.30 | 0.74 | 0.98 | |
| List first List last | 0.24 | 0.64 | 1.00 | ||
| Next letter | 0.16 | 1.00 | 0.86 | ||
| Prev letter | 0.10 | 0.36 | 0.42 | ||
| To lower | 0.00 | 0.46 | 1.00 | ||
| Knowledge | 0.00 | 0.94 | 1.00 | ||
| To upper | 0.19 | 0.72 | 0.80 | ||
| Country capital | 0.03 | 0.58 | 0.70 | ||
| Location continent | 0.09 | 0.68 | 0.78 | ||
| Location religion | 0.02 | 0.82 | 0.82 | ||
| Linguistic | Person language | 0.43 | 0.68 | 0.78 | |
| Antonyms | 0.90 | 0.98 | |||
| Plural singular | 0.08 0.00 | 0.88 | |||
| Present simple gerund | 0.76 | 0.98 0.96 | |||
| Present simple past simple | 0.02 0.14 | 0.56 | |||
| LLaMA 13B | Translation En es En fr | 0.16 | 0.34 0.36 | 0.54 | |
| Es en | 0.70 | 0.74 | |||
| Fr en | 0.06 0.13 | 0.66 | 0.76 | ||
| Algorithmic List first | 0.77 | 1.00 | 1.00 | ||
| List last | 0.07 | 0.70 | 0.92 | ||
| Next letter | 0.31 | 1.00 | 0.94 | ||
| Prev letter | 0.05 | 0.34 | 0.50 | ||
| Knowledge | To lower | 0.00 | 0.94 | 1.00 | |
| To upper | 0.00 | 0.94 | 1.00 | ||
| Country capital | 0.17 | 0.84 | 0.86 | ||
| Location continent | 0.01 | 0.70 | 0.80 | ||
| Location religion | 0.10 | 0.74 | 0.84 | ||
| 0.76 | 0.88 | ||||
| Linguistic | Person language | 0.02 0.19 | 0.74 | 0.80 | |
| Antonyms Plural singular | 0.24 | 0.84 | 0.88 | ||
| Present simple gerund | 0.00 | 0.96 | 0.96 | ||
| Translation | Present simple past simple | 0.01 | 1.00 | 0.98 | |
| En es | 0.05 | 0.78 | 0.82 | ||
| En fr | 0.15 | 0.70 | 0.84 | ||
| Es en | 0.29 | 0.76 | 0.88 | ||
| LLaMA30B | Fren | 0.25 | 0.54 | 0.72 | |
| Algorithmic List first | 0.96 | 0.98 | 1.00 | ||
| List last | 0.02 | 0.64 | 0.96 | ||
| Next letter | 0.30 | 0.98 | 0.96 | ||
| Prev letter | 0.02 | 0.56 | 0.80 | ||
| To lower | 0.00 | 1.00 | 1.00 | ||
| To upper | 0.00 | 0.90 | 1.00 | ||
| Knowledge Country capital | 0.27 | 0.72 | 0.88 | ||
| Location religion | Location continent | 0.01 0.70 0.05 | 0.86 | ||
| Person language | 0.70 0.72 | 0.88 | |||
| Linguistic | 0.01 0.76 | 0.90 | |||
| Antonyms | 0.37 0.84 | 0.82 | |||
| Plural singular | 0.21 | 0.90 | |||
| Translation | Present simple gerund | 0.00 | 0.76 | 0.98 | |
| Present simple past simple En es | 0.02 | 0.98 | 1.00 | ||
| 0.07 | 0.74 0.80 | 0.78 | |||
| En fr | 0.10 | 0.86 | |||
| Es en | 0.24 | 0.70 | 0.88 | ||
| LLaMA7B | Algorithmic | Fren | 0.20 | 0.62 | 0.78 |
| List first | 0.87 | 0.98 | 1.00 | ||
| List last Next letter | 0.03 0.03 | 1.00 0.94 | 1.00 0.88 | ||
| Prev letter | 0.04 | 0.52 | 0.58 | ||
| To lower | 0.00 | 0.74 | 1.00 | ||
| Toupper | 0.00 | 0.60 | 1.00 | ||
| Knowledge | 0.82 | 0.86 | |||
| Country capital | 0.28 | ||||
| Linguistic | Location continent | 0.02 | 0.68 | 0.72 | |
| Location religion | 0.12 | 0.84 | 0.94 | ||
| Person language | 0.02 | 0.68 | 0.78 | ||
| Antonyms Plural singular | 0.33 0.15 | 0.74 0.84 | 0.76 0.88 | ||
| Model | Task type | method Task name | Baseline | Hypothesis | Regular |
| Pythia 12B | Present simple gerund | 0.00 | 0.74 | 0.90 | |
| Present simple past simple | 0.02 | 0.94 | 0.92 | ||
| Translation | En es | 0.07 | 0.78 | 0.76 | |
| En fr | 0.04 | 0.78 | 0.88 | ||
| Es en | 0.21 | 0.68 | 0.92 | ||
| Fr en | 0.15 | 0.66 | 0.70 | ||
| Algorithmic | 0.53 | 0.98 | 0.96 | ||
| List first List last | 0.09 | 0.98 | 1.00 | ||
| Next letter | 0.15 | 0.96 | 0.76 | ||
| Prev letter | 0.00 | 0.24 | 0.42 | ||
| To lower | 0.02 | 1.00 | 1.00 | ||
| To upper | 0.00 | 0.98 | 1.00 | ||
| Knowledge Country capital | 0.19 | 0.58 0.82 | |||
| Location continent | 0.01 | 0.68 | 0.80 | ||
| Location religion | 0.07 | 0.64 | 0.78 | ||
| Linguistic | Person language | 0.01 | 0.72 | 0.86 | |
| Antonyms | 0.34 | 0.72 | 0.74 | ||
| Plural singular | 0.18 | 0.80 | 0.84 | ||
| Present simple gerund | 0.00 | 0.86 0.96 | |||
| Translation En es | Present simple past simple 0.01 0.10 | 0.76 0.44 | 0.94 | ||
| En fr | 0.48 | 0.72 0.54 | |||
| Es en | 0.16 | 0.68 | 0.80 | ||
| Fr en | 0.05 0.14 | 0.68 | 0.80 | ||
| Pythia 2.8B | Algorithmic List first | ||||
| 0.69 0.06 | 0.96 0.98 | 1.00 1.00 | |||
| List last Next letter | 0.42 | 0.86 | 0.90 | ||
| 0.22 | 0.48 | ||||
| Prev letter To lower | 0.01 | 1.00 | 1.00 | ||
| To upper | 0.00 | 1.00 | 1.00 | ||
| Knowledge | Country capital | 0.00 0.18 | 0.70 | ||
| Location continent | 0.01 | 0.62 | 0.76 0.72 | ||
| Location religion | 0.08 | 0.76 | 0.82 | ||
| Linguistic | Person language | 0.00 | 0.82 | 0.82 | |
| Antonyms | 0.37 | 0.68 | 0.76 | ||
| Plural singular | 0.13 | 0.70 | 0.78 | ||
| Translation | Present simple gerund | 0.00 | 0.86 | 0.96 | |
| Pythia 6.9B | Present simple past simple 0.03 | 0.80 | 0.92 | ||
| En es | 0.10 | 0.26 | 0.76 | ||
| En fr | 0.16 | 0.28 | 0.60 | ||
| Es en | 0.08 | 0.76 | 0.82 | ||
| Fr en | 0.10 | 0.64 | 0.82 | ||
| Algorithmic List first | 0.43 | 1.00 | 0.98 | ||
| List last Next letter | 0.08 | 0.60 | 0.98 | ||
| 0.01 | 0.66 | 0.86 | |||
| Prev letter | 0.04 | 0.28 | 0.32 | ||
| To lower | 0.00 | 1.00 | 1.00 | ||
| To upper | 0.00 | 0.94 | 1.00 | ||
| Country capital | 0.21 | 0.76 | 0.82 | ||
| Knowledge | 0.78 | ||||
| Location continent Location religion | 0.01 0.10 | 0.62 0.80 | 0.80 | ||
| Person language | 0.01 | 0.76 | 0.80 | ||
| Linguistic Antonyms | 0.33 | 0.72 | 0.74 | ||
| Translation | Plural singular | 0.14 | 0.78 | 0.88 | |
| Present simple gerund | 0.00 | 0.82 | 0.94 | ||
| Present simple past simple | 0.02 | 0.88 | 0.96 | ||
| En es | 0.11 | 0.46 | 0.70 | ||
| En fr | 0.36 | 0.60 | |||
| 0.21 | |||||
| Es en | 0.06 | 0.72 | 0.82 | ||
| Fr en | 0.14 | ||||
| 0.66 | 0.74 |
| Model | Task | Tokens |
| LLaMA 13B | Prev Letter | e,y,unknown,alphabet,preceding,c,Cad,zA,dit,bill,closer,etc, Stuart,aa,null,cin,ads,g,ulo,Ku |
| FR-EN | Mason,gram,immer,Santi,latin,utter, Span,Conc,English, equivalent,engl,Usage,none,pron,ulo,translate,adu,Wiel,grammar, ML | |
| Present Simple to Gerund | e cin, thats,gram,Lorenzo,cian,Isabel,uld,berto,partici,Sah, reporting,eing,tc,Roberto,habit,Writing,etc,ientos,ores,Dutch | |
| Country Capital | Paris,its,capital,central,Conc,cities,administrative,Los,Madrid, London,San,Isabel,exec,Ar,Bel,Wars,name,capit,Battle,History | |
| Pythia 12B | Prev Letter | r,b,a,d,m,e,p,n,t,u,h,f,c,in,g,s,the,ar,l,× |
| FR-EN | in,and,m,d,a,or,out,the,t,o,so,c,con,have,act,e,s,is, all,to | |
| to Gerund | Present Simple in,t,m,r,a,and,the,ing,action,d,o,e,current,simple,te,w, not,have,out,what | |
| CountryCapital the,in,a,C,N,B,L,M,T,P,S,R,G,and,F,I,K,U,D,H | ||
| GPT-J 6B | Prev Letter | b,c,ν,g,s,name,i,ro,n,j,d,t,A,ai,com,m,ust,test, active,k |
| FR-EN | other,name,the,true,is,social,s,active,time,car,type,money, F,force,a,public,heart,one,ms,life | |
| Present Simple to Gerund | getting, storing,working,moving,playing,doing,making,driving, shooting,picking, being, sending,putting,selling,watching, changing,taking,collecting,feeding,reading | |
| Country Capital | London,Paris,New,West,Berlin,South,Tokyo,San,Chicago,City, Moscow,Jerusalem, Amsterdam,Philadelphia,East, Madrid,Vienna, Beijing,Mexico,Germany |
| Block | Width | Depth | Kernel size Expansion ratio SE Stride NumberofWindows | ||||
| Conv | {16,24} | - | 3 | = | - | 2 | |
| MBConv-1 | {16,24} | {1,2} | {3,5} | 1 | N | 1 | |
| MBConv-2 | {24,32} | {3,4,5} | {3,5} | {4,5,6} | N | 2 | |
| MBConv-3 | {32,40} | {3,4,5,6} | {3,5} | {4,5,6} | Y | 2 | |
| Transformer-4 | {64,72} | {3,4,5,6} | - | {1,2} | - | 2 | 1 |
| Transformer-5 | {112,120,128} | {3,4,5,6,7,8} | = | {1,2} | 2 | 1 | |
| Transformer-6 | {160,168,176,184} | {3,4,5,6,7,8} | {1,2} | 1 | 1 | ||
| Transformer-7 | {208,216,224} | {3,4,5,6} | {1,2} | 2 | |||
| MBPool | {1792,1984} | = | 1 | 6 | |||
| Input resolution | {192,224,256,288} | ||||||
| FLOPs (M) | 190 | 208 | 309 | 591 |
| Scratch | 77.2 | 77.5 | 79.1 | 80.4 |
| Supernet | 76.4 | 76.6 | 78.5 | 80.6 |
| AlphaNet | DeiT | LeViT | |
| Smallest | 77.0 | 76.6 | 76.8 |
| Largest | 82.4 | 82.2 | 82.2 |
| Epoch | 1st | 90th | 180th | 270th | 360th |
| AlphaNet | 27% | 20% | 21% | 24% | 28% |
| ViT | 36% | 27% | 27% | 32% | 34% |
| Method | Dataaugmentation | Weight decay | DropConnect | Dropout |
| AlphaNet | AutoAugment | 10-5 | 0.2 | 0.2 |
| Ours | RandAugment (n = 1,m = 1) | 0 | 0 | 0 |
| Group | Method | M FLOPs | Top-1 accuracy (%) |
| 200-300 (M) | AlphaNet-A0 | 203 | 77.9 |
| NASViT-A0 (ours) | 208 | 78.2 | |
| 300-400 (M) | LeViT (Graham et al.,2021) | 300 | 76.6 |
| NASViT-A1 (ours) | 309 | 79.7 | |
| AlphaNet-A2 | 317 | 79.4 | |
| FBNetV3 (Dai et al.,2020) | 357 | 79.6 | |
| 400-500 (M) | LeViT | 406 | 78.6 |
| NASViT-A2 (ours) | 421 | 80.5 | |
| AlphaNet-A4 | 444 | 80.4 | |
| 500-600 (M) | NASViT-A3 (ours) FBNetV3 | 528 | 81.0 |
| NASViT-A4 (ours) | 557 | 80.8 | |
| 591 | 81.4 | ||
| AlphaNet | 596 | 81.1 | |
| 600 - 1000 (M) | LeViT NASViT-A5 (ours) | 658 | 80.0 |
| FBNetV3 | 757 | 81.8 | |
| AutoFormer* (Chen et al.,2021a) | 762 | 81.5 | |
| >1000 (M) | PiT-XS (Heo et al.,2021) | 1,300 1,400 | 74.7 |
| ViTAS-D* (Su et al., 2021) | 79.1 | ||
| 1,600 | 76.2 | ||
| NASViT (supernet) (ours) | 1,881 | 82.9 | |
| CVT-13-NAS*(Wu et al., 2021) | 4,100 | 82.2 | |
| Swin-Tiny* (Liu et al., 2021) | 4,500 | 81.3 | |
| CVT-13* (Wu et al., 2021) | 4,500 | 81.6 | |
| T2T-ViT-14* (Yuan et al., 2021a) | 5,200 | 81.5 | |
| Deep ViT (Zhou et al.,2021) | 6,200 | 82.3 |
| Baseline | Weak DA &Reg | Switchable scaling | Prioritize (sub) | Prioritize (sup) | |
| Top-1 (smallest) | 76.6 | 77.4 | 77.6 | 78.1 | 77.9 |
| Top-1 (supernet) | 82.2 | 82.5 | 82.6 | 82.9 | 83.0 |
| Negative Cosine Similarity Ratio | 34% | 30% | 29% | 0% | 0% |
| Architectures (M FLOPs) | A0 (203M) A1(279M)A2(317M) A3(357M) A4(444M) A5 (491M) A6 (709M) | ||||||
| AlphaNet (w/ KL) | 77.0 | 78.2 | 78.5 | 78.8 | 79.3 | 79.6 | 80.1 |
| AlphaNet (w/ KL) + Ours | 77.5 | 78.6 | 78.9 | 79.2 | 79.8 | 80.1 | 80.7 |
| AlphaNet (w/ α-div) | 77.8 | 78.9 | 79.2 | 79.4 | 80.0 | 80.3 | 80.8 |
| AlphaNet (w/ α-div) + Ours | 77.8 | 78.9 | 79.2 | 79.4 | 80.0 | 80.4 | 80.9 |
| Global | Swin | CSwin | VOLO | LSH |
| 78.5 | 78.0 | 77.9 | 78.1 | 78.0 |
| #strides | 2 | 3 | 4 |
| Top-1 Accuracy | 78.1 | 78.5 | 78.4 |
| HeadDimension | 8 | 16 | 32 |
| Top-1 Accuracy | 78.5 | 78.3 | 78.2 |
| while not converged do | |
| 1: 2: | Sample a mini-batch data from dataset |
| 3: | Sample the supernet (i.e., the biggest sub-network) from the search space and train the |
| 4: | supernet with with ground truth labels (or with KD from an external teacher model) Sample k random sub-networks from the search space and train them with KD by using the supernet network as the teacher model |
| Model | Accuracy (%) | CPU latency (ms) |
| NASViT-B0 | 78.2 | 21.0±0.4 |
| NASViT-A0 | 78.2 | 21.6±0.5 |
| AlphaNet-A0 | 77.9 | 21.4±0.5 |
| NASViT-B1 | 79.6 | 26.6±0.6 |
| NASViT-A1 | 79.7 | 27.2±0.6 |
| AlphaNet-A2 | 79.4 | 27.4±0.5 |
| NASViT-B2 | 80.6 | 29.1±0.6 |
| NASViT-A2 | 80.5 | 29.8±0.5 |
| AlphaNet-A4 | 80.4 | 30.4±0.5 |
| NASViT-A1 | NASViT-A2 | NASViT-A3 | NASViT-A4 | |
| Conv | c: 16 | c: 16 | c: 16 | c:16 |
| d:1 | d:1 | d:1 | d:1 | |
| ks: 3 | ks: 3 | ks: 3 | ks: 3 | |
| s: 2 | s: 2 | s: 2 | s:2 | |
| MBConv-1 | c: 16 | c: 16 | c: 16 | c: 16 |
| d:1 | d:1 | d:1 | d:1 | |
| ks: 3 | ks: 3 | ks: 3 | ks: 3 | |
| e:3 | e:3 | e:3 | e:3 | |
| s:1 | s:1 | s:1 | s:1 | |
| MBConv-2 | c: 24 | c: 24 | c: 24 | c: 24 |
| d:3 | d:3 | d:3 | d:3 | |
| ks: 3 | ks: 3 | ks:3 | ks: 3 | |
| e:4 | e: 4 | e:5 | e:4 | |
| s:2 | s:2 | s:2 | s:2 | |
| MBConv-3 | c: 32 | c: 32 | c: 32 | c: 32 |
| d: 3 | d:3 | d: 3 | d: 3 | |
| ks: 3 | ks: 3 | ks: 3 | ks: 3 | |
| e:4 | e: 6 | e:5 | e:6 | |
| s: 2 | s: 2 | s: 2 | s: 2 | |
| Transformer-4 | c: 64 | c: 64 | c: 64 | c: 64 |
| d: 4 | d: 4 | d: 4 | d: 4 | |
| k: 1 | k:1 | k:1 | k: 1 | |
| e:1 | e:1 | e:1 | e:1 | |
| s:2 | s: 2 | s: 2 | s: 2 | |
| Transformer-5 | c: 112 | c: 112 | c: 112 | c: 120 |
| d: 3 | d: 3 | d: 4 | d: 3 | |
| e:1 | e:1 | e:1 | e:1 | |
| s: 2 | s: 2 | s: 2 | s: 2 | |
| Transformer-6 | c: 160 | c: 160 | c: 160 | c: 160 |
| d: 3 | d: 5 | d: 7 | d: 6 | |
| e:1 | e:1 | e: 1 | e:1 | |
| s: 1 | s:1 | s: 1 | s:1 | |
| Transformer-7 | c: 216 | c: 208 | c: 216 | c: 216 |
| d: 3 | d: 4 | d: 5 | d: 6 | |
| e:1 | e:1 | e:1 | e: 1 | |
| s:2 | s: 2 | s: 2 | s:2 | |
| MBPool | c: 1792 | c: 1792 | c: 1984 | c: 1984 |
| Resolution | 192 | 224 | 256 | 288 |
| ResNet-10 (L=1), ResNet-18 (L=2), ResNet-26 (L=3) | ||||
| Layers Names | Output Size | ResNet | Applying HSA | e.g.(ResNet-10) |
| conv1 max pool | 112 × 112 × 64 56 × 56×64 | 7 × 7,64, stride 2 3 × 3, stride 2 | No | |
| conv2_x | 56 × 56× 64 | [3×3, 64] ×L [3×3, 64 | No | |
| conv3_x | 28 × 28× 128 | 3×3, 128] ×L [3×3, 128] | Yes | conv-s5t2 |
| conv4_x | 14 ×14× 256 | [3×3, 256] ×L 256 [3×3, | Yes | conv-s5t2 |
| conv5_x | 7×7×512 | [3×3, 512] ×L [3×3, 512] | Yes | conv-s1t1 |
| average pool fully connected | 1×1×512 1000 | 7×7 512 × 1000 fc | ||
| (av,ac,as,Ss) | ResNet-18 on ImageNet | ResNet-26 on ImageNet | ||||
| Accuracy | P_ratio | MA_ratio | Accuracy | P_ratio | MA_ratio | |
| (0.50,0.50,0.50,max) | 69.922 | 0.4065 | 0.4614 | 72.038 | 0.4241 | 0.4618 |
| (0.50,0.75,0.75, max) | 69.782 | 0.6014 | 0.6597 | 72.326 | 0.6308 | 0.6722 |
| (0.50,0.50, 0.50,min) | 67.572 | 0.1264 | 0.2700 | 69.970 | 0.1250 | 0.2668 |
| (0.50,0.75,0.75,min) | 67.540 | 0.1246 | 0.2986 | 69.922 | 0.1243 | 0.2910 |
| Standard (replaced layers) | 70.728 | 10.8M | 1213.8M | 72.604 | 17.03M | 1907.4M |
| Dataset | Models | Accuracy | Parameters | MAdds |
| ImageNet | ResNet-10 standard | 63.386 | 4.64M | 520M |
| BlkSConv-ResNet-18 (0.5,0.5,0.5,max) | 69.922 | 4.39M | 560M | |
| BlkSConv-ResNet-26 (0.5, 0.5, 0.5, min) | 69.970 | 2.13M | 509M | |
| CIFAR 100 | ResNet-20 standard | 67.994 | 202752 | 12.97M |
| BlkSConv-ResNet-20 (0.5, 0.5, 0.5, max) | 67.078 | 72704 | 5.04M | |
| BlkSConv-ResNet-56 (0.5,0.5, 0.5, min) | 69.994 | 149440 | 10.61M |
| (av,ac,αs,Ss) | ResNet-56 on CIFAR 10 | ResNet-56 on CIFAR 100 | ||||
| Accuracy | P_ratio | MA_ratio | Accuracy | P_ratio | MA_ratio | |
| (0.5, 0.5,0.5, max) | 93.372 | 0.3734 | 0.3829 | 70.636 | 0.3734 | 0.3829 |
| (0.5,0.75,0.75,max) | 93.338 | 0.6196 | 0.6292 | 70.668 | 0.6196 | 0.6292 |
| (0.5,0.5,0.5,min) | 93.324 | 0.2335 | 0.2462 | 69.994 | 0.2316 | 0.2570 |
| (0.5,0.75,0.75,min) | 93.324 | 0.2335 | 0.2462 | 69.994 | 0.2316 | 0.2570 |
| Standard (replaced layers) | 93.218 | 645120 | 41.28M | 70.998 | 645120 | 41.28M |
| (av,ac,as,Ss) | Stanford Dogs | Oxford 102 Flowers | ||||
| Accuracy | P_ratio | MA_ratio | Accuracy | P_ratio | MA_ratio | |
| (0.5, 0.5,0.5, max) | 53.005 | 0.5327 | 0.5394 | 65.546 | 0.5327 | 0.5394 |
| (0.5,0.75,0.75,max) | 53.359 | 0.7277 | 0.7377 | 64.567 | 0.7277 | 0.7377 |
| (0.5,0.5,0.5,min) | 53.159 | 0.3273 | 0.4006 | 65.289 | 0.4835 | 0.5179 |
| (0.5,0.75,0.75,min) | 53.615 | 0.3171 | 0.4611 | 63.217 | 0.4743 | 0.5872 |
| Standard (replaced layers) | 52.436 | 10.8M | 1213.8M | 62.238 | 10.8M | 1213.8M |
| ResNet-18 on Stanford Dogs (Qv, αc = 0.75,αs = 0.75, SS = min) | |||||||
| au | Accuracy | P_ratio | MA_ratio | au | Accuracy | P_ratio | MA_ratio |
| 0.0 | 50.918 | 0.0397 | 0.1781 | 0.3 | 52.839 | 0.1074 | 0.2377 |
| 0.1 | 50.499 | 0.0449 | 0.1777 | 0.4 | 52.035 | 0.2751 | 0.4684 |
| 0.2 | 51.040 | 0.0767 | 0.2458 | 0.5 | 53.615 | 0.3171 | 0.4611 |
| Standard | 52.436 | 10.8M | 1213.8M | ||||
| 19 | eyebrow eye ear nose mouth neck cloth hair 5 41 21 | 32 | ||||
| torso 25 eye 3 | 46 | 3462 4 | ||||
| ear 142 | 2 | 13 | ||||
| nose 81 | 4 | 10 | ||||
| (b) FFHQ2Dog |
| Resolution | 4 | 8 | 16 | 32 | 64 | 128 | 256 | 512 |
| LPIPS | 0.156 | 0.385 | 0.390 | 0.432 | 0.440 | 0.405 | 0.369 | 0.355 |
| eyebrow 19 | eye 5 | ear 41 | nose 21 | mouth 32 | neck 46 | cloth 34 | hair 62 | ||
| eyebrow | 29 | 8 | 1 | 1 | |||||
| eye | 9 | 3 | |||||||
| ear | 45 | 20 | |||||||
| nose | 23 | 8 | |||||||
| mouth | 55 | 11 | |||||||
| neck | 61 | 1 | 15 | ||||||
| cloth | 65 | 19 | |||||||
| hair | 70 | 33 |
| eyebrow 19 | eye 5 | ear 41 | nose 21 | mouth 32 | neck 46 | cloth 34 | hair 62 | ||
| eyebrow | 22 | 1 | 1 | ||||||
| eye | 1 | ||||||||
| ear | 44 | 1 | 1 | 1 | |||||
| nose | 19 | ||||||||
| mouth | 28 | 1 | 1 | 1 | |||||
| neck | 43 | 1 | 1 | 1 | 1 | ||||
| cloth | 32 | 2 | 1 | 1 | |||||
| hair | 85 | 1 | 2 | 2 |
| Zenc | Z+enc | Zopt | Zenc | Z+enc | Zopt | ||
| cat2dog | 48.8 | 68.5 | 34.2 | 16.1 | 34.8 | 7.36 | |
| dog2wild | 22.1 | 24.8 | 10.9 | 10.9 | 12.2 | 2.19 | |
| wild2dogdog2cat | 60.030.4 | 62.5 | 34.7 | 15.5 | 22.7 | 5.983.79 | |
| 21.1 | 17.9 | 14.5 | 5.49 | ||||
| (a)FID (b)KID×10³ | |||||||
| W+ enc | Z+ enc | Z enc | z_opts | W+ enc | Z+ enc | Z enc | z_opts | |
| dog2cat | 10.3 | 11.2 | 13.7 | 9.22 | 4.87 | 4.85 | 6.56 | 3.43 |
| wild2dogcat2dogdog2wild | 44.542.137.9 | 37.644.328.3 | 36.6 | 27.4 | 27.2 | 21.128.216.2 | 18.827.012.5 | 14.817.03.64 |
| 40.718.5 | 30.49.65 | 27.621.5 | ||||||
| (a)FID (b) KID×10³ | 03 | |||||||
| Mega | Metface | Dog | Cat | Wild | Unrelated FFHQ | |
| L1 in W | 0.033 | 0.057 | 0.162 | 0.172 | 0.141 | 0.391 |
| Source | Target | Avg. Likelihood | #Exact match | |||
| Original | Transfer | × | Original | Transfer | ||
| OPT-1.3B | OPT-350M | 1.05×10-11 | 1.08×10-10 | 7.5 | 0 | 0 |
| OPT-1.3B | 6.06×10-8 | 3.47×10-6 | 57.3 | 5 | 3 | |
| OPT-2.7B | 1.39×10-7 | 2.18×10-6 | 15.6 | 14 | 15 | |
| Dataset | Resolution | Type | Test size | Classes |
| ImageNet (Russakovsky et al., 2015) | 224×224 | - | 50,000 | 1,000 |
| CIFAR-10 (Krizhevsky, 2009) | 32×32 | coarse-grained | 10.000 | 10 |
| CIFAR-100 (Krizhevsky,2009) | 32×32 | coarse-grained | 10.000 | 100 |
| STL-10 (Coates et al., 2011) | 96×96 | coarse-grained | 8.000 | 10 |
| SVHN (Netzer et al., 2011) | 32×32 | coarse-grained | 26.032 | 10 |
| CUB-200-2011 (Wah et al.,2011) | 448×448 | fine-grained | 5,740 | 200 |
| Stanford Cars (Krause etal., 2013) | 448×448 | fine-grained | 8.041 | 196 |
| FGVC Aircraft (Maji et al., 2013) | 448×448 | fine-grained | 3,333 | 100 |
| Model | Attacks Clean | CIFAR-10 93.78 | CIFAR-100 74.27 | STL-10 77.59 | SVHN 96.03 | AVG. 85.42 | Model | Attacks Clean | CIFAR-10 93.78 | CIFAR-100 74.27 | STL-10 77.59 | SVHN 96.03 | AVG. 85.42 |
| 91-99A | PGD | 79.63 | 48.02 | 74.32 | 94.66 | 74.16 | PGD DIM | 86.17 | 56.38 | 74.51 | 93.94 | 77.75 | |
| DIM | 77.16 | 44.75 | 72.74 | 91.53 | 71.55 | 80.50 | 48.03 | 71.2 | 90.87 | 72.65 | |||
| DR | 72.49 | 39.04 | 72.56 | 93.27 | 69.34 | DR | 78.86 | 48.62 | 71.66 | 93.26 | 73.10 | ||
| SSP | 68.54 | 33.63 | 72.77 | 93.98 | 67.23 | SSP | 75.54 | 42.38 | 72.66 | 92.63 | 70.80 | ||
| CDA | 66.41 | 32.37 | 72.91 | 92.17 | 65.97 | P2s3152 | CDA | 66.47 | 39.30 | 69.81 | 88.09 | 65.92 | |
| CDA+gs | 86.70 | 59.43 | 73.38 | 91.61 | 77.78 | CDA+gs | 85.61 | 57.27 | 73.06 | 90.34 | 76.57 | ||
| BIA (Ours) | 57.38 | 22.47 | 69.45 | 90.44 | 59.94 | BIA(Ours) | 65.49 | 33.48 | 69.91 | 89.46 | 64.59 | ||
| BIA+DA(Ours) | 55.16 | 21.71 | 70.00 | 91.76 | 59.66 | BIA+DA(Ours) | 65.34 | 32.68 | 69.65 | 91.38 | 64.76 | ||
| BIA+RN(Ours) | 52.81 | 20.82 | 67.55 | 88.03 | 57.30 | BIA+RN(Ours) | 61.23 | 32.84 | 68.04 | 85.79 | 61.98 | ||
| 6-D5A | PGD | 79.15 | 47.73 | 74.71 | 94.86 | 74.11 | PGD | 84.55 | 54.29 | 74.55 | 93.83 | 76.81 | |
| DIM | 77.54 | 91.68 | 71.50 | DIM | 80.89 | 73.02 | |||||||
| DR | 70.72 | 43.81 | 72.96 | 93.73 | 68.51 | DR | 78.24 | 49.06 48.67 | 72.64 | 89.47 93.2 | 72.72 | ||
| SSP | 70.46 | 37.59 35.28 | 71.98 73.21 | 93.67 | 68.16 | GD-sse50 | SSP | 77.13 | 42.18 | 70.75 | 91.64 | 70.87 | |
| CDA | 81.60 | 51.53 | 71.43 | 92.64 | 74.30 | CDA | 67.75 | 35.03 | 72.53 69.00 | 88.76 | 65.14 | ||
| CDA+gs | 88.55 | 61.90 | 73.64 | 92.18 | 79.07 | CDA+gs | 85.01 | 54.71 | 72.61 | 88.69 | 75.26 | ||
| BIA (Ours) | 57.88 | 23.12 | 59.93 | BIA (Ours) | 72.02 | ||||||||
| BIA+DA(Ours) | 57.26 | 23.04 | 69.84 70.16 | 88.89 90.08 | 60.14 | BIA+DA(Ours) | 71.69 | 38.99 38.95 | 69.80 | 86.12 | 66.73 | ||
| BIA+RN (Ours) | 54.47 | 22.61 | 68.23 | 88.08 | 58.35 | BIA+RN (Ours) | 66.67 | 34.41 | 70.60 68.79 | 88.02 81.54 | 67.32 62.85 |
| Model | Attacks | CUB-200-2011 | Stanford Cars | FGVC Aircraft | AVG. | ||||||
| Res-50 | SENet154 | SE-Res101 | Res-50 | SENet154 | SE-Res101 | Res-50 | SENet154SE-Res101 | ||||
| 9I-55A | Clean | 87.35 | 86.81 | 86.56 | 94.35 | 93.36 | 92.97 | 92.23 | 92.08 | 91.90 | 90.85 |
| PGD | 80.65 | 79.58 | 80.69 | 87.45 | 89.04 | 90.30 | 84.88 | 83.92 | 82.15 | 84.30 | |
| DIM | 70.02 | 62.86 | 70.57 | 74.72 | 78.10 | 84.33 | 73.54 | 66.88 | 62.38 | 71.49 | |
| DR | 81.08 | 82.05 | 82.52 | 90.82 | 90.59 | 91.12 | 84.97 | 87.55 | 85.54 | 86.25 | |
| SSP | 62.27 | 60.44 | 71.52 | 58.02 | 75.71 | 83.02 | 54.91 | 68.74 | 63.79 | 66.49 | |
| CDA | 69.69 70.19 | 62.51 | 71.00 | 75.94 | 72.45 | 84.64 | 71.53 | 58.33 | 63.39 | 69.94 | |
| CDA+gs | 63.19 | 68.92 | 85.03 | 79.52 | 83.52 | 78.55 | 65.62 | 68.38 | 73.66 | ||
| BIA (Ours) | 32.74 25.00 | 52.99 | 58.04 | 39.61 | 69.90 | 70.17 | 28.92 | 60.31 | 46.92 | 51.07 | |
| BIA+DA(Ours) BIA+RN(Ours) | 26.13 | 40.27 46.15 | 53.24 55.07 | 22.24 20.61 | 59.48 62.64 | 61.52 63.38 | 15.36 16.50 | 47.91 52.54 | 40.83 45.48 | 40.65 43.17 | |
| 6--555 | PGD | 80.98 | 79.00 | 80.60 | 87.54 | 88.87 | 90.56 | 84.70 | 84.01 | 83.35 | 84.40 |
| DIM | 69.93 | 61.60 | 70.90 | 75.02 | 78.55 | 84.63 | 74.59 | 67.69 | 65.26 | 72.02 | |
| DR | 80.83 | 81.57 | 81.95 | 91.00 | 90.23 | 91.15 | 84.43 | 85.96 | 84.57 | 85.74 | |
| SSP | 62.94 | 58.34 | 70.45 | 61.90 | 76.27 | 83.77 | 58.78 | 69.52 | 66.88 | 67.65 | |
| CDA | 59.48 | 61.08 | 68.50 | 58.53 | 70.70 | 80.70 | 59.26 | 52.24 | 62.26 | 63.64 | |
| CDA+gs | 67.88 | 59.42 | 67.57 | 82.83 | 78.61 | 82.64 | 79.36 | 65.89 | 68.35 | 72.51 | |
| BIA (Ours) | 48.90 | 52.33 | 56.47 | 66.34 | 72.45 | ||||||
| BIA+DA(Ours) | 27.46 | 37.61 | 50.14 | 35.27 | 61.40 | 75.08 64.41 | 50.95 17.97 | 54.04 45.81 | 51.79 44.01 | 58.71 | |
| BIA+RN(Ours) | 31.77 | 43.41 | 51.09 | 42.81 | 68.90 | 66.27 | 27.75 | 52.48 | 45.57 | 42.68 47.78 | |
| P1s-152 | PGD | 73.21 | 75.89 | 76.11 | 83.99 | 86.89 | 88.24 | 79.00 | 79.30 | 75.64 | 79.81 |
| DIM | 56.30 | 59.35 | 63.36 | 67.88 | 76.37 | 79.77 | 64.21 | 62.95 | 54.49 | 64.96 | |
| DR | 77.58 | 82.07 | 80.60 | 87.96 | 90.11 | 90.67 | 77.89 | 82.27 | 80.08 | 83.25 | |
| SSP | 47.64 | 66.17 | 67.90 | 53.29 | 79.09 | 83.45 | 58.35 | 73.36 | 69.34 | 66.51 | |
| CDA | 45.15 | 53.69 | 52.86 | 57.72 | 63.09 | 73.05 | 64.87 | 46.74 | 59.11 | 57.36 | |
| CDA+gs | 59.32 | 64.24 | 60.08 | 81.15 | 82.91 | 82.56 | 75.52 | 73.72 | 66.58 | 71.79 | |
| BIA(Ours) | 43.55 | 49.50 | 55.54 | 31.65 | 54.12 | ||||||
| BIA+DA(Ours) | 25.80 | 39.71 | 52.83 | 33.43 | 52.41 | 67.21 68.03 | 38.49 27.51 | 32.46 27.42 | 51.19 47.28 | 47.08 41.60 | |
| GDr-ese5g | BIA+RN(Ours) | 23.54 | 40.13 | 51.36 | 12.39 | 47.92 | 60.59 | 32.34 | 32.85 | 46.89 | 38.67 |
| PGD | 79.29 | ||||||||||
| DIM | 63.17 | 81.14 62.01 | 79.74 65.96 | 87.66 72.88 | 90.13 | 89.80 | 83.23 | 84.31 | 81.24 | 84.06 | |
| DR | 74.84 | 78.89 | 77.93 | 86.54 | 78.29 88.82 | 81.25 89.52 | 68.89 77.80 | 65.26 78.73 | 54.79 74.47 | 68.06 80.84 | |
| SSP | 41.80 | 49.95 | 59.72 | 26.65 | 68.71 | 74.36 | 16.80 | 55.78 | 44.07 | 48.65 | |
| CDA | 52.92 | 60.96 | 57.04 | 53.64 | 73.66 | 75.51 | 62.23 | 61.42 | 59.83 | 61.91 | |
| CDA+gs | 60.86 | 61.34 | 60.10 | 74.95 | 76.35 | 78.86 | 72.94 | 68.68 | 64.66 | 68.75 | |
| BIA (Ours) | 21.79 | 29.29 | 39.13 | 9.58 | 44.46 | 49.06 | 8.04 | 27.84 | 33.87 | 29.23 | |
| BIA+DA(Ours) 12.36 | |||||||||||
| Model | Attack Clean | VGG-16 70.14 | Dense-169 75.75 | VGG-19 70.95 | Res-50 74.61 | Res-152 77.34 | Dense-121 74.22 | Inc-v3 76.19 | AVG. 74.17 |
| VGG-16 | PGD | 2.49* | 53.22 | 4.27 | 45.52 | 58.69 | 48.28 | 61.08 | 39.08 |
| DIM | 3.32* | 22.72 | 3.39 | 19.13 | 33.03 | 18.50 | 30.05 | 18.59 | |
| DR | 20.95* | 67.91 | 43.59 | 64.90 | 70.00 | 64.73 | 69.22 | 57.33 | |
| SSP | 0.95* | 39.56 | 3.42 | 26.22 | 41.68 | 34.45 | 47.46 | 27.68 | |
| CDA | 0.40* | 42.67 | 0.77 | 36.27 | 51.05 | 38.89 | 54.02 | 32.01 | |
| CDA+gs | 12.10* | 57.09 | 20.48 | 51.87 | 60.83 | 52.21 | 55.12 | 44.24 | |
| BIA (Ours) | 1.55* | 32.35 | 3.61 | 25.36 | 42.98 | 26.97 | 41.20 | 24.86 | |
| BIA+DA(Ours) | 1.04* | 24.52 | 2.07 | 18.63 | 36.43 | 19.97 | 34.54 | 19.60 | |
| BIA+RN(Ours) | 1.44* | 25.96 | 2.58 | 16.52 | 31.80 | 18.25 | 28.54 | 17.87 | |
| Dense-169 | PGD | 38.51 | 5.03* | 40.53 | 33.91 | 44.97 | 21.18 | 58.30 | 34.63 |
| DIM | 12.31 | 5.25* | 13.20 | 8.98 | 12.93 | 5.91 | 21.44 | 11.43 | |
| DR | 38.45 | 23.99* | 41.59 | 50.19 | 58.70 | 49.95 | 63.70 | 46.65 | |
| SSP | 11.53 | 1.32* | 12.54 | 12.97 | 25.66 | 9.74 | 25.58 | 14.19 | |
| CDA | 7.26 | 0.63* | 7.91 | 6.46 | 15.56 | 5.13 | 43.78 | 12.39 | |
| CDA+gs | 27.98 | 22.95* | 28.56 | 31.52 | 43.67 | 31.66 | 49.18 | 33.65 | |
| BIA (Ours) | 4.76 | 6.45* | 7.15 | 6.97 | 13.83 | 6.60 | 38.58 | 12.05 | |
| BIA+DA(Ours) | 3.17 | 3.32* | 4.09 | 4.44 | 5.85 | 3.98 | 26.51 | 7.34 | |
| BIA+RN(Ours) | 3.66 | 4.05* | 5.23 | 6.91 | 13.25 | 4.21 | 14.24 | 7.36 |
| Attacks | CUB-200-2011 | CIFAR-10 | CIFAR-100 | STL-10 | SVHN | Stanford Cars | FGVC Aircraft | ImageNet |
| CDA | 64.34 | 83.61 | 54.83 | 70.49 | 91.81 | 74.37 | 71.17 | 50.99 |
| BIA | 40.40 | 82.62 | 54.57 | 71.43 | 91.67 | 68.25 | 57.90 | 44.15 |
| BIA+DA | 29.55 | 82.94 | 54.15 | 71.70 | 87.28 | 54.65 | 55.15 | 37.15 |
| BIA+RN | 42.88 | 83.84 | 53.63 | 69.79 | 87.95 | 57.01 | 50.85 | 39.43 |
| CIFAR-10 | CIFAR-100 | STL-10 | SVHN | CUB-200-2011 (SENet-154) | Stanford Cars (SENet-154) | FGVC Aircraft (SENet-154) | |
| BIA | 57.38 | 22.47 | 69.45 | 90.44 | 52.99 | 69.90 | 60.31 |
| BIA (ensemble) | 54.96 | 21.73 | 68.94 | 85.85 | 29.15 | 46.47 | 36.87 |
| BIA+DA | 55.16 | 21.71 | 70.00 | 91.76 | 40.27 | 59.48 | 47.91 |
| BIA+DA (ensemble) | 52.97 | 20.51 | 69.51 | 89.15 | 18.47 | 38.99 | 25.41 |
| BIA+RN | 52.81 | 20.82 | 67.55 | 88.03 | 46.15 | 62.64 | 52.54 |
| BIA+RN (ensemble) | 50.99 | 22.35 | 66.06 | 81.66 | 35.40 | 41.87 | 27.81 |
| CIFAR-10 | CIFAR-100 | STL-10 | SVHN | CUB-200-2011 (SENet-154) | Stanford Cars (SENet-154) | FGVC Aircraft (SENet-154) | ImageNet (Res-152) | ||
| BIA | Paper report (Table2&3&4) | 57.38 | 22.47 | 69.45 | 90.44 | 52.99 | 69.90 | 60.31 | 42.98 |
| Random runs | 56.75 | 21.86 | 69.61 | 90.48 | 52.47 | 68.13 | 61.45 | 44.04 | |
| 57.31 | 22.32 | 69.83 | 90.35 | 52.04 | 69.48 | 57.37 | 41.67 | ||
| 57.13 | 22.82 | 70.06 | 90.01 | 53.27 | 69.74 | 58.54 | 41.02 | ||
| 57.03 | 22.38 | 70.05 | 90.43 | 53.04 | 71.91 | 62.02 | 43.09 | ||
| 57.36 | 22.50 | 69.55 | 89.78 | 51.48 | 68.15 | 58.36 | 42.00 | ||
| Paper report | 57.16 ± 0.24 | 22.39 ± 0.31 | 69.76±0.26 | 90.25±0.29 | 52.55 ± 0.69 | 69.55 ± 1.39 | 59.68 ± 1.86 | 42.47 ± 1.10 | |
| BIA+RN | (Table2&3&4) | 52.81 | 20.82 | 67.55 | 88.03 | 46.15 | 62.64 | 52.54 | 31.80 |
| 52.56 | 21.12 | 66.75 | 88.63 | 44.84 | 63.18 | 51.55 | 30.71 | ||
| 52.63 | 21.43 | 67.24 | 87.93 | 48.38 | 63.4 | 55.09 | 33.76 | ||
| 52.29 | 21.42 | 67.3 | 88.64 | 45.41 | 63.2 | 52.99 | 31.8 | ||
| 52.60 | 21.68 21.66 | 67.21 | 87.69 | 48.88 | 63.98 | 52.69 | 29.56 | ||
| 53.41 52.70±0.42 | 67.00 | 88.40 | 49.01 | 64.16 | 55.96 | 30.96 | |||
| BIA+DA | Paper report (Table2&3&4) | 21.46 ± 0.23 | 67.10± 0.23 70.00 | 88.26 ± 0.43 91.76 | 47.30 ± 2.01 40.27 | 63.58 ± 0.46 59.48 | 53.66 ± 1.81 | 31.36 ± 1.56 | |
| 55.16 | 21.71 | 47.91 | 36.43 | ||||||
| Random runs | 55.05 | 21.8 | 69.93 | 91.43 | 41.01 | 57.70 | 48.15 | 35.81 | |
| 54.47 | 21.34 | 70.05 | 91.5 | 39.96 | 57.98 | 48.30 | 35.23 | ||
| 55.30 | 21.37 | 69.79 | 92.15 | 41.89 | 58.65 | 48.78 | 36.73 | ||
| 55.15 | 21.31 | 70.05 | 92.18 | 40.73 | 59.15 | 49.08 | 36.09 | ||
| 54.52 | 21.36 | 69.71 | 92.04 | 41.92 | 58.57 | 49.02 | 37.67 | ||
| Paper report | Result 54.90 ± 0.38 | 21.44 ± 0.20 | 69.91 ± 0.15 | 91.86 ± 0.37 | 41.10 ± 0.83 | 58.41 ± 0.57 | 48.67 ± 0.42 | 36.31 ± 0.93 | |
| BIA+DA+RN | (Table 2&3&4) | 50.29 | 20.03 | 67.51 | 90.4 | 40.46 | 57.26 | 37.01 | 24.91 |
| Random runs | |||||||||
| 51.35 | 20.18 20.17 | 67.53 | 91.34 | 43.56 43.03 | 58.29 58.4 | 36.11 38.07 | 22.25 23.83 | ||
| 50.49 50.64 | 20.27 | 67.5 67.64 | 90.95 90.56 | 41.18 | 56.46 | 37.05 | 26.07 | ||
| 50.72 | 20.48 | 67.10 | 90.65 | 39.11 | 55.69 | 36.00 | 24.31 | ||
| 50.49 | 20.09 | 67.15 | 90.77 | 38.78 | 55.81 | 37.38 | 25.27 | ||
| Result | 50.74±0.36 20.24±0.15 67.38±0.24 90.85± 0.31 | 41.13 ± 2.19 | 56.93 ± 1.33 | 36.92 ± 0.87 | 24.35± 1.46 |
| SENT Demonstration Template |
| Original text in [SRC LANG]: source sentence |
| Translation into [TRG LANG]: target sentence |
| PARA_SENT Demonstration Template |
| Original text in [SRC LANG]: |
| source sentence 1 <translate> source sentence 2</translate> |
| source sentence 3 |
| Translation into [TRG LANG]: |
| target sentence 1 <translated> target sentence 2 </translated> |
| PARA Demonstration Template |
| Original text in [SRC LANG]: source paragraph |
| Translation into [TRG LANG]: |
| target paragraph |
| TYPE | TRG LANG | PARA | SENT | PARA_SENT | GTR |
| MISTRANSLATION | EN | 88 | 109 | 82 | 155 |
| JA | 224 | 295 | 223 | 334 | |
| PL | 171 | 229 | 157 | 275 | |
| TOTAL | 483 | 633 | 462 | 764 | |
| GRAMMAR | EN | 5 | 20 | 9 | 18 |
| JA | 43 | 49 | 38 | 65 | |
| PL | 57 | 87 | 66 | 59 | |
| TOTAL | 105 | 156 | 113 | 142 | |
| INCONSISTENCY | EN | 0 | 5 | 0 | 1 |
| JA | 1 | 7 | 2 | 7 | |
| PL | 1 | 19 | 1 | 7 | |
| TOTAL | 2 | 31 | 3 | 15 | |
| UNTRANSLATED | EN | 13 | 5 | 14 | 6 |
| JA | 23 | 30 | 33 | 24 | |
| PL | 23 | 16 | 25 | 4 | |
| TOTAL | 59 | 51 | 72 | 34 | |
| REGISTER | EN | 0 | 0 | 0 | 0 |
| JA | 7 | 25 | 13 | 71 | |
| PL | 0 | 0 | 0 | 0 | |
| TOTAL | 7 | 25 | 13 | 71 | |
| FORMAT | EN | 0 | n/a | n/a | 1 |
| JA | 0 | n/a | n/a | 117 | |
| PL | 0 | n/a | n/a | 8 | |
| TOTAL | 0 | n/a | n/a | 126 |
| TYPE | SOURCE | GPT-3.5 SENT TRANSLATION | GPT-3.5PARA TRANSLATION | COMMENT |
| PPNOOOUN | POMaHbI, KaK H3BeCTHO, IIeYaTa/IHCb Ha pa3Hoi 6yMare [paper]. I ropeTb OHa [she] MO>KeT IO-pa3HOMy. -RUSSIAN SOURCE (from Manaraga) | GPT-3.5 SENT (POLISH) | Romany, jak wiadomo, drukowano Jak wiadomo,powiesci drukowano na roznym papierze [paper].Imozena roznym papierze [paper]. I moze ona [she] tesknic na rózne sposoby.on [he] palic sie na rózne sposoby. GPT-3.5 PARA (POLISH) | “Paper”is a feminine noun in Russan and referred to as “she,” whereas it is a masculine noun in Polish and should be referred to as “he,”as in PARA. The absence of context in SENT leads to an incorrect translation. |
| CSIUUAUCIUNILAS | 「気付」 いやいや、(.).古倉んは每 日勤務のに拔いら -! [lit.Ms.Furukura works every day] -JAPANESE SOURCE (from Comvenience Store Womat) | “T'm sorry I didn't notice." “No, no, (..). Furukura-san works hard every day without taking any shortcuts!” GPT-3.5 SENT (ENGLISH) | “T'm sorry I didn't notice." “No, no,(..). You work every day, but you never slack off!” GPT-3.5 PARA (ENGLISH) | “Furukura-san”or“Miss Furukura” in the last source sentence is used instead of the second-person “you”as per Japanese convention. Trans- lating this sentence without context into English results in a confusing translation (SENT) that implies that the speaker refers to some other “Furukura”rather than their listener. PARA correctly translates “Furukura” as“you.” |
| JITIITTS | ,,Ne,ted'udelas nadobi!"[(you) will do thedishes!] ,Neudelam!"[(I) won't do!] ,Udelas!“[(You) will do!] CZECH SOURCE (from Crows) | Nie,teraz zrobisz zmywanie! [(you) willdo the washing] Nie zrobie![(I) won't do!] -Zrobisz to![(You) will do it!] GPT-3.5 SENT (POLISH) | Nie, teraz umyjesz naczynia [(You) will wash the dishes]! Nieumyje[(I) won't wash]! Umyjesz [(You) will wash]! GPT-3.5 PARA (POLISH) | Czech uses the same collocation as English,“do the dishes,”which is invalid in Polish. Hence, the ellipses in the last two sentences in the source text require a broader context to be translated correctly. PARA does it properly, translating both as “wash,” while SENT unsurprisingly fails to choose the correct collocation. |
| SPSIPTT ITTSTSS | back into some mother's Fendi handbag.do jakiejs torebki Fendi nalezacej We watched her apply it, unaware. | do matki.Patrzylismy, jak to robi, nieswiadomi [unaware (we)] tego. GPT-3.5 SENT (POLISH) | When we were done,the lipstick wentGdy skofczylismy,szminka wrocilaKiedyskonczylismy,szminka( wrócila do torebki Fendi jakiejs matki. Patrzylismy, jak ja naklada, nieswiadoma [unaware(she)] naszych dzialaf. GPT-3.5 PARA (POLISH) | Only from the broader context we can deduce that “unaware" refers to the mother,not the“we”(referring to children) watching her. PARA correctly attributes the state of being“unaware”to the mother, which is exhibited by its usage of the singular feminine form of the adjective. In contrast, SENT mistranslates it using the plural masculine form of the adjective “unaware,”which implies that it refers to “we”rather than the “mother.” |
| COSSSSSSSNV | Alles zu vergessen, ist gewiss schlimm [bad].Noch schlimmer [worse] ist, nichts zu vergessen (...). | とは確に悲 惨な[tragic]ことす。らに [worse]のは、何忘 。 GPT-3.5 SENT (JAPANESE) | するとは確 にい[bad]ことす。と い[worseことは、何も忘 二。 GPT-3.5 PARA (JAPANESE) | The German source translates into English as “To forget everything is bad,certainly.Worse still is to forget nothing.”It isarguably important for the translation to repeat the same word which is an equivalent of the German “schlimm"("bad"). PARA does it well, translating both as 惠 い),or“bad,”while SENT uses two different words,“tragic”and “bad" |
| PPOSETN | Bce POIJIO XOPOIIO. KHHra IPOYH- TaHa H/Jea/IbHO - He 6bICTPO H He MeJJIeHHO, MHHHMYM /JbIMa. KJIaCCH- Ka. 6bLI B copMe [in shape]. | Wszystko poszlo dobrze. Ksiazka zostala przeczytana idealnie - nie szybko i nie wolno, minimalna ilosc dymu. Klasyka. Bylem W mundurze [in uniform]. GPT-3.5 SENT (POLISH) | Wszystko poszlo dobrze. Ksiazka przeczytana idealnie - nie szybko i Klasyka. Bylem w formie [in shape]. | The ambiguity stems here from multiple meanings of the Russian noun ΦopMa ,which can mean either“shape”or“uniform.” Since one can be nie wolno,minimalna ilosc dymu.“in shape”as wellas “in a uniform”,only from the context it becomes clear which meaning was intended by the author. PARA translates it correctly as“shape”while SENT mistranslates it as“uniform.” |
| PPPPTIIPTITISS | 「あ、と煙草の5番」 「かしこまました』 [lit. (I) un- derstood] JAPANESE SOURCE (from Convenience Store Woman) | “Ah,and one pack of cigaretes,“Ah,and one pack of cigarettes, number five." "Understood." —GPT-3.5 SENT (ENGLISH) | number five.” “Right away.” GPT-3.5 PARA (ENGLISH) | This conversation is between a clerk and a customer. The Japanese expression かしこま まし is an honorific that literally means “un- derstood.”However, when choosing the best equivalent, the translator needs to consider the situation at hand to best reflect its meaning in the target language.“Understood”in SENT is technically correct, but it is an unfortunate word choice for the clerk to employ. On the other hand, |
| LANGUAGE | LANGUAGE FAMILY | MORPHOLOGICAL FEATURES | WRITING SYSTEM |
| ENGLISH | Indo-European (Germanic) | Analytic | Latin Alphabet |
| GERMAN | Indo-European (Germanic) | Fusional | Latin Alphabet |
| FRENCH | Indo-European (Romance) | Fusional | Latin Alphabet |
| POLISH | Indo-European (Slavic) | Fusional | Latin Alphabet |
| CZECH | Indo-European (Slavic) | Fusional | Latin Alphabet |
| RUSSIAN | Indo-European (Slavic) | Fusional | Cyrillic |
| JAPANESE | Japonic | Agglutinative | Kanji /Hiragana/Katakana |
| CHINESE | Sino-Tibetan | Analytic | Hanzi |
| LANG | #SENT | SRC | HUM | PARA | SENT | PARA_SENT | GTR |
| cs-pl | 163 | 2,154 | 2,027 | 2,122 | 2,123 | 2,259 | 2,065 |
| de-pl | 153 | 3,172 | 2,997 | 2,785 | 2,899 | 2,835 | 2,764 |
| ru-pl | 170 | 2,350 | 2,471 | 2,467 | 2,463 | 2,458 | 2,375 |
| ja-pl | 111 | 2,627 | 1,855 | 1,782 | 1,907 | 1,830 | 1,800 |
| en-pl | 127 | 1,702 | 1,526 | 1,444 | 1,513 | 1,483 | 1,462 |
| fr-pl | 119 | 3,253 | 2,789 | 2,641 | 2,673 | 2,654 | 2,543 |
| de-ja | 75 | 3,530 | 5,329 | 4,807 | 5,116 | 4,652 | 4,703 |
| en-ja | 176 | 1,959 | 2.617 | 2,538 | 2,653 | 2,617 | 2,634 |
| zh-ja | 194 | 2,998 | 4,124 | 3,861 | 4,249 | 3,957 | 3,978 |
| ru-ja | 193 | 2,539 | 4,753 | 3,982 | 4,348 | 4,088 | 3,921 |
| fr-ja | 195 | 2,510 | 3,426 | 3,110 | 3,355 | 3,106 | 2,958 |
| pl-ja | 188 | 1,953 | 2,944 | 3,083 | 3,418 | 3,199 | 2,972 |
| ja-en | 111 | 2,622 | 2,293 | 2,062 | 2,322 | 2,257 | 2,140 |
| pl-en | 148 | 2,696 | 3,430 | 3,234 | 3,290 | 3,273 | 3,213 |
| ru-en | 117 | 1,693 | 2.008 | 2,029 | 2,056 | 2,028 | 2,019 |
| fr-en | 120 | 3,253 | 3,123 | 3,067 | 3,150 | 3,064 | 3,098 |
| de-en | 153 | 3,172 | 3,346 | 3,361 | 3,413 | 3,325 | 3,314 |
| zh-en | 127 | 2,235 | 2.002 | 2,427 | 2,396 | 2,351 | 2,360 |
| Total | 2,640 | 46,418 | 53,060 | 50,802 | 53,344 | 51,436 | 50,319 |
| LANG | SOURCE | TARGET | PARA | SENT | PARA_SENT | GTR |
| cs-pl | 168 | 177 | 167 | 169 | 181 | 168 |
| de-en | 155 | 182 | 166 | 167 | 164 | 155 |
| de-ja | 69 | 133 | 135 | 121 | 117 | 132 |
| de-pl | 155 | 170 | 166 | 167 | 169 | 157 |
| en-ja | 169 | 168 | 166 | 161 | 169 | 169 |
| en-pl | 131 | 127 | 130 | 132 | 130 | 131 |
| fr-en | 122 | 138 | 126 | 122 | 124 | 123 |
| fr-ja | 193 | 199 | 207 | 220 | 185 | 201 |
| fr-pl | 122 | 125 | 125 | 125 | 126 | 123 |
| ja-en | 101 | 120 | 116 | 116 | 116 | 111 |
| ja-pl | 101 | 127 | 117 | 115 | 118 | 108 |
| pl-en | 148 | 156 | 149 | 145 | 151 | 145 |
| pl-ja | 189 | 153 | 174 | 196 | 178 | 191 |
| ru-en | 123 | 119 | 121 | 124 | 121 | 123 |
| ru-ja | 144 | 155 | 158 | 161 | 164 | 196 |
| ru-pl | 168 | 172 | 170 | 171 | 172 | 172 |
| zh-en | 127 | 130 | 146 | 141 | 140 | 135 |
| zh-ja | 195 | 234 | 225 | 229 | 215 | 202 |
| TOTAL | 2,580 | 2,785 | 2,764 | 2,782 | 2,740 | 2,742 |
| BOOK TITLE | AUTHOR | TRANSLATOR(S) | LANGUAGE | YEAR PUBLISHED | ||
| SOURCE | TARGET | TRANSLATION | ORIGINAL | |||
| A Children's Bible | Lydia Millet | Aga Zano | en | pl | 2022 | 2020 |
| What Can You See From Here | Mariana Leky | Agnieszka Walczy | de | pl | 2021 | 2017 |
| The Years | Annie Ernaux | Krzysztof Jarosz & | fr | pl | 2022 | 2008 |
| Manaraga | Wladimir Sorokin | Magdalena Budzinska Agnieszka Lubomira Piotrowska | pl | 2018 | 2017 | |
| Crows | Petra Dvorakova | Miroslaw Smigielski | ru cs | pl | 2020 | 2020 |
| Convenience Store Woman | Sayaka Murata | Dariusz Latos | ja | pl | 2019 | 2016 |
| Sixteen Horses | ||||||
| An Inventory of Losses | Greg Buchanan Judith Schalansky | Fuji Yoshiko Naoko Hosoi | en | ja | 2022 | 2021 |
| Dear Reader | Paul Fournel | Kei Takahashi | de | ja | 2022 | 2018 |
| The Shooting Party | Anton Chekhov | Takuya Hara | fr | ja | 2022 | 2011 |
| Sword of Destiny | ru | ja | 2022 | 1884 | ||
| Bare burial | Andrzej Sapkowski Fang Fang | Yasuko Kawano Shin'ichi Watanabe | pl zh | ja ja | 2022 2022 | 1992 |
| 2016 | ||||||
| What Can You See From Here The Years | Mariana Leky | Tess Lewis | de | en | 2021 | 2017 |
| The Story of a Life | Annie Ernaux Konstantin Paustovsky | Alison L. Strayer Douglas Smith | fr | en | 2017 2022 | 2008 |
| The Books of Jacob | Olga Yokarczuk | Jennifer Croft | ru pl | en en | 2022 | 1956 2014 |
| Convenience Store Woman | Sayaka Murata | Ginny Tapley Takemori | ja | en | 2018 | 2016 |
| Cocoon | Zhang Yueran | Jeremy Tiang | zh | en | 2022 | 2018 |
| Book An Inventory of Losses | Lang Pair | Source | Target ち以前に特筆墓所訪た |
| de-ja | Natirlich hatte ich schon davor andere bemerkenswerte Begräb- nisstätten besucht: die Toteninsel San Michele etwa,wie sie mit hohen,roten Backsteinmauern aus dem blaugrinen Wasser der Lagune von Venedig emporragt gleich einer uneinnehmbaren Fes- tung,oder das grelle Jahrmarktstreiben des Hollywood Forever Cemetery am alljährlich von der mexikanischen Bevolkerung be- gangenen Dia de los Muertos mit den orange-gelb geschmuickten Gräbern und den von der fortgeschrittenen Verwesung auf ewig zum Grinsen verdammten Totenschädeln aus bunt gefarbtem Zucker und Pappmaché.Doch keine hat mich so berihrt wie der Friedhof jener Fischersiedlung,in dessen eigentimlichem Grundriss—einer ArtKompromiss aus Kreis und Quadrat ich nichts anderes als ein Sinnbild der ungeheuerlichen Utopie zu erkennen glaubte,die ich dort verwirklicht sah: mit dem Tod vor Augen zu leben.Lange Zeit war ich üiberzeugt,an diesem Ort,dessen dänischer Name >kleine Insel" oder >vom Wasser umgeben bedeutet, sei man dem Leben näher,gerade weil seine Bewohner die Toten wortwortlich in ihre Mitte geholt hatten,anstatt sie wie sonst in unseren Breitengraden üblich- aus dem Innersten der Gemeinden vor die Stadttore zu verbannen,auch wenn der urbane Raum sich die Gräberstätten durch sein ungehemmtes Anwachsen oft nur wenig später wieder | ことはあた。たとはアの干潟の青緑色の水 中から、赤煉瓦の高い壁にまれ難攻不落の要塞のよう にた死者の島、サンシ。あるいはメキ 住民毎年にやかに死者の日を祝う、ド ーー墓地。墓はオレンと黄色の花飾られ、 カ砂糖菓子張子細工の頭蓋骨は、腐敗進 永遠の笑顔を浮かいるようた。けとも、この漁師 町の墓地ほとに私の動墓所はかた。る門 と四角の間の妥協のようの独特の輪郭、私はに ートビアの象徵見たように思た。死の前にしつ きるというートビアが、に現いた。長 いこと私確信しいた。ンマーク語「小島」と か「水にた」という意味の名前持の場所に住 は、同らいの緯度の国通常行わいよ うに、共同体の內部ら市門の外と死者た追放 代わに、死者たち文字通り町の中心に之た。 から、よ生に近いのと。とも都空間もまた 人口膨張のために、とし墓地をたた內部と 取込まる得るのけ。 | |
| A Children's Bible | en-pl | einverleibt hat. The lady urinated. “Oh, poor old thing,she has a nervous bladder!" exclaimed some- one's chubby mother.“Is that a Persian rug?" Whose mother was it? Unclear.No one would cop to it,of course. We canceled the performance. “Admit it, that was your mother,”said a kid named Rafe to a kid named Sukey, when the parents had filed out. Some of their goblets, highball glasses,and beer bottles were completely empty. Drained. Those parents were in a hurry, then. “No way,”said Sukey firmly,and shook her head. “Then who is your mother? The one with the big ass? Or the one with the clubfoot?" "Neither,”said Sukey.“So fuck you.” | Dama sie posikala. -Och,biedactwo,ma wrazliwy pecherz!-wykrzyknela czyjas pulchna matka.- Zaraz, to perski dywan? Czyja matka byla?Nie wiadomo.Oczywiscie nikt nie chcial sie przyznac. Odwolalismy przedstawienie. -No dawaj,to twoja- powiedzial chtopiec imieniem Rafe do dziewczynki imieniem Sukey, kiedy rodzice sobie poszli. Zostawili po sobie kieliszki, wysokie szklanki i butelki po piwie.Niektóre byly zupelnie puste.Do ostatniej kropelki. Tym z rodzicow sie zatem spieszylo. -W zyciu-odparla Sukey stanowczo i pokrecila glowa. - To która? Ta z wielkim dupskiem? Czy ze szpotawa stopa? -Ani jedna,ani druga. Spierdalaj. |
| LANG PAIR | TITLE | AUTHOR | TRANSLATOR(S) | YEAR PUBLISHED | |
| TRANSLATION | ORIGINAL | ||||
| ja-pl | Norwegian Wood | Haruki Murakami | Dorota Marczewska & Anna Zielinska-Elliott | 1987 | 2006 |
| de-pl | The Trial | Franz Kafka | Jakub Ekier | 1925 | 2008 |
| fr-pl | Les Miserables | Victor Hugo | Krystyna Byczewska | 1862 | 1966 |
| fr-pl | The Little Prince | Antoine de Saint-Exupéry | Jan Szwykowski | 1862 | 1967 |
| en-pl | The Valley of Fear | Arthur Conan Doyle | Tadeusz Evert | 1915 | 1927 |
| ru-pl | War and Peace | Leo Tolstoy | Andrzej Stawar | 1869 | 1958 |
| cs-pl | War with Newts | Karel Capek | Jadwiga Bulakowska | 1936 | 1949 |
| pl-ja | Solaris | Stanislaw Lem | Mitsuyoshi Numano | 1961 | 2004 |
| ru-ja | Anna Karenina | Leo Tolstoy | Hakuyo Nakamura | 1878 | 2004 |
| de-ja | Der Steppenwolf | Hermann Hesse | Fujio Nagano | 1927 | 2000 |
| fr-ja | Around the World in 80 Days | Jules Verne | Yu Takano | 1873 | 2009 |
| en-ja | Animal Farm | George Orwell | Eitaro Sayama | 1945 | 1998 |
| zh-ja | Medicine | Lu Xun | Kobai Inoue | 1919 | 1919 |
| zh-ja | The True Story of Ah Q | Lu Xun | Kobai Inoue | 1921 | 1923 |
| zh-ja | Diary of a Madman | Lu Xun | Kobai Inoue | 1921 | 1923 |
| ru-en | Confession | Leo Tolstoy | Peter Carson | 1882 | 2013 |
| zh-en | The Day the Sun Died | Yan Lianke | Carlos Rojas | 2015 | 2018 |
| ja-en | Kokoro | Natsume Soseki | Edwin McClelan | 1914 | 1957 |
| ja-en | Kokoro | Natsume Soseki | Meredith McKinney | 1914 | 2010 |
| de-en | Venus in Furs | Ritter von Leopold Sacher-Masoch | Fernanda Savage | 1870 | unclear |
| fr-en | The Debacle | Emile Zola | Leonard Tancock | 1870 | 1972 |
| LANG PAIR | NATIVE LANG | BOOKFAMILIARITY | GENDER |
| zh-en | Chinese | X | Male |
| ja-en | English | X | Male |
| de-en | Polish/English | X | Female |
| fr-en | English | X | Female |
| ru-en | Russian | X | Female |
| pl-en | Polish/English | X | Female |
| en-ja | Japanese | X | Female |
| fr-ja | Japanese | X | Male |
| de-ja | Japanese | X | Female |
| pl-ja | Polish (author) | X | Female |
| ru-ja | Japanese | X | Male |
| zh-ja | Japanese | X | Male |
| de-pl | Polish/English | X | Female |
| en-pl | Polish (author) | X | Female |
| ru-pl | Polish/Russian | X | Female |
| cs-pl | Czech | X | Male |
| ja-pl | Polish (author) | X | Female |
| fr-pl | Polish | √ | Female |
| LANGUAGE PAIR | SENT | PARA | PARA_SENT | PARA | GTR | PARA |
| Russian- English | 0 | 10 | 5 | 5 | 4 | 6 |
| Chinese - English | 1 | 9 | 3 | 7 | 3 | 7 |
| Polish - English | 4 | 6 | 4 | 6 | 1 | 9 |
| French - English | 5 | 5 | 4 | 6 | 2 | 8 |
| Japanese- English | 1 | 9 | 2 | 8 | 1 | 9 |
| German - English | 5 | 5 | 3 | 7 | 4 | 6 |
| TOTAL | 16 | 44 | 21 | 39 | 15 | 45 |
| PERCENTAGE | 26.67% | 73.33% | 35.00% | 65.00% | 25.00% | 75.00% |
| German - Japanese | 6 | 4 | 3 | 7 | 1 | 9 |
| Russian- Japanese | 4 | 6 | 4 | 6 | 2 | 8 |
| French-Japanese | 2 | 8 | 1 | 9 | 0 | 10 |
| Polish - Japanese | 2 | 8 | 4 | 6 | 0 | 10 |
| English - Japanese | 3 | 7 | 2 | 8 | 1 | 9 |
| Chinese- Japanese | 4 | 6 | 4 | 6 | 0 | 10 |
| TOTAL | 21 | 39 | 18 | 42 | 4 | 56 |
| PERCENTAGE | 35.00% | 65.00% | 30.00% | 70.00% | 6.67% | 93.33% |
| English-Polish | 0 | 10 | 3 | 7 | 4 | 6 |
| Japanese-Polish | 3 | 7 | 5 | 5 | 1 | 9 |
| French-Polish | 4 | 6 | 4 | 6 | 2 | 8 |
| Czech -Polish | 3 | 7 | 2 | 8 | 0 | 10 |
| Russian- Polish | 1 | 9 | 4 | 6 | 3 | 7 |
| German - Polish | 3 | 7 | 3 | 7 | 1 | 9 |
| TOTAL | 14 | 46 | 21 | 39 | 11 | 49 |
| PERCENTAGE | 23.33% | 76.67% | 35.00% | 65.00% | 18.33% | 81.67% |
| TOTAL | ||||||
| 51 | 129 | 60 | 120 | 30 | 150 | |
| PERCENTAGE | 28.33% | 71.67% | 33.33% | 66.67% | 16.67% | 83.33% |
| Language Pair | PARA | SENT | PARA_SENT | GTR |
| Russian-English | 0 | 0 | 1 | 0 |
| Chinese-English | 1 | 0 | 1 | 0 |
| Polish-English | 0 | 0 | 0 | 0 |
| French-English | 1 | 0 | 2 | 0 |
| Japanese-English | 2 | 1 | 2 | 3 |
| German-English | 0 | 0 | 0 | 0 |
| German-Japanese | 8 | 2 | 6 | 8 |
| Russian-Japanese | 10 | 4 | 6 | 4 |
| French-Japanese | 3 | 1 | 4 | 4 |
| Polish-Japanese | 4 | 1 | 3 | 0 |
| English-Japanese | 2 | 2 | 1 | 0 |
| Chinese-Japanese | 2 | 0 | 0 | 1 |
| English-Polish | 0 | 1 | 2 | 0 |
| Japanese-Polish | 0 | 0 | 1 | 1 |
| French-Polish | 2 | 2 | 1 | 1 |
| Czech-Polish | 1 | 2 | 1 | 0 |
| Russian-Polish | 1 | 1 | 1 | 0 |
| German-Polish | 0 | 0 | 0 | 0 |
| Total | 37 | 17 | 32 | 22 |
| SYSTEM | COMET | BLEURT | BERTSCORE | COMET-QE |
| PARA | 0.785 | 0.485 | 0.840 | 0.038 |
| SENT | 0.779 | 0.469 | 0.839 | -0.052 |
| PARA_SENT | 0.780 | 0.480 | 0.838 | -0.062 |
| GTR | 0.735 | 0.443 | 0.832 | -0.156 |
| METRIC | Acc | T | Acc (conf) | T (conf) |
| COMET | 67.41% | 0.348 | 72.78% | 0.456 |
| COMET-QE | 64.44% | 0.289 | 70.64% | 0.413 |
| BLEURT | 61.30% | 0.226 | 66.36% | 0.327 |
| BARTSCORE | 58.52% | 0.170 | 63.91% | 0.278 |
| SOURCE | TARGET | PARA | PARA_PIVOT |
| Czech | Polish | 11 | 9 |
| German | Japanese | 13 | 7 |
| German | Polish | 12 | 8 |
| French | Japanese | 9 | 11 |
| French | Polish | 11 | 9 |
| Japanese | Polish | 10 | 10 |
| Polish | Japanese | 3 | 17 |
| Russian | Japanese | 10 | 10 |
| Russian | Polish | 8 | 12 |
| Chinese | Japanese | 9 | 11 |
| TOTAL | 96 | 104 |
| TYPE | DESCRIPTION | TRG LANG | PARA SENT | PARA_SENT | GTR | |
| CONTEXT (SENTENCE) | A mistranslation that results most likely from lack of "understanding"the sentence-level context (e.g.,translating“guide”as“doradca,”or“adviser" | Japanese | 114 | 118 67 | 107 | 158 |
| instead of "przewodnik,”or“guide").This can include translating a word or a phrase into one that is semantically related but does not convey the intended | Polish English | 64 30 | 36 | 49 44 | 82 59 | |
| meaning,or translation which appear to be an outcome of translating a word semantically related to the source word, instead of the source word itself. | ||||||
| CONTEXT (PARAGRAPH) | A mistranslation that results from lack of a beyond-sentence context. This include issues such as polysemy,employment of correct pronouns,or | Japanese | 6 | 36 | 6 | 38 |
| translating elliptical expressions. | Polish English | 13 2 | 51 | 15 0 | 59 | |
| 25 | 48 | |||||
| MINOR ISSUE | A minor issue which does not significantly affect the text and can be disputable, | Japanese | 34 | 25 | 26 | 16 |
| such as translating“barked”as “howl.” | Polish | 33 | 26 | 16 | 13 | |
| English | 18 | 11 | 12 | 9 | ||
| SURFACE SIMILARITY | A translation by word which is similar to the correct translation on the surface | Japanese | 8 | 6 | 7 | 2 |
| level, but has a different meaning (e.g.,“Wilczak,”a Polish surname,instead of “wilczarz,'a"wolfhound"). | Polish | 14 | 13 | 16 | 5 | |
| English | 5 | 5 | 6 | 2 | ||
| A translation of longer phrase which is overly literal resulting in confusing and | Japanese | 15 | 52 | 34 | 84 | |
| WORD-BY-WORD | incorrect translation. | Polish | 17 | 23 | 18 | 33 |
| English | 7 | 13 | 5 | 20 | ||
| A translation with unrelated word such as “klnie"("swear") instead of “zapuka” | Japanese | 3 | 2 | 5 | 4 | |
| UNRELATED WORD | ("knock") where no apparent semantic relation could be found. | Polish | 5 | 14 | 10 | 12 |
| English | 1 | 3 | 1 | 2 | ||
| Change of subject.In the case of PARA,it occurs mostly due to merging two | Japanese | 5 | 2 | 2 | 0 | |
| SUBJECT CHANGED | sentences with two distinctive subjects where all states and/or actions are then assigned to one of them. | Polish | 6 | 0 | 5 | |
| English | 7 | 2 | 5 | |||
| A translation that results in change in factuality, such as translating afrmative | Japanese | 4 | 11 | 5 | ||
| FACTUALITY | sentence as negation or translating word by its antonym. | Polish | 0 | 2 | 1 | |
| English | 1 | 2 | 1 | |||
| NON-WORD | A translation by a non-existent (made up) word. Some examples include | Japanese | 1 | 2 | 2 | 0 |
| skillfully constructed words like 火炎棒 which was generated instead of a "torch.”While this word does not exist in Japanese (or Chinese) it follows the | Polish | 6 | 8 | 9 | ||
| compositionality rules of these languages and is fully intelligible to a native speaker(火炎“fire”and 棒“stick.") | English | 0 | 0 | 0 | ||
| Change in the grammatical mood with regard to the source text. Note that the | ||||||
| MOOD | sentence here is still grammatically correct but does not reflect the meaning | Japanese Polish | 4 1 | 9 3 | 1 4 | |
| intended by the author. | English | 0 | 0 | 0 | ||
| A translation of text which should be left untranslated such as some proper | Japanese | 0 | 0 | 0 | ||
| UNNECESSARY TRANSLATION | names. | Polish | 0 | 3 | 0 | |
| English | 1 | 1 | 1 | |||
| LANGUAGE MISMATCH | A translation into a language diferent than the target language (e.g., Chinese | Japanese | 2 | 3 | 3 | |
| instead of Japanese).Note that leaving the word in the source language classifies as an “untranslated" error. | Polish | 2 | 0 | 2 | ||
| English | 0 | 0 | 0 | 0 | ||
| NUMBER/TIME | A translation which changes number or time expression,such as translating | Japanese | 3 | 2 | 4 | |
| 1h15min as 1h30min. Note that these rarely affect the overall meaning of the text.We have not observe cases where this would be a critical issue. | Polish | 0 | 0 | 0 | 0 | |
| English | 5 | 2 | 1 | 3 | ||
| PIVOT TRANSLATION (Czech) | A mistranslation that stems from pivoting on English (annotated for cs-pl | Polish | 0 | 0 | 0 | 43 |
| language pair). | Japanese | 24 | 26 | 27 | 17 | |
| OTHER | Other issues which do not fit into any of the above. | Polish | 9 | 14 | 10 | 13 |
| English | 10 | 4 | 5 | 4 | ||
| TOTAL (Japanese) | 223 | 294 | 229 | 334 | ||
| TOTAL (Polish) | 170 | 224 | 155 | 273 | ||
| TOTAL (English) | 87 | 104 | 81 | 150 |
| TRG LANG | TYPE | SUBTYPE | PARA | SENTS | PARA_SENTS | GTR |
| JAPANESE | PARTICLE | wrong or missing | 21 | 22 | 13 | 12 |
| ADJECTIVE | wrong continuative | 0 | 2 | 3 | 0 | |
| other | 0 | 0 | 2 | 0 | ||
| VERB | tense | 3 | 7 | 1 | 14 | |
| mood | 2 | 1 | 4 | 5 | ||
| finite/non-finite | 5 | 2 | 1 | 3 | ||
| other | 2 | 5 | 6 | 0 | ||
| ORDER | wrong order | 1 | 6 | 1 | 16 | |
| OTHER | 8 | 5 | 6 | 13 | ||
| TOTAL | 42 | 50 | 37 | 63 | ||
| POLISH | ADJECTIVE | gender | 7 | 14 | 8 | 4 |
| case | 2 | 1 | 1 | 0 | ||
| NOUN | other | 1 | 1 | 1 | 1 | |
| case | 9 | 13 | 9 | 1 | ||
| PRONOUN | other | 3 | 3 | 3 | 2 | |
| omitted or wrong | 5 | 8 | 3 | 2 | ||
| VERB | case or gender | 1 | 6 | 4 | 5 | |
| aspect | 1 | 5 | 1 | 12 | ||
| person or gender | 2 | 8 | 5 | 2 | ||
| conjugation | 1 | 0 | 7 | 3 | ||
| other | 2 | 4 | 1 | 13 | ||
| PREPOSITION | omitted or wrong | 14 | 15 | 15 | 4 | |
| NUMERAL | case or gender | 2 | 1 | 0 | 1 | |
| ORDER | wrong order | 2 | 4 | 2 | 4 | |
| OTHER | 3 | 3 | 4 | 5 | ||
| TOTAL | 55 | 86 | 64 | 59 | ||
| ARTICLE | omitted or wrong | 1 | 9 | 2 | 8 | |
| ENGLISH | PREPOSITION | omitted or wrong | 3 | 7 | ||
| OTHER | 3 | 5 | ||||
| 1 | 4 | 4 | 5 | |||
| TOTAL | 5 | 20 | 9 | 18 |
| Predictors | BLEURT | ||
| Estimates | CI | p-value | |
| (Intercept) | 0.48 | 0.47-0.50 | <0.001 |
| PARA_SENT | -0.00 | -0.01-0.00 | 0.130 |
| SENT | -0.02 | -0.02-(-0.01) | <0.001 |
| GTR | -0.04 | -0.05-(-0.04) | <0.001 |
| BLEURT | |||||
| Contrast | Estimate | SE | df | t-ratio | p-value |
| PARA - PARA_SENT | 0.00477 | 0.00315 | 1074 | 1.515 | 0.780 |
| PARA - SENT | 0.01641 | 0.00315 | 1074 | 5.215 | <0.001 |
| PARA - GTR | 0.04155 | 0.00315 | 1074 | 13.205 | <0.001 |
| PARA_SENT - SENT | 0.01164 | 0.00315 | 1074 | 3.700 | 0.001 |
| PARA_SENT-GTR | 0.03678 | 0.00315 | 1074 | 11.690 | <0.001 |
| SENT - GTR | 0.02514 | 0.00315 | 1074 | 7.990 | <0.001 |
| Predictors | COMET | ||
| Estimates | CI | p-value | |
| (Intercept) | 0.79 | 0.77-0.80 | <0.001 |
| PARA_SENT | -0.01 | -0.01-(-0.00) | 0.019 |
| SENT | -0.01 | -0.01-(-0.00) | 0.004 |
| GTR | -0.05 | -0.05-(-0.05) | <0.001 |
| COMET | |||||
| Contrast | Estimate | SE | df | t-ratio | p-value |
| PARA-PARA_SENT | 0.00563 | 0.00239 | 1074 | 2.356 | 0.112 |
| PARA - SENT | 0.00691 | 0.00239 | 1074 | 2.893 | 0.023 |
| PARA - GTR | 0.04998 | 0.00239 | 1074 | 20.928 | <.001 |
| PARA_SENT-SENT | 0.00128 | 0.00239 | 1074 | 0.536 | 1.000 |
| PARA_SENT -GTR | 0.04435 | 0.00239 | 1074 | 18.571 | <.001 |
| SENT -GTR | 0.04307 | 0.00239 | 1074 | 18.035 | <.001 |
| Predictors | COMET-QE | ||
| Estimates | CI | p-value | |
| (Intercept) | -0.04 | -0.06 - -0.01 | 0.004 |
| PARA_SENT | -0.01 | -0.03 - -0.00 | 0.026 |
| SENT | -0.02 | -0.04- -0.01 | <0.001 |
| GTR | -0.12 | -0.13 - -0.11 | <0.001 |
| COMET-QE | |||||
| Contrast | Estimate | SE | df | t-ratio | p-value |
| PARA -PARA_SENT | 0.01464 | 0.00655 | 1074 | 2.235 | 0.154 |
| PARA - SENT | 0.02376 | 0.00655 | 1074 | 3.628 | 0.002 |
| PARA - GTR | 0.11848 | 0.00655 | 1074 | 18.092 | <.001 |
| PARA_SENT-SENT | 0.00912 | 0.00655 | 1074 | 1.392 | 0.9844 |
| PARA_SENT-GTR | 0.10384 | 0.00655 | 1074 | 15.857 | <.001 |
| SENT -GTR | 0.09472 | 0.00655 | 1074 | 14.464 | <.001 |
| Predictors | BERTSCORE | ||
| Estimates | CI | p-value | |
| (Intercept) | 0.84 | 0.83-0.85 | <0.001 |
| PARA_SENT | -0.00 | -0.00-0.00 | 0.037 |
| SENT | -0.00 | -0.00-0.00 | 0.522 |
| GTR | -0.01 | -0.01-0.01 | <0.001 |
| BERTSCORE | |||||
| Contrast | Estimate | SE | df | t-ratio | p-value |
| PARA -PARA_SENT | 0.002422 | 0.00116 | 1074 | 2.082 | 0.225 |
| PARA - SENT | 0.000745 | 0.00116 | 1074 | 0.640 | 1.000 |
| PARA - GTR | 0.007508 | 0.00116 | 1074 | 6.454 | <0.001 |
| PARA_SENT - SENT | -0.001678 | 0.00116 | 1074 | -1.442 | 0.897 |
| PARA_SENT -GTR | 0.005086 | 0.00116 | 1074 | 4.372 | <0.001 |
| SENT - GTR | 0.006763 | 0.00116 | 1074 | 5.814 | <0.001 |
| Methods | APped | APdivider | APboundary | mAP |
| STSU (Can et al., 2021) | 7.0 | 11.6 | 16.5 | 11.7 |
| HDMapNet (Camera) (Li et al., 2021) | 14.4 | 21.7 | 33.0 | 23.0 |
| HDMapNet (LiDAR) (Li et al.,2021) | 10.4 | 24.1 | 37.9 | 24.1 |
| HDMapNet (Fusion) (Li et al., 2021) | 16.3 | 29.6 | 46.7 | 31.0 |
| VectorMapNet (Camera) | 36.1 | 47.3 | 39.3 | 40.9 |
| VectorMapNet (Camera) + fine-tune | 42.5 | 51.4 | 44.1 | 46.0 |
| VectorMapNet (LiDAR) | 25.7 | 37.6 | 38.6 | 34.0 |
| VectorMapNet (Fusion) | 37.6 | 50.5 | 47.5 | 45.2 |
| VectorMapNet (Fusion) + fine-tune | 48.2 | 60.1 | 53.0 | 53.7 |
| Frechet Distance | Chamfer Distance | ||||||||
| Keypoint Representaion | #dim | APped | APdivider | APboundary | mAP | APped | APdivider | APboundary | mAP |
| HDMapNet (Camera) Li et al. (2021) | 2 | - | 1 | - | 13.1 | 5.7 | 37.6 | 18.8 | |
| VectorMapNet (Camera) | 2 | 43.2 | 45.5 | 52.0 | 46.9 | 38.3 | 36.1 | 39.2 | 37.9 |
| VectorMapNet (Camera) | 3 | 41.7 | 42.3 | 49.9 | 44.6 | 36.5 | 35.0 | 36.2 | 35.8 |
| Fréchet Distance | Chamfer Distance | ||||||||
| Keypoint Representaion | k | APped | APdivider | APboundary | mAP | APped | APdivider | APboundary | mAP |
| Bbox | 2 | 47.4 | 46.9 | 62.8 | 52.4 | 36.1 | 47.3 | 39.3 | 40.9 |
| SME | 3 | 47.0 | 47.4 | 56.9 | 50.4 | 27.6 | 34.4 | 35.4 | 32.5 |
| Extreme | 4 | 41.7 | 47.3 | 59.0 | 49.4 | 30.4 | 33.1 | 37.3 | 33.6 |
| Prediction Model Inputs | minADE↓ | minFDE↓ | MR@2m↓ |
| Traj. | 0.909 | 1.577 | 19.6 |
| Traj. + G.T. Map | 0.779 | 1.390 | 18.0 |
| Traj. + Pred. Map | 0.826 | 1.477 | 18.2 |
| Frechet Distance | Chamfer Distance | |||||||
| Vertex Sampling Method | APped | APdivider | APboundary | mAP | APped | APdivider | APboundary | mAP |
| curvature-based | 47.0 | 47.4 | 56.9 | 50.4 | 27.6 | 34.4 | 35.4 | 32.5 |
| fixed interval | 26.0 | 23.6 | 37.1 | 28.9 | 14.6 | 17.6 | 18.7 | 17.0 |
| Frechet Distance | Chamfer Distance | |||||||
| Modeling Method | APped | APdivider | APboundary | mAP | APped | APdivider | APboundary | mAP |
| discrete | 47.0 | 47.4 | 56.9 | 50.4 | 27.6 | 34.4 | 35.4 | 32.5 |
| continuous | 38.0 | 41.6 | 46.1 | 41.9 | 26.5 | 28.1 | 30.1 | 26.5 |
| Key: Pretrained | Examplesseen duringpretraining | UsesFixRes[24] | 3Other non-contrastivetrainingobjective | ||||||
| Architecture | Batch | Examples seen | Parameters | ImageNet top-1 % Test | |||||
| Image | Text | size | per token | V2 | R | ||||
| COCA [25] | ViT-g | T-g | 65k | 32.8B | 1.1B | 86.3 | 80.7 | 96.5 | 90.2 |
| BASIC[18] | CoAtNet-7* | T-H* | 65k | 19.7BPT +32.8B | 1.5B | 85.7 | 80.6 | 95.7 | 85.6 |
| LIT[16] | ViT-g* | T-g | 32k | 25.8BPT + 18.2B | 1.1B | 84.5 | 78.7 | 93.9 | 79.4 |
| ALIGN [10] | EffNet-L2 | T-L* | 16k | 19.8B | ~410M | 76.4 | 70.1 | 92.2 | 75.8 |
| CLIP [7] | ViT-L/14† | T-B | 32k | 12.8B | ~ 200M | 76.2 | 70.1 | 88.9 | 77.2 |
| LIMoE | H/14 | 21k | 23.3B | 675M | 84.1 | 77.7 | 94.9 | 78.7 | |
| Model | i1k Oshot | i1k 10shot | coco t2i | coco i2t | |
| dense one-tower | 49.8 50.4 49.2 | 43.8 44.3 43.3 | 23.7 24 | 23.4 | 36.7 38.9 34.6 |
| dense two-tower | 54.7 55.2 54.1 | 47.1 47.6 46.7 | 26.6 27.1 26.2 | 41.3 42.0 40.6 | |
| LIMoE | B71 56.9 | 58 50.5 | 25.6 | 28 | 39.7 42.2 37.1 |
| Validation | Oshot | 10shot | ||||
| Auxiliary loss | X | √ | X | √ | X | √ |
| Importance | 70.5 | 70.6 | 55.4 | 56.2 | 51.1 | 51.3 |
| Load | 70.3 | 70.6 | 56.2 | 55.7 | 51.3 | 51.1 |
| Z-Loss | 70.3 | 70.6 | 55.8 | 56.2 | 50.5 | 51.3 |
| Global Ent Image | 70.6 | 70.5 | 56.0 | 56.2 | 50.8 | 51.3 |
| Global Ent Text | 69.1 | 70.6 | 54.3 | 56.2 | 51.1 | 51.3 |
| Local Ent Image | 70.6 | 68.7 | 56.2 | 53.5 | 51.3 | 47.5 |
| Local Ent Text | 67.2 | 70.6 | 53.3 | 56.2 | 47.5 | 51.3 |
| Position Method | Train Length | Speed (↑) | Memory (↓) | |
| Train | Eval. | |||
| Sinusoidal | 512 | 28.5k | 82.1k | 15.3 GB |
| 1024 | 26.0k | 77.8k | 19.2 GB | |
| 3072 | 15.3k | 42.4k | 15.1 GB | |
| Rotary | 512 | 20.0k | 43.4k | 17.8 GB |
| 1024 | 17.7k | 39.4k | 22.8 GB | |
| 3072 | 11.5k | 29.5k | 17.8 GB | |
| T5 Bias | 512 | 14.4k | 21.8k | 16.9 GB |
| 1024 | 13.0k | 20.2k | 20.9 GB | |
| 3072 | 4.3k | 4.9k | 15.9 GB | |
| ALiBi | 512 | 28.3k | 85.8k | 15.3 GB |
| 1024 | 25.8k | 76.4k | 19.3 GB | |
| 3072 | 15.5k | 42.2k | 15.2 GB | |
| Sinusoidal | Rotary | T5 Bias | ALiBi | |||||
| Inputs | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) |
| 512 | 20.05 | 15046 | 20.07 | 10839 | 19.65 | 11724 | 19.73 | 14726 |
| 513 | 19.98 | 14925 | 20.01 | 10806 | 19.57 | 10491 | 19.62 | 14965 |
| 522 | 19.93 | 15116 | 20.02 | 11295 | 19.57 | 9970 | 19.64 | 15316 |
| 532 | 19.91 | 15358 | 19.98 | 10854 | 19.53 | 10382 | 19.61 | 15383 |
| 542 | 19.91 | 15076 | 19.94 | 10795 | 19.47 | 12270 | 19.57 | 15301 |
| 552 | 19.91 | 16394 | 19.93 | 12267 | 19.47 | 13000 | 19.54 | 16540 |
| 562 | 19.91 | 16646 | 19.87 | 12481 | 19.39 | 12201 | 19.49 | 16385 |
| 572 | 19.95 | 16934 | 19.83 | 12668 | 19.36 | 12851 | 19.46 | 16881 |
| 582 | 20.13 | 16961 | 19.88 | 12594 | 19.41 | 13904 | 19.48 | 17064 |
| 592 | 20.18 | 17243 | 19.84 | 13007 | 19.36 | 13706 | 19.43 | 17289 |
| 602 | 20.40 | 17502 | 19.81 | 12788 | 19.33 | 14102 | 19.38 | 17141 |
| 612 | 20.59 | 17637 | 19.81 | 12601 | 19.27 | 14573 | 19.38 | 17661 |
| 712 | 24.86 | 15614 | 19.79 | 12676 | 19.10 | 13818 | 19.14 | 15637 |
| 812 | 30.82 | 17151 | 20.17 | 13954 | 18.94 | 14377 | 18.99 | 17210 |
| 912 | 37.42 | 17200 | 20.73 | 13887 | 18.86 | 15345 | 18.88 | 17619 |
| 1012 | 43.54 | 16304 | 21.37 | 13759 | 18.79 | 14240 | 18.73 | 16059 |
| 1112 | 50.36 | 16424 | 22.01 | 13891 | 18.77 | 14014 | 18.68 | 16659 |
| 1212 | 58.01 | 17294 | 23.02 | 15245 | 18.87 | 14589 | 18.67 | 17372 |
| 1312 | 63.62 | 15314 | 23.93 | 13698 | 18.84 | 13138 | 18.60 | 15698 |
| 1412 | 70.75 | 15663 | 24.81 | 13928 | 18.87 | 12857 | 18.59 | 15860 |
| 1512 | 76.23 | 15812 | 25.99 | 14248 | 18.91 | 13752 | 18.52 | 16225 |
| 2512 | 132.41 | 15254 | 31.58 | 13456 | 20.41 | 9948 | 18.41 | 15204 |
| 3512 | 178.97 | 13293 | 35.54 | 11850 | 22.91 | 7847 | 18.40 | 13329 |
| 4512 | 209.37 | 11767 | 39.15 | 10485 | 25.91 | 6146 | 18.41 | 11738 |
| 5512 | 240.44 | 10168 | 43.14 | 9020 | 29.54 | 5309 | 18.36 | 9986 |
| 6512 | 271.40 | 9052 | 47.81 | 8108 | 34.48 | 4680 | 18.35 | 9022 |
| 7512 | 293.02 | 8315 | 51.12 | 7483 | 39.29 | 4102 | 18.33 | 8324 |
| 8512 | 305.65 | 7259 | 54.98 | 6718 | 43.08 | 3660 | 18.34 | 7366 |
| 9512 | 336.02 | 6672 | 57.85 | 6211 | 48.90 | 3370 | 18.34 | 6555 |
| 10512 | 341.53 | 6126 | 60.77 | 5575 | 52.95 | 3010 | 18.32 | 6030 |
| 11512 | 362.74 | 5994 | 66.62 | 5445 | 61.38 | 2873 | 18.32 | 5882 |
| 12512 | 373.17 | 5421 | 69.70 | 4988 | 64.94 | 2602 | 18.31 | 5287 |
| 13512 | 382.91 | 5174 | 73.27 | 4692 | OOM | 18.31 | 4962 | |
| 14512 | 399.98 | 4351 | 75.52 | 4103 | OOM | = = | 18.31 | 4352 |
| 15512 | 406.01 | 4291 | 79.25 | 3969 | OOM | 18.31 | 4289 | |
| Sinusoidal | Rotary | T5 Bias | ALiBi | |||||
| Inputs | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) |
| 1024 | 19.34 | 17002 | 19.33 | 14690 | 18.80 | 14973 | 18.66 | 16951 |
| 1025 | 19.33 | 16630 | 19.34 | 14423 | 18.82 | 14635 | 18.67 | 16690 |
| 1034 | 19.27 | 16589 | 19.28 | 14351 | 18.74 | 14435 | 18.60 | 16707 |
| 1044 | 19.26 | 16760 | 19.27 | 14491 | 18.72 | 14644 | 18.60 | 16667 |
| 1054 | 19.23 | 16747 | 19.26 | 14503 | 18.71 | 14800 | 18.58 | 16833 |
| 1064 | 19.21 | 16676 | 19.22 | 14623 | 18.70 | 14498 | 18.55 | 16941 |
| 1074 | 19.19 | 16879 | 19.19 | 14464 | 18.65 | 14670 | 18.49 | 16936 |
| 1084 | 19.22 | 16942 | 19.23 | 14650 | 18.70 | 14607 | 18.56 | 17090 |
| 1094 | 19.24 | 16771 | 19.22 | 14629 | 18.69 | 14517 | 18.54 | 16880 |
| 1104 | 19.28 | 16870 | 19.27 | 14837 | 18.69 | 14635 | 18.52 | 17009 |
| 1114 | 19.29 | 16795 | 19.27 | 14879 | 18.69 | 14540 | 18.52 | 17050 |
| 1124 | 19.26 | 17312 | 19.18 | 15121 | 18.62 | 14480 | 18.46 | 17571 |
| 1224 | 20.54 | 17901 | 19.38 | 15584 | 18.58 | 14956 | 18.40 | 18013 |
| 1324 | 23.13 | 16308 | 19.96 | 14386 | 18.52 | 13726 | 18.33 | 16422 |
| 1424 | 26.45 | 16217 | 21.27 | 14385 | 18.48 | 13516 | 18.28 | 16121 |
| 1524 | 29.82 | 16377 | 22.59 | 14693 | 18.42 | 13587 | 18.22 | 16659 |
| 1624 | 34.27 | 15928 | 24.34 | 14228 | 18.40 | 12979 | 18.17 | 16053 |
| 1724 | 38.24 | 16640 | 25.66 | 14686 | 18.35 | 12976 | 18.15 | 16607 |
| 1824 | 42.23 | 16840 | 27.63 | 14918 | 18.30 | 13071 | 18.08 | 16846 |
| 1924 | 46.46 | 15071 | 29.64 | 13452 | 18.31 | 11843 | 18.08 | 15118 |
| 2024 | 51.09 | 15591 | 31.17 | 13706 | 18.34 | 11906 | 18.05 | 15557 |
| 3024 | 96.46 | 13639 | 35.67 | 12256 | 18.62 | 8480 | 17.92 | 13668 |
| 4024 | 144.00 | 12441 | 44.30 | 11203 | 19.44 | 7443 | 17.95 | 12402 |
| 5024 | 182.31 | 11431 | 48.31 | 10324 | 20.47 | 6384 | 17.92 | 11394 |
| 6024 | 214.02 | 10238 | 54.78 | 9117 | 21.76 | 5577 | 18.01 | 10119 |
| 7024 | 261.86 | 8785 | 62.83 | 7950 | 23.64 | 4867 | 17.93 | 8779 |
| 8024 | 284.88 | 8132 | 64.91 | 7355 | 25.79 | 4377 | 17.96 | 8086 |
| 9024 | 310.04 | 7045 | 71.91 | 6380 | 27.54 | 3787 | 17.98 | 7001 |
| 10024 | 337.48 | 6633 | 77.70 | 6016 | 29.54 | 3582 | 17.97 | 6583 |
| 11024 | 358.43 | 5722 | 81.15 | 5219 | 31.94 | 3170 | 18.02 | 5641 |
| 12024 | 375.95 | 5560 | 87.51 | 5072 | 33.35 | 2940 | 18.01 | 5294 |
| 13024 | 393.57 | 4691 | 94.74 | 4383 | OOM | 17.98 | 4621 | |
| 14024 | 403.52 | 4905 | 96.10 | 4546 | OOM | - | 18.01 | 4827 |
| 15024 | 431.66 | 4518 | 99.78 | 4170 | OOM | = | 17.96 | 4447 |
| 16024 | 453.32 | 4239 | 106.99 | 3878 | OOM | 17.98 | 4153 | |
| Inputs | Sinusoidal | Rotary | T5Bias | ALiBi | ||||
| PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | PPL (↓) | WPS (↑) | |
| 3072 | 18.67 | 13380 | 18.57 | 12548 | 18.01 | 8828 | 17.60 | 13866 |
| 3073 | 18.67 | 13773 | 18.57 | 12474 | 18.01 | 8483 | 17.59 | 13793 |
| 3082 | 18.62 | 13741 | 18.54 | 12388 | 17.95 | 8698 | 17.59 | 13778 |
| 3092 | 18.60 | 13742 | 18.48 | 12458 | 17.92 | 8361 | 17.55 | 13783 |
| 3102 | 18.65 | 13701 | 18.52 | 12365 | 17.94 | 8764 | 17.59 | 13747 |
| 3112 | 18.64 | 13809 | 18.51 | 12449 | 17.96 | 8665 | 17.59 | 13827 |
| 3122 | 18.68 | 13722 | 18.52 | 12432 | 17.98 | 8437 | 17.58 | 13795 |
| 3132 | 18.67 | 13825 | 18.54 | 12490 | 17.97 | 8653 | 17.58 | 13784 |
| 3142 | 18.69 | 13543 | 18.52 | 12230 | 17.97 | 8282 | 17.61 | 13572 |
| 3152 | 18.66 | 13520 | 18.56 | 12240 | 17.98 | 8608 | 17.59 | 13523 |
| 3162 | 18.71 | 13501 | 18.56 | 12253 | 18.04 | 8589 | 17.62 | 13598 |
| 3172 | 18.72 | 13563 | 18.55 | 12297 | 17.99 | 8583 | 17.59 | 13625 |
| 3272 | 18.87 | 13453 | 18.55 | 12148 | 17.93 | 8144 | 17.59 | 13482 |
| 3372 | 19.46 | 13533 | 18.50 | 12254 | 17.88 | 8442 | 17.52 | 13565 |
| 3472 | 20.55 | 13047 | 18.52 | 11868 | 17.95 | 7857 | 17.54 | 13107 |
| 3572 | 21.84 | 13128 | 18.50 | 11882 | 17.86 | 7814 | 17.50 | 13170 |
| 3672 | 23.04 | 13106 | 18.49 | 11859 | 17.87 | 7719 | 17.48 | 13196 |
| 3772 | 24.47 | 13287 | 18.54 | 11942 | 17.85 | 7579 | 17.49 | 13312 |
| 3872 | 25.85 | 12621 | 18.40 | 11272 | 17.82 | 7581 | 17.41 | 12566 |
| 3972 | 27.21 | 12379 | 18.48 | 11151 | 17.84 | 7483 | 17.41 | 12324 |
| 4072 | 28.59 | 12178 | 18.59 | 11019 | 17.88 | 6974 | 17.48 | 12212 |
| 5072 | 45.53 | 11076 | 18.80 | 9887 | 17.76 | 6230 | 17.33 | 10938 |
| 6072 | 65.01 | 10114 | 19.50 | 9049 | 17.68 | 5554 | 17.26 | 10133 |
| 7072 | 85.96 | 8647 | 20.60 | 7861 | 17.83 | 4820 | 17.22 | 8670 |
| 8072 | 102.74 | 7755 | 21.60 | 6991 | 18.06 | 4281 | 17.30 | 7729 |
| 9072 | 125.99 | 6953 | 22.14 | 6360 | 18.12 | 3823 | 17.26 | 6939 |
| 10072 | 133.68 | 6646 | 23.21 | 6068 | 18.37 | 3579 | 17.28 | 6597 |
| 11072 | 161.29 | 5663 | 24.39 | 5158 | 18.64 | 3119 | 17.26 | 5585 |
| 12072 | 169.55 | 5567 | 26.70 | 5111 | 18.93 | 2920 | 17.24 | 5397 |
| 13072 | 189.43 | 5044 | 29.33 | 4658 | 19.10 | 2735 | 17.15 | 4809 |
| 14072 | 203.86 | 4915 | 32.21 | 4616 | OOM | 17.22 | 4866 | |
| 15072 | 221.14 | 4561 | 33.47 | 4292 | OOM | - = | 17.23 | 4491 |
| 16072 | 231.29 | 4382 | 34.51 | 4099 | OOM | = | 17.22 | 4312 |
| ALiBi | Evaluation Length | |||||||
| Train Length | 64 | 128 | 256 | 512 | 1024 | 1536 | 2048 | 3072 |
| 64 | 28.46 | 24.70 | 22.88 | 22.09 | 21.73 | 21.63 | 21.59 | 21.53 |
| 128 | 1 | 23.98 | 21.70 | 20.67 | 20.36 | 20.29 | 20.31 | 20.28 |
| 256 | 1 | 1 | 21.29 | 19.89 | 19.29 | 19.13 | 19.10 | 19.03 |
| 512 | - | - | 19.73 | 18.81 | 18.50 | 18.48 | 18.40 | |
| 1024 | = | - | 18.66 | 18.20 | 18.05 | 17.96 | ||
| 1536 | = | = | = | 18.12 | 17.90 | 17.72 | ||
| 2048 | = | = | = | 17.91 | 17.64 | |||
| 3072 | 1 | = | = | = | = | - | 1 | 17.60 |
| Sinusoidal* | 28.03 | 23.81 | 21.45 | 20.05 | 19.34 | 19.05 | 18.87 | 18.67 |
| Rotary* | 1 | 1 | - | 20.07 | 19.33 | 1 | 1 | 18.57 |
| T5 Bias* | = | = | 1 | 19.65 | 18.80 | 1 | = | 18.01 |
| Model | Param.↓ | Train | Inference | ||
| Speed↑ | Speed ↑ | Valid ↓ | Test↓ | ||
| Sinusoidal, L = 3072 | 247M | 15.3k | 13.6k | 18.67 | 19.38 |
| Rotary,L = 3072 | 247M | 11.5k | 12.2k | 18.57 | 19.28 |
| T5 Bias,L = 3072 | 247M | 4.3k | 7.3k | 18.01 | 18.73 |
| L = 512,Lvalid = 3072 | 247M | 28.3k | 13.6k | 18.40 | 19.08 |
| A L=3072,Lvalid =3072 | 247M | 15.5k | 13.6k | 17.60 | 18.30 |
| Model | Param.↓ | Valid↓ | Test↓ |
| Adaptive Inputs (Baevski & Auli, 2018) | 247M | 17.97 | 18.70 |
| Transformer-XL (Dai et al., 2019) | 257M | 1 | 18.3 |
| Shortformer (Press et al.,2021) | 247M | 17.47 | 18.15 |
| Sandwich Transformer (Press et al., 2020) | 247M | 1 | 17.96 |
| Staged Training (Press et al.,2021) | 247M | 17.56 | |
| Compressive Transformer (Rae et al., 2020) | 329M | 17.1 | |
| Routing Transformer (Roy et al., 2020) | = | 15.8 | |
| kNN-LM (Khandelwal et al., 2020) | 247M | 15.81 | 15.79 |
| Sinusoidal, L = 3072 | 247M | 17.95 | 18.67 |
| Rotary,L = 3072 | 247M | 17.98 | 18.72 |
| T5 Bias,L = 3072 | 247M | 17.37 | 18.12 |
| L = 512,Lyalid = 3072 | 247M | 18.30 | 19.01 |
| L = 3072,Lvalid = 3072 | 247M | 16.97 | 17.66 |
| Train Length 512 | Evaluation Length 1024 3072 | ||
| 512 | 14.29 | 13.64 | 13.55 |
| 1024 | 1 | 13.86 | 13.52 |
| 3072 | 1 | 1 | 13.15 |
| Sinusoidal* | 14.80 | 14.73 | 14.46 |
| Model | Param.↓ | Valid ↓ | Test↓ |
| Sinusoidal,L = 3072 | 247M | 14.46 | 11.67 |
| B Ltrain =512,Lvalid =3072 | 247M | 13.55 | 10.98 |
| A Ltrain =3072,Lvalid=3072 | 247M | 13.15 | 10.73 |
| Model | Param.↓ | Valid ↓ | Test↓ |
| kNN-LM (Khandelwal et al., 2020) | 247M | 14.20 | 10.89 |
| Shortformer (Press et al.,2021) | 247M | 13.40 | 10.88 |
| Sandwich (Press et al., 2020) | 247M | = | 10.83 |
| Staged Training (Press et al., 2021) | 247M | 12.80 | 10.48 |
| Sinusoidal,L = 3072 | 247M | 14.06 | 11.40 |
| L =512,Lyalid = 3072 | 247M | 13.76 | 11.11 |
| L= 3072,Lalid = 3072 | 247M | 12.70 | 10.40 |
| Training | Valid PPL↓ | ||||
| Memory↓ | Updates | Hours↓ | Lvalid = 1024 | Lvalid = 2048 | |
| Sinusoidal, Ltrain = 1024 | 26.2 GB | 46.7k | 5.5k | 9.24 | 1 |
| ALiBi, Ltrain = 512 | 24.6 GB | 50.0k | 5.5k | 9.30 | 1 |
| Sinusoidal, Ltrain = 2048 | 29.3 GB | 44.2k | 5.9k | 1 | 9.01 |
| ALiBi,Ltrain = 1024 | 26.2 GB | 50.0k | 5.9k | 8.92 | |
| Training | Valid PPL ↓ | |||||
| Memory↓ | Updates | Hours← | Lvalid =512 Lvalid =1024 Lvalid=2048 | |||
| Sinusoidal, Ltrain = 512 ALiBi,Ltrain = 512 | 24.6 GB | 50.0k | 5.5k | 9.71 | 37.05 | 105.42 |
| 24.6 GB | 50.0k | 5.5k | 9.79 | 9.30 | 9.54 | |
| Sinusoidal, Ltrain = 1024 ALiBi,Ltrain = 1024 | 26.2 GB | 50.0k | 5.9k | 1 | 9.15 | 48.85 |
| 26.2 GB | 50.0k | 5.9k | = | 9.16 | 8.92 | |
| Sinusoidal, Ltrain = 2048 ALiBi,Ltrain = 2048 | 29.3 GB | 50.0k | 6.7k | - | 8.83 | |
| 29.4 GB | 50.0k | 6.7k | 1 | 8.84 | ||
| Evaluation Length (S = 1) | 3072 | ||||
| Train Length | 512 | 1024 | 1536 | 2048 | |
| 512 | 18.35 | 204.42 | 264.74 | 306.19 | 360.12 |
| 1024 | 1 | 18.05 | 206.55 | 302.6 | 393.71 |
| 3072 | - | 1 | 1 | 1 | 18.03 |
| Evaluation Length (S = 1) | |||||
| Train Length | 512 | 1024 | 1536 | 2048 | 3072 |
| 512 | 17.92 | 18.51 | 20.36 | 22.62 | 30.77 |
| 1024 | 1 | 17.65 | 17.87 | 18.51 | 20.66 |
| 3072 | 1 | 1 | 1 | 1 | 17.41 |
| Evaluation Length (S = 1) | |||||
| Train Length | 512 | 1024 | 1536 | 2048 | 3072 |
| 512 | 17.98 | 17.92 | 18.2 | 18.28 | 18.3 |
| 1024 | 1 | 17.46 | 17.47 | 17.62 | 17.92 |
| 3072 | 1 | 1 | 1 | 1 | 16.96 |
| Type | Argument | State -→ Next State |
| search | [Query] | Search-Results |
| choose | Back to search | *→Search |
| choose | Prev/Next page | Results→Results |
| choose | [Product title] | Results →Item |
| choose | [Option] | Item →Item |
| choose | Desc/Overview | Item -→Item-Detail |
| choose | Previous | Item-Detail →Item |
| choose | Buy | Item →Episode End |
| Score | Count | |||||||
| All | Att | Opt | Type | Price | State | Item | Search | |
| Rule | 45.6 | 66.6 | 0.0 | 80.5 | 86.0 | 3.0 (3/3) | 1.0 (1/1) | 1.0 (1/1) |
| IL | 59.9 | 69.3 | 45.2 | 86.4 | 84.0 | 9.4 (90/3) | 1.6 (11/1) | 1.3 (17/1) |
| IL+RL | 62.4 | 74.0 | 38.9 | 89.7 | 88.7 | 4.5 (5/1) | 1.0 (1/1) | 1.0(1/1) |
| Human Expert | 82.1 | 81.8 | 73.9 | 94.4 | 97.7 | 11.3 (114 /4) | 1.9 (16 / 1) | 1.4 (16 /1) |
| Instruction 1 | Instruction 2 | |
| I want to find white blackout shades that are 66inches in width and 66 inches in height. They needto be easy to install [...] | I need a gingko light and 2O"x2O" pillow cover that is hand painted [.] | |
| Human Actions (r = 1.O,length = 8)search[66 inches in width and 66 inches in heightwhite shades ]choose[ item:CALYX... ]choose[Back to Search]search[66 x 66 blackout shades ]choose[item:Milin...]choose[opt : 66"w x 66"h]choose[opt : cordless bottom up-blackout-white ]choose[Buy] | Human Actions (r = 1.O,length = 8)search[66 inches in width and 66 inches in heightwhite shades ]choose[ item:CALYX... ]choose[Back to Search]search[66 x 66 blackout shades ]choose[item:Milin...]choose[opt : 66"w x 66"h]choose[opt : cordless bottom up-blackout-white ] | |
| IL+RL Actions (r = O.2,length = 3)search[white blackout shades 65 inches in widthand 66 inches in height]choose[item:Window...]choose[Buy] | IL+RL Actions (r = O.25,length = 3)search[ gingko light and 2Ox2O pillow cover handpainted ]choose[ item :UPOOS..]choose[Buy] | |
| Instr. text | IL BART | Human expert (first) | Human expert (last) | |
| Score | 79.7 | 83.0 | 82.1 | 84.4 |
| Success Rate | 52.6% | 57.6% | 57.9% | 61.0% |
| Amazon | eBay | |||||||||
| Score /SR | Att | Opt | Type | Price | Score /SR | Att | Opt | Type | Price | |
| Rule | 45.8 /19% | 45.6 | 38.0 | 66.2 | 90.0 | 31.7/ 7% | 62.3 | 25.9 | 49.0 | 67.0 |
| IL | 61.5 / 27% | 60.7 | 53.7 | 85.6 | 96.0 | 58.2/21% | 60.2 | 52.3 | 85.1 | 96.9 |
| IL+RL | 65.9 / 25% | 71.6 | 47.0 | 87.8 | 100.0 | 62.3 / 21% | 69.1 | 39.5 | 91.7 | 97.0 |
| Human | 88.2 / 65% | 86.2 | 76.3 | 99.0 | 100.0 | 79.7 /40% | 80.3 | 70.1 | 99.5 | 100.0 |
| Method | Training | Inference | ||||
| Learnable Params (M) | # GPU | Batch Size | Time (h) | FPS | Mem. | |
| SAM [21] | 1191 | 128 | 128 | N/A | 5.0 | 7.6G |
| HQ-SAM | 5.1 | 8 | 32 | 4 | 4.8 | 7.6G |
| Model | DIS [35] | COIFT [29] | HRSOD [51] | ThinObject [29] | Average | |||||
| mIoU | mBIoU | mIoU | mBIoU | mIoU | mBIoU | mIoU | mBIoU | mIoU | mBIoU | |
| SAM (baseline) | 62.0 | 52.8 | 92.1 | 86.5 | 90.2 | 83.1 | 73.6 | 61.8 | 79.5 | 71.1 |
| Using SAM's mask decoder feature: | ||||||||||
| SAM+Context Token [56] | 71.5 | 62.2 | 93.0 | 87.7 | 91.8 | 85.0 | 84.5 | 73.1 | 85.2 | 77.0 |
| SAM + HQ-Output Token (× Output Token) | 75.1 | 65.8 66.4 | 93.9 | 88.9 | 93.0 | 86.1 | 86.1 | 74.6 | 87.0 | 78.9 |
| SAM + HQ-Output Token (Boundary Loss) SAM + HQ-Output Token | 75.2 75.3 | 66.0 | 94.0 94.2 | 88.9 89.2 | 92.1 93.0 | 85.7 86.1 | 87.3 86.8 | 76.0 75.4 | 87.2 87.3 | 79.3 |
| 79.2 | ||||||||||
| Using Our HQ-Feature: | ||||||||||
| SAM + HQ-Output Token (+ Context Token) | 78.5 | 70.4 | 94.6 | 89.6 | 93.6 | 87.0 | 88.9 89.5 | 79.3 | 88.9 89.1 | 81.6 |
| SAM+ HQ-Output Token | 78.6 | 70.4 | 94.8 | 90.1 | 93.6 | 86.9 | 79.9 | 81.8 | ||
| Model | Fusion conv | Decoder Mask feature | ViT Encoder Final-layer Early-layer | mIoU | Four HQ datasets mBIoU |
| SAM [21] | √ | 79.5 | 71.1 | ||
| HQ-SAM (Ours) | 广 | √ | 87.3 79.2 | ||
| √ | 87.8 | 80.1 | |||
| 15.1 | 9.0 | ||||
| √ √ | 广 | 88.6 81.3 | |||
| √ √ | √ √ √ 丁 | 88.6 89.1 | 81.1 81.8 |
| Model | Four HQ datasets mIoU mBIoU | CoCo | |||||
| APB | AP | APL | APm | APs | |||
| SAM (baseline) | 79.5 | 71.1 | 33.3 | 48.5 | 63.9 | 53.1 | 34.1 |
| Training the whole SAM | 38.0 | 12.2 | 0.2 | 5.5 | 1 | - | 1 |
| Add Context Token [56] | 85.2 | 77.0 | 31.9 | 47.2 | 65.1 | 51.2 | 31.9 |
| CascadePSP Post-refinement [6] | 80.9 | 74.6 | 2.8 | 13.4 | 43.4 | 9.4 | 0.0 |
| CRM Post-refinement [37] | 81.4 | 75.4 | 15.9 | 28.7 | = | - | - |
| Finetune SAM's decoder | 87.6 | 79.5 | 9.0 | 19.5 | 45.2 | 15.8 | 4.7 |
| Finetune SAM's output token | 87.6 | 79.7 | 33.7 | 48.7 | 66.0 | 52.3 | 33.6 |
| HQ-SAM (Ours) | 89.1 | 81.8 | 34.4 | 49.5 | 66.2 | 53.8 | 33.9 |
| Model | APsict | AP | AP6 | APB | APB75 | APB50 | AP |
| SAM | 8.6 | 3.7 | 25.6 | 17.3 | 14.4 | 37.7 | 29.7 |
| HQ-SAM | 9.9 | 5.0 | 28.2 | 18.5 | 16.3 | 38.6 | 30.1 |
| Model | GT Box Prompt mIoU | mBIoU | Mask Prompt mIoU | mBIoU |
| SAM | 81.1 | 70.4 | 66.6 | 41.8 |
| HQ-SAM | 86.0 | 75.3 | 86.9 | 75.1 |
| Model | COCO | LVIS | |||||
| APB | AP | APsiet | AP | APB | APB75 | AP | |
| SAM | 33.3 | 48.5 | 32.1 | 32.8 | 38.5 | 40.9 | 43.6 |
| HQ-SAM | 34.4 | 49.5 | 32.5 | 33.5 | 38.8 | 41.2 | 43.9 |
| Model | APB | AP | AP5 | APM | AP | AP |
| SAM | 30.2 | 19.1 | 72.9 | 60.7 | 68.1 | 90.5 |
| HQ-SAM | 34.0 | 24.3 | 79.5 | 63.6 | 70.5 | 91.1 |
| Model | CoCo | Model Params (MB) | |||||
| APB | AP | APL | APM | APs | Total | Trainable | |
| SAM | 33.3 | 48.5 | 63.9 | 53.1 | 34.1 | 1191 | 1 |
| SAM+LoRA[17] | 28.6 | 43.7 | - | - | - | 1192.5 | 1.5 |
| SAM + Encoder Adapter [48] | 29.6 | 44.8 | 63.9 | 47.8 | 29.0 | 1203 | 12.0 |
| HQ-SAM | 34.4 | 49.5 | 66.2 | 53.8 | 33.9 | 1196.1 | 5.1 |
| Model | Four HQ datasets | CoCo | Model Params (MB) | FPS | Memory | ||||||
| mIoU | mBIoU | APB | AP | APL | APM | APs | Total | Learnable | |||
| SAM-B HQ-SAM-B | 70.6 86.3 | 62.3 78.1 | 28.2 31.3 | 44.4 46.7 | 57.7 62.9 | 48.7 50.5 | 32.1 32.0 | 358 362.1 | 358 4.1 | 10.1 9.8 | 5.1G 5.1G |
| SAM-L | 79.5 | 71.1 | 33.3 | 48.5 | 63.9 | 53.1 | 34.1 | 1191 | 1191 | 5.0 | 7.6G |
| HQ-SAM-L SAM-H | 89.1 75.6 | 81.8 68.3 | 34.4 | 49.5 | 66.2 | 53.8 | 33.9 | 1196.1 | 5.1 2446 | 4.8 3.5 | 7.6G 10.3G |
| HQ-SAM-H | 89.3 | 81.5 | 34.0 34.9 | 48.9 49.9 | 64.5 | 53.3 | 34.4 | 2446 2452.1 | 6.1 | 3.4 | 10.3G |
| 66.5 | 54.0 | 34.2 | |||||||||
| MobileSAM | |||||||||||
| 69.0 | 58.8 | 28.6 | 44.3 | - | - | : | 38.6 | 38.6 | 44.8 | 3.7G | |
| Light HQ-SAM | 81.4 | 71.6 | 29.6 | 45.0 | - | - | 40.3 | 1.7 | 41.2 | 3.7G | |
| Model | YTVIS 2019 | HQ-YTVIS | ||||||
| AP | AP50 | AP75 | APL | APm | APs | APB | APM | |
| SAM | 51.8 | 82.1 | 55.4 | 65.5 | 52.0 | 34.2 | 30.2 | 60.7 |
| HQ-SAM | 53.2 | 82.9 | 58.3 | 66.4 | 53.3 | 33.7 | 34.0 | 63.6 |
| Model | J&F | J | F |
| SAM | 82.0 | 79.0 | 84.9 |
| HQ-SAM | 83.2 | 80.3 | 86.1 |
| Model | No Noise mIoU mBIoU | Noise scale 0.2 mIoU mBIoU | Noise scale 0.4 mIoU mBIoU | |||
| SAM | 79.5 | 71.1 | 65.7 | 57.1 | 46.4 | 39.8 |
| HQ-SAM | 89.1 | 81.8个10.7 | 82.8 | 73.4个16.3 | 69.9 | 60.3个20.5 |
| Dataset | DIS [35] | Thin-Object 5k [29] | FSS [26] | DUTS [46] | ECSSD [38] | MSRA-10K [8] | Total |
| Image Num. | 3000 | 4748 | 10000 | 15572 | 1000 | 10000 | 44320 |
| Model | Dataset | DIS | COIFT | HRSOD | ThinObject mBIoU | Average | |||||
| mIoU | mBIoU | mIoU | mBIoU | mIoU | mBIoU | mIoU | mIoU | mBIoU | |||
| SAM | SA-1B | 62.0 | 52.8 | 92.1 | 86.5 | 90.2 | 83.1 | 73.6 | 61.8 | 79.5 | 71.1 |
| HQ-SAM | + SA-1B-44K | 60.4 | 51.7 | 91.1 | 86.1 | 88.4 | 80.9 | 73.1 | 61.8 | 78.3 | 70.1 |
| HQ-SAM | + HQ-Seg-44K(Ours) | 78.6 | 70.4 | 94.8 | 90.1 | 93.6 | 86.9 | 89.5 | 79.9 | 89.1 | 81.8 |
| Training Setting | DIS-mIoU | DIS-mBIoU | ThinObject-mloU | ThinObject-mBIoU |
| SAM (baseline) | 62.0 | 52.8 | 73.6 | 61.8 |
| HQ-SAM (remove both DIS and ThinObject) | 72.9 | 63.1 | 82.7 | 70.7 |
| HQ-SAM (remove DIS) | 74.7 | 66.2 | (90.1) | (80.4) |
| HQ-SAM (remove ThinObject) | (78.4) | (70.3) | 83.3 | 72.1 |
| HQ-SAM (default HQSeg-44K) | (78.6) | (70.4) | (89.5) | (79.9) |
| l2Radius (CIFAR-10) | 0.25 | 0.5 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 2.0 | ACR |
| Baseline | 6.96 | 2.04 | 0.09 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.026 |
| Randg =0.5 (Cohen et al.,2019) | 51.68 | 40.38 | 30.25 | 20.81 | 13.36 | 7.71 | 3.38 | 0.0 | 0.488 |
| (Ours)Randg=0.5 +adaptation | 62.91 | 52.25 | 40.06 | 25.57 | 17.43 | 10.67 | 5.46 | 1.92 | 0.657 |
| SmoothAdvg=0.5 (Salman et al.,2019a) | 58.82 | 49.68 | 42.68 | 37.55 | 32.64 | 27.52 | 22.42 | 0.0 | 0.918 |
| (Ours) SmoothAdvg=0.5+adaptation | 59.89 | 50.4 | 41.76 | 35.5 | 30.92 | 26.1 | 20.25 | 15.05 | 1.008 |
| Advo (Rice et al.,2020) | 35.95 | 29.44 | 23.51 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.317 |
| (Ours)Advo + adaptation | 67.96 | 55.06 | 43.27 | 30.55 | 24.68 | 18.49 | 12.11 | 8.45 | 0.903 |
| Adv2 (Rice et al., 2020) | 41.89 | 34.15 | 26.7 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.359 |
| (Ours) Adv2 +adaptation | 68.84 | 58.77 | 49.71 | 37.74 | 33.37 | 28.82 | 23.65 | 18.23 | 1.198 |
| MARCERg=0.5 (Zhai et al.,2020) | 60.0 | 53.0 | 46.0 | 38.0 | 29.0 | 19.0 | 12.0 | 0.0 | 0.726 |
| Consistancyg=0.5 (Jeong& Shin,2020) | 48.9 | 45.1 | 41.3 | 37.8 | 33.9 | 29.9 | 25.2 | 0.0 | 0.726 |
| (a) IMAGENET | ||||
| Model | σ=0 | σ=0.25 | σ=0.5 | σ=0.75 |
| Baseline | 75.2±0.0 | 11.8±0.22 | 0.3±0.01 | 0.1±0.0 |
| + adaptive BN | 74.4±0.04 | 31.0±0.27 | 7.7±0.24 | 2.4±0.01 |
| Advo∞≤4/255] | 62.8±0.0 | 3.9±0.03 | 0.4±0.01 | 0.2±0.01 |
| + adaptive BN | 60.8±0.16 | 53.4±0.15 | 44.9±0.08 | 33.7±0.28 |
| Adv2≤3] | 59.8±0.0 | 9.8±0.08 | 0.9±0.01 | 0.3±0.0 |
| +adaptive BN | 58.3±0.08 | 53.7±0.14 | 47.3±0.14 | 39.8±0.18 |
| Rand g=0.5 | 22.0±0.0 | 32.8±0.11 | 60.9±0.04 | 0.9±0.06 |
| + adaptive BN | 62.7±0.03 | 62.3±0.18 | 59.5±0.11 | 51.4±0.27 |
| (b) CIFAR-10 | ||||
| Model | σ=0 | g=0.25 | σ=0.5 | g=0.75 |
| Baseline + adaptive BN | 95.2±0.0 | 10.9±0.88 | 10.6±0.76 | 10.5±1.19 |
| 95.0±0.57 | 40.1±0.97 | 22.0±0.83 | 17.2±0.66 | |
| Advo≤8/255] | 82.1±0.0 | 40.2±4.56 | 16.1±7.85 | 12.2±5.23 |
| + adaptive BN | 81.6±0.96 | 74.2±0.95 | 62.4±0.64 | 51.0±1.03 |
| Adv2[≤1] | 81.6±0.0 | 47.5±5.1 | 21.5±7.79 | 14.3±5.63 |
| + adaptive BN | 81.8±0.7 | 75.8±0.43 | 64.9±0.73 | 53.5±1.71 |
| Rand g=0.5 | 66.7±0.0 | 69.1±1.01 | 61.2±0.84 | 25.9±1.41 |
| + adaptive BN | 74.0±2.1 | 73.0±2.04 | 66.8±2.01 | 56.7±0.94 |
| (b) CIFAR-10 | ||
| p | Advo | Adv2 |
| 0.0 (No adaptation) | 16.1±7.85 | 21.5±7.79 |
| 0.1 | 45.1±0.49 | 46.9±0.48 |
| 0.3 | 59.2±0.42 | 60.8±0.33 |
| 0.5 | 62.4±0.27 | 64.4±0.6 |
| 0.7 | 62.8±0.52 | 64.9±0.31 |
| 0.9 | 62.8±0.71 | 64.9±0.31 |
| 1.0 (Full adaptation) | 62.4±0.64 | 64.9±0.73 |
| (a)IMAGENET | ||
| p | Advo | Adv2 |
| 0.0 (No adaptation) | 0.4±0.01 | 0.9±0.01 |
| 0.1 | 2.1±0.04 | 7.7±0.09 |
| 0.3 | 20.6±0.16 | 36.6±0.09 |
| 0.5 | 41.1±0.09 | 45.5±0.13 |
| 0.7 | 43.5±0.14 | 46.7±0.13 |
| 0.9 | 44.2±0.12 | 46.8±0.13 |
| 1.0 (Full adaptation) | 44.8±0.13 | 47.2±0.14 |
| (b) CIFAR-10 | ||
| Batch Size | Advo | Adv2 |
| w/o BNadapt | 16.1±7.85 | 21.5±7.79 |
| 8 | 57.2±1.23 | 59.5±0.38 |
| 16 | 60.2±0.79 | 62.3±0.87 |
| 32 | 61.5±0.46 | 63.6±0.55 |
| 64 | 62.3±0.5 | 64.0±0.38 |
| 128 | 62.7±0.68 | 64.4±0.53 |
| 256 | 62.7±0.68 | 64.9±0.48 |
| 512 | 62.4±0.64 | 64.9±0.73 |
| (a)IMAGENET | ||
| Batch Size | Advo | Adv2 |
| w/oBNadapt | 0.4±0.01 | 0.9±0.01 |
| 8 | 11.5±0.22 | 9.1±0.15 |
| 16 | 28.1±0.22 | 26.7±0.14 |
| 32 | 37.1±0.24 | 37.6±0.2 |
| 64 | 41.4±0.26 | 42.9±0.12 |
| 128 | 43.3±0.15 | 45.4±0.13 |
| 256 | 44.4±0.21 | 46.7±0.07 |
| 512 | 44.8±0.13 | 47.2±0.14 |
| IMAGENET | |||||||||||||
| Model | BN adaption | Certification | 0.5 | l2 Radius 1.25 | 1.75 | ||||||||
| 0.25 | 0.75 | 1.0 | 1.5 | 2.0 | 2.25 | 2.5 | 2.75 | ACR | |||||
| Baseline | = | atσ=0.25 | 7.8 | 4.8 | 3.0 | 0.0 | 0.0 | 0.0 | 0.0 0.0 | 0.0 | 0.0 | 0.0 | 1 0.054 |
| Advo[l∞o ≤4/255] | at σ = 0.25 at σ = 0.50 | at g = 0.25 | 50.0 | 46.4 | 41.6 | 0.0 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.445 0.607 |
| at σ=0.75 | at σ = 0.50 at σ=0.75 | 43.6 31.6 | 39.4 | 35.8 | 31.4 27.6 | 23.4 | 18.2 | 0.0 | 0.0 | 0.0 | 0.0 | 0.443 | |
| 26.4 | 22.4 | 18.6 16.8 | 14.4 | 11.8 | 9.4 | 7.6 | 5.6 | 3.6 | |||||
| Advo[loo≤4/255]+adapt[BestRadij(Ours) | 50.0 | 46.4 | 41.6 | 31.4 27.6 | 23.4 | 18.2 | 9.4 | 7.6 | 5.6 | 3.6 | 0.759 | ||
| at σ = 0.25 | 0.480 | ||||||||||||
| Adv2[l2 ≤3.00] | at σ= 0.50 | at σ = 0.25 | 53.2 47.0 | 50.2 43.0 | 46.8 39.0 | 0.0 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.711 |
| at σ=0.75 | at g = 0.50 at g=0.75 | 37.8 | 32.2 | 36.4 32.8 | 30.8 20.2 | 27.0 | 0.0 | 0.0 14.2 | 0.0 12.0 | 0.0 9.6 | 0.639 | ||
| Adv2l2≤3.00] +adapt[Best Radii](Ours) | 53.2 | 28.4 | 26.0 22.4 32.8 | 30.8 | 19.0 27.0 | 17.4 17.4 | 14.2 | 12.0 | 9.6 | 0.930 | |||
| 50.2 | 46.8 | 36.4 | |||||||||||
| Randg=0.5 Cohen et al. (2019) | = | at σ=0.50 | 60.8 | 54.4 | 47.8 | 39.0 | 34.2 29.0 | 23.8 | 0.0 | 0.0 | 0.0 | 0.0 | 0.809 |
| at σ = 0.25 | at g = 0.25 | 59.8 | 53.6 | 46.6 | 0.0 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.507 | |
| + adaptation | at σ= 0.50 | at σ= 0.50 | 58.6 | 51.0 | 43.8 | 37.4 | 32.2 27.4 | 22.4 | 0.0 | 0.0 | 0.0 | 0.0 | 0.768 |
| at σ=0.75 | at σ=0.75 | 48.6 | 41.6 | 36.6 | 31.2 | 26.2 22.4 | 18.6 | 16.8 | 12.8 | 8.6 | 5.4 | 0.720 | |
| Randg=0.5+adapt[BestRadii](Ours) | 22.4 | 16.8 | |||||||||||
| 59.8 | 53.6 | 46.6 | 37.4 | 32.2 | 27.4 | 12.8 | 8.6 | 5.4 | 0.973 | ||||
| CIFAR-10 | |||||||||||
| Model | BN adaption | Certification | 0.25 | 0.5 | 0.75 | l2Radius 1.0 1.25 | 1.5 | 1.75 | 2.0 | ACR | |
| Baseline | 0.0 0.0 | 0.0 | 0.0 0.0 | 0.026 | |||||||
| Advo[∞ ≤ 4/255] | at σ = 0.25 at σ = 0.50 at σ = 0.75 | at σ = 0.25 at σ = 0.50 at σ = 0.75 | 67.96 47.34 26.89 | 50.46 31.83 15.92 | 31.96 18.78 8.44 | 0.0 9.98 4.31 | 0.0 4.44 2.03 | 0.0 1.62 0.28 0.79 0.23 | 0.0 0.0 0.0 | 0.08 | 0.485 0.350 0.146 |
| Advo[lo ≤8/255] | at σ =0.25 at σ = 0.50 at σ =0.75 | at g =0.25 at σ = 0.50 at σ =0.75 | 66.43 53.65 39.96 | 55.06 42.91 30.76 | 42.86 32.58 22.01 | 0.0 22.68 14.64 | 0.0 14.24 8.84 | 0.0 7.88 4.81 | 0.0 2.94 | 0.0 0.0 1.15 | 0.527 0.515 0.352 |
| Advoo[∞ ≤12/255] | atσ =0.25 at σ = 0.50 at σ = 0.75 | at σ= 0.25 at σ = 0.50 atσ=0.75 | 60.52 51.53 42.61 | 52.42 43.94 35.56 | 43.27 36.41 28.47 | 0.0 28.69 22.39 | 0.0 21.25 16.69 | 0.0 14.53 | 2.29 0.0 8.03 | 0.0 0.0 4.43 | 0.499 0.581 0.482 |
| Adv[l∞o ≤ 16/255] | at σ =0.25 at σ = 0.50 at σ =0.75 | atg=0.25 at σ = 0.50 atσ =0.75 | 53.75 48.07 42.05 67.96 | 47.57 42.51 36.42 55.06 | 41.18 36.54 31.24 43.27 | 0.0 30.55 26.05 30.55 | 0.0 24.68 20.74 | 11.58 0.0 18.49 16.15 12.01 | 7.42 0.0 12.11 | 0.0 0.0 8.45 | 0.454 0.598 0.557 0.903 |
| Advo + adapt [Best Radii] (Ours) 24.68 | 18.49 12.11 8.45 0.0 | ||||||||||
| Adv2[l2 ≤0.50] | at σ = 0.25 at σ = 0.50 atg =0.75 | at σ = 0.25 at σ = 0.50 atσ =0.75 | 68.84 48.81 27.38 | 54.04 33.82 16.15 | 37.13 20.95 9.23 | 0.0 11.5 4.56 | 0.0 5.64 2.06 | 0.0 2.29 0.62 0.91 0.33 | 0.0 0.0 0.08 | 0.518 0.382 0.153 | |
| Adv2[l2 ≤ 1.00] | at σ = 0.25 atσ =0.50 atσ =0.75 | at σ = 0.25 at σ = 0.50 atg =0.75 | 68.02 56.45 43.04 | 58.54 46.24 33.08 | 46.98 35.6 24.81 | 0.0 26.89 17.68 | 0.0 18.73 11.39 | 0.0 11.37 6.6 | 0.0 5.41 3.57 | 0.0 0.0 1.95 | 0.551 0.580 0.405 |
| Adv2[l2 ≤1.25] | at g =0.25 at σ =0.50 atσ =0.75 | at σ= 0.25 at σ = 0.50 at σ = 0.75 | 67.13 57.73 46.54 | 58.77 48.8 37.53 | 49.43 39.64 29.35 | 0.0 31.07 22.0 | 0.0 22.61 15.62 | 0.0 15.82 10.51 | 0.0 8.96 6.55 | 0.0 0.0 3.68 | 0.557 0.647 0.496 |
| Adv2[l2 ≤ 1.50] | at σ = 0.25 at σ =0.50 at σ=0.75 | at g = 0.25 at σ = 0.50 atg =0.75 | 64.21 56.55 47.73 | 57.13 49.19 40.89 | 49.71 41.72 33.78 | 0.0 34.47 27.22 | 0.0 27.36 20.78 | 0.0 20.23 15.19 | 0.0 12.98 10.51 | 0.0 0.0 6.77 | 0.543 0.689 0.585 |
| Adv2[l2 ≤ 2.00] | at σ= 0.25 at σ =0.50 at σ =0.75 | at σ = 0.25 at σ = 0.50 at σ = 0.75 | 60.4 54.27 47.96 | 54.71 48.89 42.54 | 48.35 43.1 37.04 | 0.0 37.34 31.65 | 0.0 31.52 26.17 | 0.0 25.74 21.12 | 0.0 19.14 16.59 | 0.0 0.0 12.44 | 0.523 0.731 0.698 |
| Adv2[l2 ≤ 2.25] | at σ = 0.25 at σ = 0.50 at σ = 0.75 | atg=0.25 at σ = 0.50 at σ = 0.75 | 57.08 52.1 46.45 | 52.5 46.99 41.71 | 47.11 42.26 36.75 | 0.0 36.9 31.88 | 0.0 31.58 26.95 | 0.0 26.08 22.33 | 0.0 20.03 17.82 | 0.0 0.0 13.55 | 0.504 0.724 0.713 |
| Adv2[l2 ≤ 2.50] | at σ =0.25 at σ = 0.50 atσ=0.75 | at σ = 0.25 at σ = 0.50 at σ =0.75 | 54.88 50.53 45.95 | 50.79 46.26 41.89 | 46.29 41.84 37.53 | 0.0 37.74 33.55 | 0.0 33.2 29.31 | 0.0 28.69 25.27 | 0.0 23.34 20.98 | 0.0 0.0 17.28 | 0.487 0.734 0.765 |
| Adv2[2 ≤ 3.00] | at σ =0.25 atσ =0.50 at σ = 0.75 | at σ =0.25 at σ = 0.50 atσ =0.75 | 53.82 49.41 45.37 | 49.69 45.57 41.54 | 45.04 41.52 37.75 | 0.0 37.43 33.49 | 0.0 33.37 29.35 | 0.0 28.82 25.62 | 0.0 23.65 21.83 | 0.0 0.0 18.23 | 0.475 0.720 0.771 1.198 |
| Adv2+adapt[Best Radii](Ours) 68.84 58.77 | |||||||||||
| Randg=0.5 | at g= 0.50 at σ = 0.25 | 51.68 62.91 | 40.38 52.25 | 49.71 30.25 40.06 | 37.74 20.81 0.0 | 33.37 13.36 0.0 | 28.82 7.71 0.0 | 23.65 3.38 0.0 | 18.23 0.0 0.0 | 0.488 0.497 0.575 | |
| at σ = 0.25 + adaptation at σ =0.50 at σ = 0.50 | |||||||||||
| Randg=0.5 +adapt[Best Radi] (Ours) | atσ =0.75 | at σ =0.75 | 57.58 46.4 62.91 | 46.46 35.63 52.25 | 35.5 26.06 40.06 | 25.57 18.17 25.57 | 17.43 11.61 17.43 | 10.67 5.46 6.86 3.64 10.67 5.46 | 0.0 1.92 1.92 | 0.427 0.657 0.609 | |
| SmoothAdvg=0.5,=0.25 at g= 0.50 | |||||||||||
| + adaptation | at σ = 0.25 at σ = 0.50 | at σ = 0.25 at g = 0.50 | 57.8 58.74 54.0 | 47.63 48.29 42.91 | 37.41 36.7 32.63 | 27.88 0.0 23.6 | 20.33 13.53 0.0 16.08 | 8.03 0.0 0.0 9.93 5.5 | 0.0 0.0 0.0 | 0.464 0.535 0.390 | |
| SmoothAdvg=0.5,∈=0.50 | at σ = 0.75 at σ = 0.25 | at σ = 0.75 atσ =0.50 at σ = 0.25 | 43.15 58.82 59.89 | 32.29 49.68 50.4 | 23.61 40.35 39.99 | 16.41 31.93 0.0 | 10.94 24.18 0.0 | 6.65 17.05 0.0 | 3.88 10.57 0.0 | 2.16 0.0 0.0 | 0.661 0.483 0.592 |
| + adaptation | |||||||||||
| SmoothAdvg=0.5,∈=1.0 | at σ = 0.50 at σ = 0.75 | at σ = 0.50 at σ =0.75 atg=0.50 | 55.73 46.25 56.53 | 45.79 36.72 49.53 | 36.6 28.2 41.38 | 27.4 20.9 34.63 | 20.03 14.73 27.81 | 13.48 9.46 21.22 | 7.85 5.78 14.41 | 0.0 3.32 0.0 | 0.470 0.691 0.467 |
| + adaptation | - at σ = 0.25 at σ = 0.50 | at σ = 0.25 at σ = 0.50 at σ = 0.75 | 57.13 53.56 47.17 | 48.38 45.79 39.4 | 39.7 37.64 | 0.0 30.06 | 0.0 23.12 18.93 | 0.0 17.27 13.7 | 0.0 11.14 9.26 | 0.0 0.0 5.8 | 0.620 0.545 |
| at σ = 0.75 | |||||||||||
| SmoothAdvg=0.5,∈=2.0 | at σ = 0.25 at σ =0.50 | atg =0.50 at σ = 0.25 | 52.82 52.23 | 47.67 47.24 | 31.8 42.68 41.76 | 24.74 37.55 0.0 | 32.64 0.0 | 27.52 0.0 | 22.42 0.0 | 0.0 0.0 | 0.732 0.451 0.692 |
| + adaptation at σ =0.75 | at σ = 0.50 atσ =0.75 | 50.28 46.7 | 45.24 41.79 | 40.39 37.14 | 35.5 33.05 | 30.92 28.28 | 26.1 23.88 | 20.25 19.34 | 0.0 15.05 | 0.727 | |
| SmoothAdvg=0.5[BestRadii] | 58.82 59.89 | 49.68 50.4 | 42.68 41.76 | 37.55 35.5 | 32.64 30.92 | 27.52 26.1 | 22.42 20.25 | 0.0 15.05 | 0.918 1.008 | ||
| SmoothAdvg=0.5 + adapt [Best Radii] (Ours) MARCERg=0.5 (Zhai et al.,2020) 60.0 53.0 | |||||||||||
| Method | DMF | D7P | MSK | PH2 | SON | UDA | AVG |
| DeepAll | 0.2492 ±0.0127 | 0.5680±0.0181 | 0.6674±0.0083 | 0.8000±0.0167 | 0.8613±0.0296 | 0.6264±0.0312 | 0.6287 |
| MASF Dou et al. (2019) | 0.2692±0.0146 | 0.5678±0.0361 | 0.6815±0.0122 | 0.7833±0.0101 | 0.9204±0.0227 | 0.6538±0.0196 | 0.6460 |
| LDDG Li et al. (2020a) | 0.2793±0.0244 | 0.6007±0.0187 | 0.6967±0.0211 | 0.8167±0.0209 | 0.9272±0.0117 | 0.6978±0.0182 | 0.6697 |
| SWAD Cha et al.(2021) | 0.3582 ±0.0234 | 0.5491 ±0.0231 | 0.6842 ±0.0156 | 0.9167 ±0.0121 | 0.9824 ±0.0012 | 0.7240 ±0.0251 | 0.7024 |
| BDIL Xiao et al. (2021) | 0.2985±0.0452 | 0.6204±0.0212 | 0.7059±0.0145 | 0.8967±0.0096 | 0.9860±0.0198 | 0.7219±0.0284 | 0.7049 |
| DNA Chu et al. (2022) | 0.3532 ±0.0133 | 0.5581 ±0.0178 | 0.7120 ±0.0194 | 0.9333 ±0.0045 | 0.9851 ±0.0032 | 0.7314 ±0.0141 | 0.7122 |
| DSU Li et al. (2022c) | 0.3830 ±0.0267 | 0.5739 ±0.0147 | 0.6935 ±0.0165 | 0.8833 ±0.0231 | 0.9841 ±0.0098 | 0.7201 ±0.0121 | 0.7063 |
| Ours | 0.3781±0.0136 | 0.6120±0.0115 | 0.7276 ±0.0201 | 0.9416±0.0103 | 0.9889±0.0041 | 0.7486 ±0.0123 | 0.7328 |
| Method | IHC | NKI | VGH | Average (%) |
| DeepAll | 73.29 ±0.13 | 70.60 ±0.15 | 79.56±0.11 | 74.48 |
| MASF Dou et al. (2019) | 80.45±0.10 | 76.10±0.11 | 84.44±0.12 | 80.33 |
| SWAD Cha et al. (2021) | 79.74±0.15 | 74.84±0.13 | 84.29±0.12 | 79.62 |
| BDIL Xiao et al. (2021) | 85.56±0.12 | 71.89±0.14 | 85.90±0.18 | 81.05 |
| DNA Chu et al. (2022) | 83.93±0.18 | 73.94±0.15 | 85.57±0.17 | 81.14 |
| DSULi et al. (2022c) | 81.56±0.14 | 72.47±0.12 | 83.94±0.16 | 79.32 |
| Ours (in this paper) | 88.82±0.09 | 76.71±0.10 | 86.92±0.14 | 84.06 |
| Method | DeepAll | MASF Dou et al. (2019) | LDGG Li et al. (2020a) | KDGG | DSU Li et al. (2022c) | Ours |
| Wang et al. (2021) | ||||||
| DSC↑ | 0.7425 | 0.7710 | 0.7881 | 0.7886 | 0.7921 | 0.7957 |
| CC个 | -11.4 | 23.52 | 34.86 | 33.43 | 34.65 | 35.76 |
| JI个 | 0.6160 | 0.6502 | 0.6667 | 0.6667 | 0.6775 | 0.6828 |
| TPR↑ | 0.7667 | 0.7803 | 0.8058 | 0.8075 | 0.8225 | 0.8260 |
| ASD | 0.5265 | 0.5505 | 0.4076 | 0.3553 | 0.4362 | 0.3356 |
| Backbone (Unet) | Bayesian Layers | Local Alignment Alignment | Global | Bayesian Classifier | DSC | CC | JI | TPR | ASD |
| √ | X | = | = | X | 0.7223 | 26.21 | 0.5789 | 0.8109 | 0.0992 |
| = | 0.7934 47.19 0.6595 | 0.8133 | 30.0692 | ||||||
| 0.8268 57.52 | 0.7067 | 0.8156 | 50.0501 | ||||||
| ? | X | 0.8364 | 60.72 | 0.7195 | 0.8267 | 0.0486 | |||
| ss> | × | 0.837160.57 0.7217 | 0.8152 | 20.0510 | |||||
| ? | Xxs>>> | 0.848563.78 0.7389 0.8401 | 10.0401 |
| Method | Reference | Art | Cartoon | Photo | Sketch | Average (%) |
| RSCHuang et al. (2020) L2A-OT Zhou et al. (2020) | ECCV2020 ECCV2020 | 78.9 83.3 | 76.9 78.2 | 94.1 96.2 | 76.8 73.6 | 81 .7 82.8 |
| MatchDG Mahajan et al. (2021) pAdaIN Nuriel et al. (2021) | ICML2020 CVPR 2021 | 81.2 81.7 | 80.4 76.6 | 96.8 | 77.2 | 83.9 |
| MixStyle Zhou et al. (2021) | 96.3 | 75.1 | 82.5 | |||
| ICLR2021 | 86.8 | 79.0 | 96.6 | 78.5 | 85.2 | |
| SagNet Nam et al. (2021) | CVPR2021 | 87.4 | 80.7 | 97.1 | 80.0 | 86.3 |
| SWAD Cha et al. (2021) | NeurIPS2021 | 89.3 | 83.4 | 97.3 | 82.5 | 88.1 |
| DNA Chu et al. (2022) | ICML2022 | 89.8 | 83.4 | 97.7 | 82.6 | |
| 88.4 | ||||||
| Bayesian | = | 89.4 | 83.5 | 97.3 | 82.3 | 88.1 |
| Ours | 90.2 | 85.2 | 98.7 | 83.6 | 89.4 |
| (a) DeepAll | (b) KDDG | |||||||||||||||
| source | target | DSC | CC | JI | TPR | ASD | source | target | DSC | CC | JI | TPR | ASD | |||
| 2,3,4 | 1 | 0.8560 | 65.34 | 0.7520 | 0.8746 | 0.0809 | 2,3,4 | 1 | 0.8745 | 70.75 | 0.7795 | 0.8949 | 0.0539 | |||
| 1,3,4 | 2 | 0.7323 | 26.21 | 0.5789 | 0.8109 | 0.0992 | 1,3,4 | 2 | 0.8229 | 56.71 | 0.6997 | 0.8226 | 0.0490 | |||
| 1,2,4 | 3 | 0.5041 | -209 | 0.3504 | 0.4926 | 1.8661 | 1,2.4 | 3 | 0.5676 | -63.1 | 0.3866 | 0.5904 | 1.2805 | |||
| 1,2.3 | 4 | 0.8775 | 71.92 | 0.7827 | 0.8888 | 0.0599 | 1,2,3 | 4 | 0.8894 | 75.06 | 0.8011 | 0.9222 | 0.0377 | |||
| Average | 0.7425 | -11.4 | 0.6160 | 0.7667 | 0.5265 | Average | 0.7886 | 34.86 | 0.6667 | 0.8075 | 0.3553 | |||||
| (c) MASF | (d) LDDG | |||||||||||||||
| source | target | DSC | CC | JI | TPR | ASD | source | target | DSC | CC | JI | TPR | ASD | |||
| 2.3,4 | 1 | 0.8502 | 64.22 | 0.7415 | 0.8903 | 0.2274 | 2.3,4 | 1 | 0.8708 | 69.29 | 0.7753 | 0.8978 | 0.0411 | |||
| 1,3,4 | 2 | 0.8115 | 53.04 | 0.6844 | 0.8161 | 0.0826 | 1,3,4 | 2 | 0.8364 | 60.58 | 0.7199 | 0.8485 | 0.0416 | |||
| 1,2,4 | 3 | 0.5285 | -99.3 | 0.3665 | 0.5155 | 1.8554 | 1,2,4 | 3 | 0.5543 | -71.6 | 0.3889 | 0.5923 | 1.5187 | |||
| 1,2,3 | 4 | 0.8938 | 76.14 | 0.8083 | 0.8991 | 0.0366 | 1,2.3 | 4 | 0.8910 | 75.46 | 0.8039 | 0.8844 | 0.0289 | |||
| Average | 0.7710 | 23.52 | 0.6502 | 0.7803 | 0.5505 | Average | 0.7881 | 33.43 | 0.6720 | 0.8058 | 0.4076 | |||||
| (e)DSU | (f) Ours | |||||||||||||||
| source | target | DSC | CC | JI | TPR | ASD | source | target | DSC | CC | JI | TPR | ASD | |||
| 2,3,4 | 1 | 0.8739 | 70.32 | 0.7794 | 0.9210 | 0.0793 | 2,3,4 | 1 | 0.8786 | 71.57 | 0.7873 | 0.9293 | 0.0422 | |||
| 1,3,4 | 2 | 0.8474 | 63.58 | 0.7367 | 0.8502 | 0.0494 | 1,3,4 | 2 | 0.8485 | 63.78 | 0.7389 | 0.8401 | 0.0401 | |||
| 1,2,4 | 3 | 0.5574 | -70.4 | 0.3923 | 0.6097 | 1.5049 | 1,2,4 | 3 | 0.5634 | -68.0 | 0.3992 | 0.6103 | 1.2239 | |||
| 1,2,3 | 4 | 0.8897 | 75.10 | 0.8018 | 0.9225 | 0.0415 | 1,2.3 | 4 | 0.8921 | 75.69 | 0.8058 | 0.9245 | 0.0362 | |||
| Average | 0.7921 | 34.65 | 0.6775 | 0.8225 | 0.4362 | Average | 0.7957 | 35.76 | 0.6828 | 0.8260 | 0.3356 | |||||
| Algorithm | Art | Clipart | Product | Real | Avg |
| Mixstyle Zhou et al. (2021) | 51.1 | 53.2 | 68.2 | 69.2 | 60.4 |
| RSC Huang et al. (2020) | 60.7 | 51.4 | 74.8 | 75.1 | 65.5 |
| DANN Ganin et al. (2016) | 59.9 | 53.0 | 73.6 | 76.9 | 65.9 |
| GroupDRO Sagawa et al. (2019) | 60.4 | 52.7 | 75.0 | 76.0 | 66.0 |
| MTL Blanchard et al. (2021) | 61.5 | 52.4 | 74.9 | 76.8 | 66.4 |
| VREx Krueger et al. (2021) | 60.7 | 53.0 | 75.3 | 76.6 | 66.4 |
| MLDG Balaji et al. (2018) | 61.5 | 53.2 | 75.0 | 77.5 | 66.8 |
| SagNet Qian et al. (2021) | 63.4 | 54.8 | 75.8 | 78.3 | 68.1 |
| CORAL Sun & Saenko (2016) | 65.3 | 54.4 | 76.5 | 78.4 | 68.7 |
| SWAD Cha et al. (2021) | 66.1 | 57.7 | 78.4 | 80.2 | 70.6 |
| DNA Chu et al. (2022) | 67.7 | 57.7 | 78.9 | 80.5 | 71.2 |
| Bayesian | 67.0 | 58.0 | 79.3 | 80.4 | 71.2 |
| Ours | 68.2 | 58.9 | 80.2 | 80.7 | 72.0 |
| Algorithm | C | L ? | S | V | Ayg |
| Mixstyle Zhou et al. (2021) RSC Huang et al. (2020) DANN Ganin et al. (2016) GroupDRO Sagawa et al. (2019) | 98.3 97.9 99.0 97.3 97.8 | 64.8 62.5 65.1 63.4 64.3 64.4 | 72.1 72.3 73.1 69.5 71.5 74.1 | 74.3 75.6 77.2 76.7 75.3 76.2 | 77.4 77.1 78.6 76.7 77.2 78.3 |
| Task | Datasets Domains) and Corresponding Size | Number of Class |
| Skin Lesion Classification | HAM:10015: DMF:1212: D7P:1926:MSK:3551;UDA:601: PH2:200: SON:9251 | 7 |
| Epithelium Stroma Classification | NKI: 671: IHC:1376: VGH: 615 | 2 |
| Spnal Cord GM Segementation | site1: 30; site2: 113: site3: 246: site4: 122 | 2 (pixel-level) |
| PACS | Art:2048: Cartoon:2344: Photo:1670: Sketch:3929 | 7 |
| OfficeHome | A total of around 15500 images for 4 domains(Art,Clipart,Product,and Real with around 3897 per domain) | 6 |
| VLCS | VOC2007 (V): 3376: LabelMe (L):2656;SUN09(S):3282: Caltech101 (C): 1415 | 5 |
| Number of sample per class for each source domain | DeepAll | DSU | BDIL | DNA? | Qurs |
| 40 | 0.5399 ±0.0156 | 0.6145 ±0.0175 | 0.5897±0.0029 0.5412 ±0.0143 0.6368±0.0074 | ||
| 30 | 0.5309±0.02010.5458±0.01840.5762_±0.01010.5132±0.02290.6138±0.0291 | ||||
| 20 | 0.5044 ±0.0129 | 0.5243 ±0.0143 | 0.5573 ±0.0011 | 0.5048 ±0.00870.6037±0.0121 |
| Fraction of sample for each source domain | DeepAll | DSU | BDIL | DNA | Qurs |
| 100% | 0.6674 ±0.0312 | 0.6935 ±0.0121 | 0.7059 ±0.0284 | 0.7121_±0.0141 | 0.7276 ±0.0123 |
| 80% | 0.6614 ±0.0123 | 0.6717 ±0.0029 | 0.6625 ±0.092 | 0.6591 ±0.0022 | 0.6975 ±0.0036 |
| 60% | 0.6249 ±0.0122 | 0.6299 ±0.0114 | 0.6468 ±0.0106 | 0.6149 ±0.0112 | 0.6641 ±0.0114 |
| 40% | 0.5911 ±0.0215 | 0.6188 ±0.0541 | 0.6491±0.0171 | 0.6065 ±0.0111 | 0.6579 ±0.0057 |
| Method | Epochs | Keeping ratio (p) | Top-k &BBR | AP50 AP75 APs APM APL | |||||||||
| AP | params FLOPs FPS | ||||||||||||
| ResNet-50 backbone: | |||||||||||||
| F-RCNN-FPNt | 109 | N/A | 42.0 | 62.1 | 45.5 | 26.6 | 45.4 | 53.4 | 42M | 180G | 26 | ||
| DETRt | 500 | 100% | 42.0 | 62.4 | 44.2 | 20.5 | 45.8 | 61.1 | 41M | 86G | 28 | ||
| DETR-DC5† | 500 | 100% | 43.3 | 63.1 | 45.9 | 22.5 | 47.3 | 61.1 | 41M | 187G | 12 | ||
| PnP-DETR$ | 500 | 33% | 41.1 | 61.5 | 43.7 | 20.8 | 44.6 | 60.0 | - | 1 | - | ||
| 500 | 50% | 41.8 | 62.1 | 44.4 | 21.2 | 45.3 | 60.8 | - | 1 | - | |||
| PnP-DETR-DC5‡ | 500 | 33% | 42.7 | 62.8 | 45.1 | 22.4 | 46.2 | 60 | - | - | - | ||
| 500 | 50% | 43.1 | 63.4 | 45.3 | 22.7 | 46.5 | 61.1 | 1 | - | - | |||
| Deformable-DETR | 50 | 100% | 43.9 | 62.8 | 47.8 | 26.1 | 47.4 | 58.0 | 40M | 173G | 19.1 | ||
| 50 | 100% 10% | √ | 46.0 | 65.2 | 49.8 | 28.2 | 49.1 | 61.0 | 41M | 177G | 18.2 | ||
| Sparse-DETR | 50 50 | √ | 45.3 65.8 | 49.3 | 28.4 | 48.3 | 60.1 | 41M | 105G | 25.3 | |||
| 20% | √ | 45.6 65.8 | 49.6 | 28.5 | 48.6 | 60.4 | 41M | 113G | 24.8 | ||||
| 50 | 30% | √ | 46.0 | 65.9 | 49.7 | 29.1 | 49.1 | 60.6 | 41M | 121G | 23.2 | ||
| 50 | 40% | √ | 46.2 | 66.0 | 50.3 | 28.7 | 49.0 | 61.4 | 41M | 128G | 21.8 | ||
| 50 | 50% | √ | 46.3 | 66.0 | 50.1 | 29.0 | 49.5 | 60.8 | 41M | 136G | 20.5 | ||
| Swin-T backbone: | |||||||||||||
| DETR | 500 | 100% | 45.4 | 66.2 | 48.1 | 22.9 | 49.5 | 65.9 | 45M | 92G | 26.8 | ||
| Deformable-DETR | 50 | 100% | 45.7 | 65.3 | 49.9 | 26.9 | 49.4 | 61.2 | 40M | 180G | 15.9 | ||
| 50 | 100% | √ | 48.0 | 68.0 | 52.0 | 30.3 | 51.4 | 63.7 | 41M | 185G | 15.4 | ||
| Sparse-DETR | 50 | 10% | √ | 48.2 | 69.2 | 52.3 | 29.8 | 51.2 | 64.5 | 41M | 113G | 21.2 | |
| 50 | 20% | √ | 48.8 | 69.4 | 53.0 | 41M | 121G | 20.0 | |||||
| 50 | 30% | √ | 49.1 | 69.5 | 30.4 31.4 | 51.9 52.5 | 64.8 65.1 | 41M | 129G | 18.9 | |||
| 50 | √ | 49.2 | 69.5 | 53.5 53.5 | 31.4 | 52.9 | 64.8 | 41M | 136G | 18.0 | |||
| 50 | 40% 50% | √ | 49.3 | 69.5 | 53.3 | 32.0 | 52.7 | 64.9 | 41M | 144G | 17.2 | ||
| Loss | Keeping ratio (p) | AP AP50 | AP75 | APs | APM | APL |
| smoothed L1 | 10% | 47.8 68.9 | 51.7 | 29.8 | 50.9 | 64.0 |
| 20% | 48.4 69.0 | 52.5 | 31.1 | 51.4 | 64.7 | |
| 30% | 48.6 69.0 | 52.6 | 31.1 | 51.9 | 64.7 | |
| 40% | 48.6 69.3 | 52.9 | 33.4 | 51.8 | 64.5 | |
| ranking | 10% | 48.0 69.1 | 52.1 | 29.9 | 51.4 | 64.6 |
| 20% | 48.7 69.5 | 53.0 | 31.1 | 51.8 | 65.1 | |
| 30% | 48.8 69.2 | 52.8 | 31.4 | 52.0 | 64.9 | |
| 40% | 48.9 69.3 | 53.1 | 31.5 | 52.2 | 64.7 | |
| BCE | 10% | 48.2 69.2 | 52.3 | 29.8 | 51.2 | 64.5 |
| 20% | 48.8 69.4 | 53.0 | 30.4 | 51.9 | 64.8 | |
| 30% | 49.1 69.5 | 53.5 | 31.4 | 52.5 | 65.1 | |
| 40% | 49.2 69.5 | 53.5 | 31.4 | 52.9 | 64.8 |
| Scoring method | Keeping | |||||||||
| ratio (p) ResNet-50 backbone: | AP AP50 | AP75 | APs | APM | APL | params | FLOPs | FPS | ||
| N/A | 100% | 46.3 | 65.8 | 50.1 | 29.0 | 49.4 | 61.7 | 41M | 177G | 18.2 |
| 0% | 42.2 | 63.0 | 45.6 | 25.9 | 45.3 | 56.5 | 36M | 91G | 35.0 | |
| random | 10% 20% | 43.6 44.0 | 64.3 64.8 | 47.2 47.8 | 26.7 27.3 | 46.9 | 58.4 | 41M | 102G | 27.7 |
| 30% | 44.0 | 47.0 | 58.4 | 41M | 110G | 25.6 | ||||
| 64.9 | 47.5 | 27.4 | 47.4 | 58.2 | 41M | 117G | 24.1 | |||
| 40% 50% | 44.5 44.4 | 65.1 | 48.0 48.0 | 27.3 27.8 | 47.8 47.4 | 59.8 59.2 | 41M 41M | 125G | 22.5 21.1 | |
| OS | 10% | 44.9 | 64.8 65.2 | 48.7 | 27.9 | 47.8 | 60.4 | 41M | 133G 106G | 26.6 |
| 20% | 45.5 | 65.5 | 49.3 | 28.7 | 48.3 | 60.5 | 41M | 114G | 24.7 | |
| 30% | 45.7 | 65.8 | 49.5 | 29.7 | 48.5 | 60.8 | 41M | 121G | 23.2 | |
| 40% | 45.8 | 65.5 | 49.8 | 29.1 | 48.8 | 41M | 21.8 | |||
| 50% | 46.0 | 65.9 | 49.8 | 28.8 | 48.9 | 60.5 60.6 | 41M | 129G 136G | 20.6 | |
| DAM | 10% | 45.3 | 65.8 | 49.3 | 28.4 | 48.3 | 60.1 | 41M | 105G | 26.5 |
| 20% | 45.6 | 65.8 | 49.6 | 28.5 | 48.6 | 60.4 | 41M | 113G | 24.8 | |
| 30% | 46.0 | 65.9 | 49.7 | 29.1 | 49.1 | 60.6 | 41M | 121G | 23.2 | |
| 40% | 46.2 | 66.0 | 50.3 | 28.7 | 49.0 | 61.4 | 41M | 128G | 21.8 | |
| 50% | 46.3 | 66.0 | 50.1 | 29.0 | 49.5 | 60.8 | 41M | 136G | 20.5 | |
| Swin-T backbone: | ||||||||||
| N/A | 100% 0% | 49.4 43.7 | 69.4 | 53.5 | 31.9 | 52.6 | 65.1 | 41M | 185G | 15.4 |
| 10% | 45.5 | 65.8 | 46.9 | 27.0 | 46.7 | 60.0 | 37M | 96G | 26.5 | |
| random | 20% | 45.6 | 67.6 | 48.8 | 28.4 | 48.5 | 62.2 | 41M | 110G | 22.1 |
| 30% | 67.5 | 49.2 | 28.6 | 49.1 | 62.2 | 41M | 118G | 20.8 | ||
| 46.2 | 68.1 | 49.7 | 29.5 | 49.7 | 63.0 | 41M | 125G | 19.7 | ||
| 40% | 46.5 | 68.2 | 50.0 | 29.9 | 49.8 | 63.0 | 41M | 133G | 18.7 | |
| 50% | 47.2 | 68.3 | 50.9 | 29.1 | 50.4 | 63.9 | 41M | 141G | 17.7 | |
| OS | 10% | 48.0 | 69.1 | 52.1 | 29.9 | 51.1 | 64.4 | 42M | 114G | 21.4 |
| 20% | 48.3 | 69.1 | 52.5 | 30.4 | 51.6 | 64.2 | 42M | 122G | 20.2 | |
| 30% | 48.6 | 69.2 | 53.0 | 31.0 | 52.0 | 64.6 | 42M | 129G | 18.6 | |
| 40% | 48.9 | 69.4 | 53.1 | 33.0 | 51.9 | 64.5 | 42M | 137G | 18.2 | |
| 50% | 49.0 | 69.2 | 53.5 | 31.2 | 52.4 | 65.0 | 42M | 145G | 17.2 | |
| DAM | 10% | 48.2 | 69.2 | 52.3 | 29.8 | 51.2 | 64.5 | 41M | 113G | 21.2 |
| 20% | 48.8 | 69.4 | 53.0 | 30.4 | 51.9 | 64.8 | 41M | 121G | 20.0 | |
| 30% | 49.1 | 69.5 | 53.5 | 31.4 | 52.5 | 65.1 | 41M | 129G | 18.9 | |
| 40% | 49.2 | 69.5 | 53.5 | 31.4 | 52.9 | 64.8 | 41M | 136G | 18.0 | |
| 50% | 49.3 | 69.5 | 53.3 | 32.0 | 52.7 | 64.9 | 41M | 144G | 17.2 | |
| #of encoder | Keeping ratio (p) | Aux. loss | |||||||||
| AP | AP50 | AP75 | APs | APM | APL | params FLOPs | FPS | ||||
| 6 | 100% | 48.0 | 68.0 | 52 | 30.3 | 51.4 | 63.7 | 41M | 185G | 15.4 | |
| 100% | √ | 49.4 | 69.4 | 53.5 | 31.9 | 52.6 | 65.1 | 41M | 185G | 15.4 | |
| 10% | 46.8 | 68.0 | 50.6 | 29.7 | 49.7 | 63.3 | 41M | 113G | 21.2 | ||
| 20% | 47.5 | 68.3 | 51.4 | 31.4 | 50.4 | 64.4 | 41M | 121G | 20.0 | ||
| 30% | 47.6 | 67.9 | 51.4 | 29.9 | 51.1 | 63.9 | 41M | 129G | 18.9 | ||
| 40% | 47.6 | 68.2 | 51.5 | 30.3 | 50.8 | 64.0 | 41M | 136G | 18.0 | ||
| 10% | 48.2 | 69.2 | 52.3 | 29.8 | 51.2 | 64.5 | 41M | 113G | 21.2 | ||
| 20% | 48.8 | 69.4 | 53.0 | 30.4 | 51.9 | 64.8 | 41M | 121G | 20.0 | ||
| 30% | √>>> | 49.1 | 69.5 | 53.5 | 31.4 | 52.5 | 65.1 | 41M | 129G | 18.9 | |
| 9 | 40% | 49.2 | 69.5 | 53.5 | 31.4 | 52.9 | 64.8 | 41M | 136G | 18.0 | |
| 12 | 100% | √ | 49.7 | 69.4 | 54.1 | 32.4 | 52.9 | 65.4 | 44M | 220G | 12.8 |
| 100% | √ | 50.1 | 69.6 | 54.6 | 32.2 | 53.4 | 65.8 | 46M | 261G | 11.0 | |
| 10% | 49.0 | 69.5 69.6 | 53.5 53.5 | 31.6 | 52.2 | 65.2 | 46M | 128G | 19.2 | ||
| 20% 30% | >> | 49.4 49.3 | 69.4 | 53.6 | 31.9 31.7 | 52.8 52.5 | 65.4 65.6 | 46M 46M | 143G 158G | 17.5 | |
| 40% | 49.8 | 69.8 | 54.3 | 33.1 | 53.4 | 65.4 | 46M | 173G | 15.6 14.6 | ||
| Method | Keeping ratio (p) | |||||||||
| AP | AP50 | AP75 | APs | APM | APL | params FLOPs | FPS | |||
| Sparse-DETR | 10% | 45.3 | 65.8 | 49.3 | 28.4 | 48.3 | 60.1 | 41M | 105G | 25.3 |
| 20% | 45.6 | 65.8 | 49.6 | 28.5 | 48.6 | 60.4 | 41M | 113G | 24.8 | |
| 30% | 46.0 | 65.9 | 49.7 | 29.1 | 49.1 | 60.6 | 41M | 121G | 23.2 | |
| 40% | 46.2 | 66.0 | 50.3 | 28.7 | 49.0 | 61.4 | 41M | 128G | 21.8 | |
| 50% | 46.3 | 66.0 | 50.1 | 29.0 | 49.5 | 60.8 | 41M | 136G | 20.5 | |
| Sparse-DETR$ | 10% | 46.9 | 67.2 | 51.0 | 30.2 | 49.7 | 62.3 | 41M | 105G | 25.3 |
| 20% | 47.3 | 67.1 | 51.4 | 29.7 | 50.3 | 62.7 | 41M | 113G | 24.8 | |
| 30% | 47.4 | 67.3 | 51.4 | 30.1 | 50.5 | 62.4 | 41M | 121G | 23.2 | |
| 40% | 47.7 | 67.4 | 51.6 | 30.0 | 50.8 | 62.9 | 41M | 128G | 21.8 | |
| 50% | 47.9 | 67.5 | 52.1 | 30.5 | 51.2 | 63.2 | 41M | 136G | 20.5 | |
| Backbone | Keeping ratio (p) | Aux. loss | FLOPs | ||||||||
| AP | AP50 | AP75 APs | APM | APL | params | FPS | |||||
| Swin-T | 100% | 48.0 | 68.0 | 52.0 | 30.3 | 51.4 | 63.7 | 41M | 185G | 15.4 | |
| 10% | 48.2 | 69.2 | 52.3 | 29.8 | 51.2 | 64.5 | 41M | 113G | 21.2 | ||
| 20% | 48.8 | 69.4 | 53.0 | 30.4 | 51.9 | 64.8 | 41M | 121G | 20.0 | ||
| 30% | 49.1 | 69.5 | 53.5 | 31.4 | 52.5 | 65.1 | 41M | 129G | 18.9 | ||
| 40% | <>>> | 49.2 | 69.5 | 53.5 | 31.4 | 52.9 | 64.8 | 41M | 136G | 18.0 | |
| Swin-B | 100% | 52.5 | 72.9 | 56.9 | 34.7 | 56.5 | 69.6 | 101M | 400G | 7.6 | |
| 10% | 52.2 | 73.5 | 57.0 | 34.0 | 56.3 | 70.3 | 101M | 335G | 8.8 | ||
| 20% | 53.1 | 73.8 | 57.9 | 34.6 | 56.9 | 70.6 | 101M | 343G | 8.6 | ||
| 30% | <>>> | 53.2 | 73.7 | 57.7 | 35.3 | 56.8 | 70.8 | 101M | 350G | 8.4 | |
| 40% | 53.3 | 73.4 | 58.0 | 36.3 | 57.2 | 70.9 | 101M | 358G | 8.2 | ||
| Keeping ratio (p) | AP AP50 | AP75 | APs | APM | APL |
| 100% | 48.0 68.0 | 52.0 | 30.3 | 51.4 | 63.7 |
| 10% | 44.0 66.0 | 47.2 | 26.9 | 46.8 | 61.0 |
| 20% | 44.9 66.3 | 48.3 | 28.2 | 48.2 | 61.4 |
| 30% | 46.5 67.3 | 50.2 | 30.9 | 49.7 | 62.3 |
| 40% | 47.3 67.9 | 51.3 | 30.7 | 50.7 | 63.4 |
| 50% | 47.9 67.8 | 52.0 | 29.8 | 51.4 | 63.7 |
| AigormmmrT.CartoonA Data: Image x ∈ [0,1]exwxt with c channels and wt pixels, classifier Φ. Result: CartoonX explanation ε ∈ [0,1]w×t for decision Φ(x). Hyperparameters: Sparsity level X > O, number of steps N, number of noise samples L. Initialize mask s := [1,..,1]T ∈ [0,1]k on DWT coefficients h = [h1,.., hk] with x = f(h), where f is the discrete inverse wavelet transform; Compute predicted label j* := arg maxi Φ(x); fori←1toNdo |
| Sample L adaptive Gaussian noise samples u(1),., u(L) ~ N(μ,o²); Compute obfuscations y(1), ),.,y(L) with y() := f(h ① s+ (1- s) ①u(i)); Clip obfuscations into [0,1]cx w ×t; |
| Approximate expected distortion D(𝑥x,s,Φ) :=∑𝑖=1(Φj+(x)- Φj(y())²/L; |
| Compute loss for the mask l(s) := D(x,s,Φ) + λ|lsll1 and gradient Vsl(s); Update mask s with gradient descent step and clip s back to [0,1]k ; |
| end Compute wavelet coefficients h for greyscale image x of x; |
| Invert wavelet mask s back to pixel space as & := f(h s) ; Clip the explanation ε into [0,1]w×t to obtain ε. Visualize ε; |
| Model | # Params | Train loss (pretraining) | Val loss (pretraining) | In-context Acc |
| Vocab size 50, k =10,n =64 | ||||
| Transformer (4 layer) | 29M | 1.49 | 1.50 | 60.2 ± 5.7 |
| Transformer(12 layer) | 85M | 1.31 | 1.33 | 81.2 ± 7.1 |
| Transformer (16 layer) | 115M | 1.31 | 1.33 | 84.7 ± 3.4 |
| LSTM | 28M | 1.31 | 1.35 | 95.8 ± 1.11 |
| Vocab size 100, k = 10,n =64 | ||||
| Transformer (4 layer) | 29M | 1.58 | 1.59 | 67.4 ± 4.7 |
| Transformer(12 layer) | 85M | 1.40 | 1.42 | 84.6 ± 3.0 |
| Transformer (16 layer) | 115M | 1.41 | 1.43 | 88.7 ± 1.6 |
| LSTM | 28M | 1.43 | 1.44 | 95.8 ± 1.54 |
| Vocab size 150, k = 10,n =64 | ||||
| Transformer (4 layer) | 29M | 1.44 | 1.45 | 92.8 ± 1.9 |
| Transformer (12 layer) | 85M | 1.27 | 1.28 | 98.4 ±0.4 |
| Transformer (16 layer) | 115M | 1.27 | 1.28 | 98.1 ± 0.5 |
| LSTM | 28M | 1.26 | 1.31 | 99.2 ± 1.06 |
| Prompt example length | Test Acc (200-300 chars) |
| 5 examples | |
| Short (200-300 chars) | 69.8 |
| Long (500-600 chars) | 70.7 |
| 10 examples | |
| Short, duplicated examples | 69.6 |
| Short, independent examples | 71.4 |
| Terrain | Average x-Displacement (↑) | Mean Time to Fall (s) | ||||||
| RMA | MLith | Noisy | Blind | RMA | MLith | Noisy | Blind | |
| Slopes | 43.98 | 44.09 | 36.14 | 34.72 | 88.99 | 85.68 | 70.25 | 67.07 |
| Stepping Stones | 18.83 | 20.72 | 1.09 | 1.02 | 34.3 | 41.32 | 2.51 | 2.49 |
| Stairs | 31.24 | 42.4 | 6.74 | 16.64 | 69.99 | 90.48 | 15.77 | 39.17 |
| Discrete Obstacles | 40.13 | 28.64 | 29.08 | 32.41 | 85.17 | 57.53 | 59.3 | 66.33 |
| Total | 134.18 | 135.85 | 73.05 | 84.79 | 278.45 | 275.01 | 147.83 | 175.06 |
| Algorithm1:Conditional Statements |
| Define intervals I CR,i=1,.,M Define functions |
| hi,,i=1,.,M+1 switch h1(x) do |
| if hi(x)∈Ii then L apply hi+1 to x |
| Algorithm | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
| ERM (ResNet50) (Vapnik,1991) | 85.7± 0.5 | 77.4±0.3 | 67.5±0.5 | 47.2±0.4 | 41.2 ±0.2 |
| IRM[ArXiv 20] (Arjovsky et al.,2019) | 83.5± 0.8 | 78.5± 0.5 | 64.3± 2.2 | 47.6± 0.8 | 33.9 ± 2.8 |
| DANN [JMLR 16] (Ganin et al., 2016) | 84.6 ± 1.1 | 78.7± 0.3 | 68.6 ± 0.4 | 46.4±0.8 | 41.8 ± 0.2 |
| CORAL [ECCV16] (Sun & Saenko,2016) | 86.0± 0.2 | 77.7±0.5 | 68.6±0.4 | 46.4±0.8 | 41.8 ± 0.2 |
| MMD [CVPR 18] (Li et al.,2018b) | 85.0± 0.2 | 76.7 ± 0.9 | 67.7 ± 0.1 | 42.2 ± 1.4 | 39.4±0.8 |
| FISH[ICLR 22] (Shi et al., 2021) | 85.5± 0.3 | 77.8± 0.3 | 68.6±0.4 | 45.1 ± 1.3 | 42.7 ± 0.2 |
| SWAD [NeurIPS 21] (Cha et al.,2021a) | 88.1± 0.1 | 79.1 ± 0.1 | 70.6± 0.2 | 50.0±0.3 | 46.5 ± 0.1 |
| Fishr [ICML 22] (Rame et al., 2021) | 85.5±0.2 | 77.8±0.2 | 68.6±0.2 | 47.4 ± 1.6 | 41.7 ± 0.0 |
| MIRO [ECCV 22] (Cha et al., 2022) | 85.4 ± 0.4 | 79.0 ± 0.0 | 70.5± 0.4 | 50.4 ± 1.1 | 44.3 ± 0.2 |
| ERM (ViT-S/16) [ICLR 21] (Dosovitskiy et al., 2021) | 86.2 ± 0.1 | 79.7 ± 0.0 | 72.2 ± 0.4 | 42.0 ± 0.8 | 47.3 ± 0.2 |
| GMoE-S/16 (Ours) | 88.1 ± 0.1 | 80.2 ± 0.2 | 74.2 ± 0.4 | 48.5 ± 0.4 | 48.7 ± 0.2 |
| Algorithms | SVIRO | Wilds-Camelyon | Wilds-FMOW | ||
| ERM (ResNet50) (Vapnik, 1991) | 85.7 ± 0.1 | 93.1± 0.2 | 40.6± 0.4 | ||
| ERM(ViT-S/16)[ICLR 21] (Dosovitskiy et al., 2021) | 89.6 ± 0.0 | 91.1 ± 0.1 | 44.8 ± 0.2 | ||
| GMoE-S/16 (Ours) | 90.3 ± 0.1 | 93.7 ± 0.2 | 46.6 ± 0.4 |
| Algorithm | DomainNet |
| GMoE | 48.7 |
| Fish (Rame et al., 2021) | 42.7 |
| GMoEw/Fish | 48.8 |
| Swad (Cha et al., 2021a) | 46.5 |
| GMoE w/Swad | 49.6 |
| MACs | Paint | Clipart | Info | Paint | Quick | Real | Sketch | IID Imp. | 00D Imp. |
| 4.1G | ResNet50 | 37.1 | 12.9 | 62.7 | 2.2 | 49.3 | 33.3 | - | 1 |
| 7.9G | ResNet101 | 40.5 | 13.1 | 63.4 | 3.1 | 51.2 | 35.4 | 1.1% | 12.4% |
| 4.6G | ViT-S/16 | 42.7 | 15.9 | 69.0 | 5.0 | 56.4 | 37.0 | 10.0% | 38.6% |
| 4.8G | GMoE-S/16 | 43.5 | 16.1 | 69.3 | 5.3 | 56.4 | 38.0 | 10.5% | 42.3% |
| Backbone | ResNet-50 | ResNet-101 | ViT-S/16 | GMoE-S/16 |
| Pretrained Dataset | ImageNet-1K | ImageNet-1K | ImageNet-1K | ImageNet-1K* |
| Parameters | 25.6M | 42.9M | 21.7M | 33.8M |
| Flops | 4.1G | 7.9G | 4.6G | 4.8G |
| Patch | Expert 1 | Expert 2 | Expert 3 | Expert 4 | Selected Expert |
| X1 | 0.9 | 0.4 | 0.1 | 0.2 | Expert 1 |
| x2 | 0.2 | 0.4 | 0.9 | 0.1 | Expert 3 |
| 33 | 0.1 | 0.4 | 0.2 | 0.9 | Expert 4 |
| Dataset | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
| # Domains | 4 | 4 | 4 | 4 | 6 |
| # Classes | 7 | 5 | 65 | 10 | 345 |
| # Examples | 9,991 | 10,729 | 15,588 | 24,788 | 586,575 |
| Dataset | SVIRO | W-Camelyon | W-FMOW | CUB-DG |
| # Domains | 10 | 5 | 6 | 4 |
| # Classes | 7 | 2 | 62 | 200 |
| # Examples | 56,000 | 455,954 | 523,846 | 47,152 |
| Hyperparameters | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
| Learning Rate | 3×10-5 | 3×10-5 | 1×10-5 | 5×10-5 | 5×10-5 |
| Weight Decay | 0 | 1×10-6 | 1 ×10-6 | 1 ×10-4 | 0 |
| Iterations | Algorithms & Models | DomainNet (train-val) | DomainNet (lodo) |
| 5K | ERM (Vapnik, 1991) | 41.2 ± 0.2 | 40.6± 0.2 |
| IRM (Arjovsky et al., 2019) | 33.9 ± 2.8 | 33.5 ± 3.0 | |
| DANN (Ganin et al., 2016) | 41.8 ± 0.2 | 38.2 ± 0.2 | |
| CORAL (Sun & Saenko, 2016) | 41.8 ± 0.2 | 41.1 ± 0.1 | |
| MMD (Li et al., 2018b) | 39.4 ± 0.8 | 23.4 ± 9.4 | |
| MLDG (Li et al., 2018a) | 41.2 ± 0.1 | 41.0 ± 0.2 | |
| Fish (Shi et al., 2021) | 42.7 ± 0.2 | 1 | |
| Fishr (Rame et al., 2021) | 41.7 ± 0.2 | 1 | |
| ViT-S/16 (Dosovitskiy et al., 2021) | 42.4 ± 0.0 | 42.0 ± 0.2 | |
| GMoE-S/16 | 44.6 ± 0.3 | 44.7 ± 0.2 | |
| 15K | Swad (Cha et al., 2021a) | 46.5 ± 0.1 | |
| MIRO (Cha et al., 2022) | 47.4 ± 0.0 | ||
| ViT-S/16 (Dosovitskiy et al., 2021) | 47.1 ± 0.3 | 46.1 ± 0.4 | |
| GMoE-S/16 | 48.7 ± 0.2 | 48.4 ± 0.3 |
| Algorithm | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
| ERM (ResNet50) | 83.0 ±0.7 | 77.2 ± 0.4 | 65.7 ± 0.5 | 41.4 ± 1.4 | 40.6 ± 0.2 |
| IRM | 81.5 ±0.8 | 76.3 ± 0.6 | 64.3 ± 1.5 | 41.2 ± 3.6 | 33.5 ± 3.0 |
| DANN | 81.0 ± 1.1 | 76.9 ± 0.4 | 64.9 ± 1.2 | 44.4 ± 1.1 | 38.2 ±0.2 |
| CORAL | 82.6 ± 0.5 | 78.7 ±0.4 | 68.5 ± 0.2 | 46.3 ± 1.7 | 41.1 ± 0.1 |
| MMD | 83.2 ± 0.2 | 77.3 ± 0.5 | 60.2 ± 5.2 | 46.5 ± 1.5 | 23.4 ± 9.5 |
| MLDG | 77.2 ± 0.9 | 82.9 ± 1.7 | 66.1 ± 0.5 | 46.2 ± 0.9 | 41.0 ± 0.2 |
| ViT-S/16 | 86.6 ± 0.1 | 78.3 ± 0.2 | 71.9 ± 0.2 | 41.9 ±0.3 | 46.1 ± 0.2 |
| GMoE-S/16 | 87.1 ± 0.3 | 80.0 ± 0.1 | 73.9 ± 0.2 | 46.7 ± 0.3 | 48.4 ± 0.2 |
| Algorithms | Backbones | PACS | VLCS | OfficeHome | TerraInc | DomainNet | Avg |
| Swad | ViT-S/16 | 88.3 | 80.4 | 72.7 | 46.9 | 48.8 | 67.4 |
| GMoE-S/16 | 88.6 | 80.8 | 74.5 | 49.1 | 49.6 | 68.5 | |
| Fish | ResNet50 | 85.5 | 77.8 | 68.6 | 45.1 | 43.1 | 64 |
| ViT-S/16 | 87.3 | 78 | 72.2 | 42.7 | 47.3 | 65.5 | |
| GMoE-S/16 | 87.9 | 79.7 | 73.0 | 46.9 | 48.8 | 67.1 | |
| Fishr | ResNet50 | 85.5 | 77.8 | 68.6 | 47.4 | 41.7 | 64.2 |
| ViT-S/16 | 85.6 | 79.1 | 71.8 | 44.1 | 47.6 | 65.6 | |
| GMoE-S/16 | 88.3 | 78.8 | 72.7 | 45.6 | 48.3 | 66.7 |
| Clipart | Clipart | Info | Paint | Quick | Real | Sketch |
| ResNet50 | 68.5 | 12.3 | 25.8 | 9.5 | 39.6 | 37.1 |
| ResNet101 | 70.4 | 12.2 | 28.6 | 12.4 | 43.0 | 40.7 |
| ViT-S/16 | 74.7 | 16.2 | 35.3 | 12.7 | 50.9 | 40.4 |
| GMoE-S/16 | 76.3 | 16.8 | 36.2 | 13.9 | 51.6 | 42.6 |
| Info | Clipart | Info | Paint | Quick | Real | Sketch |
| ResNet50 | 27.9 | 34.0 | 23.3 | 2.0 | 34.3 | 24.5 |
| ResNet101 | 29.2 | 35.6 | 24.0 | 3.1 | 34.1 | 24.6 |
| ViT-S/16 | 36.3 | 40.2 | 33.6 | 5.1 | 47.4 | 30.7 |
| GMoE-S/16 | 36.7 | 41.0 | 34.2 | 4.5 | 48.2 | 30.8 |
| Paint | Clipart | Info | Paint | Quick | Real | Sketch |
| ResNet50 | 37.1 | 12.9 | 62.7 | 2.2 | 49.3 | 33.3 |
| ResNet101 | 40.5 | 13.1 | 63.4 | 3.1 | 51.2 | 35.4 |
| ViT-S/16 | 42.7 | 15.9 | 69.0 | 5.0 | 56.4 | 37.0 |
| GMoE-S/16 | 43.5 | 16.1 | 69.3 | 5.3 | 56.4 | 38.0 |
| Quick | Clipart | Info | Paint | Quick | Real | Sketch |
| ResNet50 | 17.3 | 1.1 | 2.8 | 62.4 | 7.3 | 9.2 |
| ResNet101 | 16.0 | 1.3 | 2.7 | 64.1 | 6.3 | 8.5 |
| ViT-S/16 | 22.1 | 2.0 | 6.3 | 66.2 | 11.8 | 11.8 |
| GMoE-S/16 | 22.8 | 2.0 | 6.4 | 66.8 | 12.2 | 12.9 |
| Real | Clipart | Info | Paint | Quick | Real | Sketch |
| ResNet50 | 39.5 | 13.0 | 40.0 | 4.4 | 73.6 | 30.5 |
| ResNet101 | 43.3 | 15.6 | 43.3 | 6.8 | 75.3 | 34.0 |
| ViT-S/16 | 46.9 | 18.0 | 46.0 | 5.0 | 79.5 | 33.4 |
| GMoE-S/16 | 47.4 | 18.5 | 46.9 | 7.5 | 79.6 | 36.4 |
| Sketch | Clipart | Info | Paint | Quick | Real | Sketch |
| ResNet50 | 48.9 | 11.7 | 31.6 | 10.8 | 40.7 | 63.4 |
| ResNet101 | 48.4 | 11.8 | 31.9 | 11.3 | 38.1 | 64.3 |
| ViT-S/16 | 52.4 | 14.4 | 38.6 | 13.4 | 48.2 | 67.4 |
| GMoE-S/16 | 52.9 | 14.1 | 36.6 | 13.7 | 49.0 | 68.4 |
| Algorithm | Config. | Router | H | D | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
| GMoE-S/16 | Every 2 | Linear | 6 | 384 | 81.8 ± 0.2 | 75.0 ± 0.1 | 64.0 ± 0.4 | 32.5 ± 0.7 | 46.3 ± 0.3 |
| GMoE-S/16 | Every 2 | Cosine | 6 | 384 | 81.4 ± 0.1 | 74.8 ± 0.2 | 62.2 ± 0.4 | 40.9 ± 0.3 | 46.4 ± 0.2 |
| GMoE-S/16 | Last 2 | Linear | 6 | 384 | 87.8 ±0.2 | 80.0 ±0.0 | 72.7 ± 0.2 | 46.7 ±0.2 | 48.3 ±0.1 |
| GMoE-S/16 | Last 2 | Cosine | 6 | 384 | 88.1 ± 0.1 | 80.2 ±0.2 | 74.2 ±0.4 | 48.5 ± 0.4 | 48.7 ±0.2 |
| GMoE-B/16 | Last 2 | Cosine | 12 | 768 | 89.4 ± 0.1 | 81.2 ± 0.1 | 77.2 ±0.4 | 49.3 ± 0.3 | 51.3 ± 0.1 |
| Recipe | LR Opt. | TrivialAug. | Ep. | Rand Er. | Label Sm. | FixRes Mt. | WDT | IRT | IN1K |
| ResNet-50 V2 | 600 | 卜 | ← | 80.8 | |||||
| ViT-S/16 | ← | × | 300 | ← | × | × | × | 【 | 79.9 |
| Algorithm | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
| ERM (w/ ResNet-50 V2) | 87.2 | 78.2 | 68.7 | 49.9 | 45.3 |
| Fishr (w/ ResNet-50 V2) | 87.5 | 77.9 | 70.4 | 51.7 | 47.0 |
| ViT-S/16 | 86.2 | 79.7 | 72.2 | 42.0 | 47.1 |
| GMoE-S/16 | 88.1 | 80.1 | 74.2 | 48.5 | 48.7 |
| Training | ERM DANN | IRM Fish | Fishr | SWAD | VIT-S/16 | GMoE-S/16 |
| Step Time (s)↓ | 1.01 1.02 | 1.10 | 2.79 1.10 | 1.21 | 0.90 | 0.98 |
| Run-time Memory (GB)↓ | 13.40 13.42 | 13.40 | 3.41 15.25 | 14.32 | 11.15 | 12.28 |
| Inference | ERM DANN | IRM | Fish Fishr | SWAD | VIT-S/16 | GMoE-S/16 |
| Step Time (s)↓ | 0.32 0.33 | 0.32 | 0.33 0.34 | 0.33 | 0.28 | 0.30 |
| Run-time Memory(GB)↓ | 1.82 1.83 | 1.82 | 1.82 1.83 | 1.84 | 0.76 | 1.05 |
| Algorithm | Candy | Mosaic | Natural | Udnie |
| ERM (ResNet50) | 64.3 | 20.8 | 75.3 | 77.9 |
| DANN | 47.0 | 14.9 | 76.5 | 70.6 |
| Fish | 62.6 | 22.4 | 84.5 | 77.7 |
| Fishr | 62.4 | 24.3 | 78.8 | 73.8 |
| GMoE-S/16 | 83.1 | 42.8 | 89.3 | 82.7 |
| Dataset | = std | max |
| noisy (Ours) | 18 | 253 |
| biased (Ours) | 9 | 31 |
| diverse (D4RL) | 2 | 11 |
| play (D4RL) | 2 | 13 |
| Task &Dataset | EDAC | BEAR | CQL | IQL | Ours |
| medium-noisy | 0 | 0 | 55 | 44 | 73 |
| medium-biased | 0 | 0 | 73 | 48 | 74 |
| large-noisy | 0 | 0 | 42 | 39 | 53 |
| large-biased | 0 | 0 | 50 | 41 | 45 |
| Dataset | BC | 10%BC | DT | AWAC | Onestep RL | TD3+BC | COMBO | CQL | IQL | (Ours) |
| halfcheetah-medium-replay | 36.6 | 40.6 | 36.6 | 40.5 | 38.1 | 44.6 | 55.1 | 45.5 | 44.2 | 52.3 |
| hopper-medium-replay | 18.1 | 75.9 | 82.7 | 37.2 | 97.5 | 60.9 | 89.5 | 95.0 | 94.7 | 101.5 |
| walker2d-medium-replay | 26.0 | 62.5 | 66.6 | 27.0 | 49.5 | 81.8 | 56.0 | 77.2 | 73.9 | 85.0 |
| halfcheetah-medium-expert | 55.2 | 92.9 | 86.8 | 42.8 | 93.4 | 90.7 | 90.0 | 91.6 | 86.7 | 89.5 |
| hopper-medium-expert | 52.5 | 110.9 | 107.6 | 55.8 | 103.3 | 98.0 | 111.1 | 105.4 | 91.5 | 110.0 |
| walker2d-medium-expert | 107.5 | 109.0 | 108.1 | 74.5 | 113.0 | 110.1 | 103.3 | 108.8 | 109.6 | 112.0 |
| locomotion total | = 295.9 | 491.8 | 488.4 | 277.8 | 494.8 | 486.1 | 505 | 523.5 | 500.6 | 550.3 |
| Task | Ⅱ CQL | CQL (ReDS) | std D(π,πβ)(s) | max D(π,πβ)(s) |
| Pick &Place | 6.5± 0.4 | 15.1 ± 0.4 | 48.7 | 307.4 |
| Bin Sort (Easy) | 31.2 ± 0.3 | 31.4 ± 0.3 | 7.9 | 81.6 |
| Bin Sort (Hard) | 6.1 ± 0.2 | 23.1 ± 0.7 | 59.6 | 988.3 |
| q-data = critic(batch['observations’],batch['actions’]) | |
| next_dist = actor(batch['next_observations']) | next-pi_actions, next_log-pis = next_dist.sample() |
| target-qval = target_critic(batch['observations'], | |
| next-pi-actions) target-qval = batch['rewards’] + | |
| self.gamma * (1 - batch['dones’]) * target-qval | |
| td_loss = mse_loss(q-data, target-qval) | |
| # importance sampling term | |
| num_samples = 4 | |
| # assume env is normalized between [-1,1] | |
| random_actions = uniform_sample((num_samples, batch['actions’].shape[-1]),min=-1, max=1) | |
| random-pi = O.5 ** batch['actions’].shape[-1] | |
| dist = actor(batch['observations']) pi-actions,iog-pis = dist.sample(num_samples) | |
| rho_dist = rho(batch['observations']) | |
| rho_actions,log-probs_rho = rho_dist.sample(num_samples) | |
| q-rand_is = critic(batch['observations'], | |
| random-actions)- random-pi | |
| q-pi_is = critic(batch['observations'], | |
| q-rho_is = critic(batch['observations'], | pi_actions)- log-pis |
| rho_actions)- log-probs_rho | |
| cat-q = logsumexp(cat-q, axis=-1) | cat-q = concatenate(q-rand_is,q-pi_is, new_axis=True) |
| cat-q-rho = logsumexp(q-rho_is , axis=-1) | |
| # average between rho and pi | |
| push_down_term_reds = O.5 * (cat-q + cat_q-rho) | |
| reds_loss = td_loss +\\ | |
| ((push_down_term_reds - q-data).mean() * cql_alpha) | |
| critic_optimizer.zero-grad() | |
| reds_loss .backward() | |
| critic_optimizer.step() | |
| # Identical to CQL # return distribution of actions dist = actor(batch['observations']) # sample actions with associated log probabilities |
| # AWR style update to find rho # sample policy actions for advantage calculation |
| dist = actor(batch['observations']) pi_actions,log-pis = dist.sample() # calculate advantage qdata = critic(batch['observations'],batch['actions’]) value = critic(batch['observations'], actions) |
| advantage = (qdata - value.min(O)).mean() |
| # awr style clipping clipped_advantage = clip(advantage/self.temperature,\ min=-10,max=5) # find log rho(a|s) rho_dist = _rho(batch['observations']) log-prob_rho = rho-dist.log-prob(batch['actions']) |
| Hyperparameters | Values |
| N | 10,20,50,100 |
| m | 0,1,5,10,50,100,1000 |
| Hyperparameters | Values |
| Kernel Type | Laplacian, Gaussian |
| 0 | 1, 10, 20, 50 |
| Method | Scale Augmentation Set | Easy | Medium | Hard |
| ResNet-2.5GF BFBox-2.5GF (Liu& Tang,2020) Evolutionary-2.5GF | [0.3, 1.0] [0.3, 1.0] [0.3,1.0] | 91.87 92.22 92.30 | 89.49 90.19 90.21 | 67.32 69.41 69.62 |
| CR@backbone CR@detector CR@two-steps (CRFD-2.5GF) | [0.3,1.0] [0.3, 1.0] [0.3, 1.0] | 92.32 92.61 92.66 | 90.25 90.74 90.72 | 69.78 70.98 71.37 |
| ResNet-2.5GF ResNet-2.5GF | [0.3,2.0] SR | 93.21 93.17 | 91.11 91.14 | 74.47 74.93 |
| CR@two-steps CR@tw0-steps (SCRFD-2.5GF) | [0.3,2.0] SR | 93.78 93.76 | 92.16 92.17 | 77.87 78.35 |
| Method | Backbone | Easy | Medium | Hard | #Params(M) | #Flops(G) | Infer(ms) |
| DSFD@VGA | ResNet152 | 94.29 | 91.47 | 71.39 | 120.06 | 259.55 | 55.6 |
| DSFD @Multi-Scale | ResNet152 | 96.6 | 95.7 | 90.4 | 120.06 | 15928.5 | |
| RetinaFace@VGA | ResNet50 | 94.92 | 91.90 | 64.17 | 29.50 | 37.59 | 21.7 |
| RetinaFace@Multi-Scale | ResNet50 | 96.7 | 96.1 | 91.4 | 29.50 | 4585.98 | |
| BFBox@VGA | 1 | 94.2 | 92.1 | 70.4 | 28.6 | 39.4 | 22.4 |
| BFBox @Multi-Scale | - | 96.5 | 95.7 | 91.7 | 28.6 | 4732.8 | = |
| HAMBox @VGA | ResNet50 | 95.27 | 93.76 | 76.75 | 30.24 | 43.28 | 25.9 |
| HAMBox @Multi-Scale | ResNet50 | 97.0 | 96.4 | 93.3 | 30.24 | 5246.23 | = |
| TinaFace@VGA | ResNet50 | 95.61 | 94.25 | 81.43 | 37.98 | 172.95 | 38.9 |
| TinaFace@Multi-Scale | ResNet50 | 97.0 | 96.3 | 93.4 | 37.98 | 42333.64 | ■ |
| ResNet-34GF@VGA | ResNet50 | 95.64 | 94.22 | 84.02 | 24.81 | 34.16 | 11.8 |
| CRFD-34GF@VGA | Bottleneck Res | 96.06 | 94.92 | 85.29 | 9.80 | 34.13 | 11.7 |
| SCRFD-34GF@VGA | Bottleneck Res | 96.05 | 94.96 | 86.21 | 9.80 | 34.13 | 11.7 |
| SCRFD-34GF@Multi-Scale | Bottleneck Res | 97.20 | 96.58 | 93.53 | 9.80 | 2098.98 | - |
| ResNet-10GF@VGA | ResNet34x0.5 | 94.69 | 92.90 | 80.42 | 6.85 | 10.18 | 6.3 |
| CRFD-10GF@VGA | Basic Res | 95.16 | 93.87 | 83.05 | 3.86 | 9.98 | 4.9 |
| SCRFD-10GF@VGA | Basic Res | 95.14 | 93.96 | 83.43 | 3.86 | 9.98 | 4.9 |
| SCRFD-10GF@Multi-Scale | Basic Res | 95.93 | 94.95 | 90.81 | 3.86 | 614.14 | 1 |
| ResNet-2.5GF@VGA | ResNet34x0.25 | 93.21 | 91.11 | 74.47 | 1.62 | 2.57 | 5.4 |
| CRFD-2.5GF@VGA | Basic Res | 93.78 | 92.16 | 77.87 | 0.67 | 2.53 | 4.2 |
| SCRFD-2.5GF@VGA | Basic Res | 93.76 | 92.17 | 78.35 | 0.67 | 2.53 | 4.2 |
| SCRFD-2.5GF@Multi-Scale | Basic Res | 95.21 | 94.44 | 89.92 | 0.67 | 155.69 | - |
| Method | Backbone | Easy | Medium | Hard | #Params(M) | #Flops(G) | Infer(ms) |
| RetinaFace@VGA | MobileNet0.25 | 87.78 | 81.16 | 47.32 | 0.44 | 0.802 | 7.9 |
| RetinaFace @Original | MobileNet0.25 | 89.58 | 87.11 | 69.12 | 0.44 | 2.358 | - |
| RetinaFace@Multi-Scale | MobileNet0.25 | 91.4 | 89.2 | 82.5 | 0.44 | 49.28 | - |
| FaceBoxes @VGA | = | 76.17 | 57.17 | 24.18 | 1.01 | 0.275 | 2.5 |
| FaceBoxes @Original | 84.5 | 77.7 | 40.4 | 1.01 | 0.809 | - | |
| FaceBoxes @Multi-Scale | 85.9 | 81.6 | 55.7 | 1.01 | 16.93 | - | |
| libfacedetection @ Original | 85.6 | 84.2 | 72.7 | 2.33 | 3.25 | - | |
| LFFD@Original | 91.0 | 88.0 | 77.8 | 2.15 | 27.20 | - | |
| MobileNet-1.0GF@VGA | MobileNet0.25 | 91.66 | 89.28 | 70.46 | 0.63 | 1.024 | 4.9 |
| CRFD-1.0GF@VGA | Depth-wise Conv | 92.38 | 90.57 | 74.80 | 0.64 | 0.982 | 4.1 |
| SCRFD-1.0GF@VGA | Depth-wise Conv | 92.36 | 90.58 | 76.03 | 0.64 | 0.982 | 4.1 |
| SCRFD-1.0GF@ Original | Depth-wise Conv | 91.89 | 89.96 | 84.70 | 0.64 | 2.89 | 1 |
| SCRFD-1.0GF@Multi-Scale | Depth-wise Conv | 93.87 | 92.99 | 88.74 | 0.64 | 60.39 | - |
| MobileNet-0.5GF@VGA | MobileNet0.25 | 90.38 | 87.05 | 66.68 | 0.37 | 0.507 | 3.7 |
| CRFD-0.5GF@VGA | Depth-wise Conv | 90.57 | 88.12 | 68.51 | 0.57 | 0.508 | 3.6 |
| SCRFD-0.5GF@VGA | Depth-wise Conv | 90.80 | 88.43 | 68.82 | 0.57 | 0.508 | 3.6 |
| SCRFD-0.5GF@Original | Depth-wise Conv | 90.35 | 88.21 | 81.46 | 0.57 | 1.49 | - |
| SCRFD-0.5GF@Multi-Scale | Depth-wise Conv | 92.71 | 91.45 | 86.23 | 0.57 | 31.24 | - |
| Algorithm1: Search algorithm for computation redistribution |
| Input: Constraint of computation cost Y (in GFlops); Number of random network architectures N; Dataset for training Dtrain and validation Dual; Evaluation metric AP. Output: Best architecture A* Initialize the architecture set A = O whilelength(A)<Ndo |
| net=RandomSampling({di,wi}); /* di and wi denote block number and channel number,i =2,3,4,5. */ if net.Flops≤1.02 *Y And net.Flops≥ 0.98 *Y then |A.Append(net) end |
| end ParallelTrain(A,Dtrain) CR1= Bootstrap(A,APs) | APs = Evaluate(A,Dvat) Initialize the architecture set A= |
| while length(A)<N do net= RandomSampling({CRl,n,m,h}); |
| /* n,m,and h denote channel in neck, block and channel in head. */ |
| if net.Flops ≤1.02 *Y And net.Flops ≥ 0.98 *Y then |
| |A.Append(net) end |
| Testing Scale | AP | #Flops(G) |
| Multi-scale | 0.934 | 42333.64 |
| 1650 | 0.930 | 1021.82 |
| Original(1024) | 0.914 | 508.47 |
| 640 | 0.814 | 172.95 |
| Name | Conv Type | Stem | Backbone Depth | Backbone Width | Neck | Head |
| ResNet-34GF | Bottleneck Res | 256 | [3,4,6,3] | [256,512,1024,2048] | 128 | [256,256] |
| CRFD-34GF | Bottleneck Res | 56 | [17,16,2,8] | [56,56,144,184] | 128 | [256,256] |
| ResNet-10GF | Basic Res | 32 | [3,4,6,3] | [32,64,128,256] | 128 | [160,160] |
| CRFD-10GF | Basic Res | 56 | [3,4,2,3] | [56,88,88,224] | 56 | [80,80,80] |
| ResNet-2.5GF | Basic Res | 16 | [3,4,6,3] | [16,32,64,128] | 48 | [96,96] |
| CRFD-2.5GF | Basic Res | 24 | [3,5,3,2] | [24,48,48,80] | 24 | [64,64] |
| MobileNet-1.0GF | Depth-wise Conv | 16 | [3,3,7,3] | [32,64,128,256] | 64 | [128,128] |
| CRFD-1.0GF | Depth-wise Conv | 48 | [3,2,1,5] | [48,160,216,312] | 24 | [96,96] |
| MobileNet-0.5GF | Depth-wise Conv | 16 | [2,2,6,3] | [32,64,128,256] | 32 | [80,80] |
| CRFD-0.5GF | Depth-wise Conv | 16 | [2,3,2,6] | [40,72,152,288] | 16 | [64,64] |
| Methods | AFW | PASCAL | FDDB |
| BFBox (Liu & Tang,2020) HAMBox (Liu et al., 2020) | 99.68 99.90 | 99.43 99.50 | 98.9 99.10 |
| SCRFD-34GF | 99.945 | 99.597 | 99.25 |
| SCRFD-10GF | 99.900 | 99.461 | 99.07 |
| SCRFD-2.5GF | 99.821 | 98.911 | 99.02 |
| SCRFD-1.0GF | 99.696 | ||
| 98.601 | 98.69 | ||
| SCRFD-0.5GF | 98.603 | 98.537 | 98.14 |
| Model | Backbone | Method | 50 | 51 | 52 | 53 | mean | FB-IoU |
| OSLSM co-FCN AMP-2 | VGG16 | 1-shot 1-shot 1-shot | 33.6 36.7 41.9 | 55.2 50.6 50.2 | 40.9 44.9 46.7 | 33.5 32.4 34.7 | 40.8 41.1 43.4 | 61.3 60.1 61.9 |
| PANet PGNet | ResNet50 | 1-shot 1-shot | 44.0 56.0 | 57.5 66.9 | 50.8 50.6 | 44.0 50.4 | 49.1 56.0 | - 69.9 |
| FWB PPNet DAN PFENet RePRI | ResNet101 | 1-shot 1-shot 1-shot 1-shot 1-shot | 51.3 52.7 54.7 60.5 59.6 | 64.5 62.8 68.6 69.4 68.6 | 56.7 57.4 57.8 54.4 62.2 | 52.2 47.7 51.6 55.9 47.2 | 56.2 55.2 58.2 60.1 | 1 70.9 71.9 72.9 |
| HSNet SPNet ZS3Net | ResNet101 | 1-shot zero-shot zero-shot | 67.3 23.8 40.8 | 72.3 17.0 39.4 | 62.0 14.1 39.3 | 63.1 18.3 33.6 | 66.2 18.3 38.3 | 77.6 44.3 57.7 |
| LSeg LSeg | ResNet101 ViT-L/16 | zero-shot zero-shot | 52.8 61.3 | 53.8 63.6 | 44.4 43.1 | 38.5 41.0 | 47.4 52.3 | 64.1 67.0 |
| Model | Backbone | Method | 200 | 201 | 20² | 203 | mean | FB-IoU |
| PPNet PMM RPMM RePRI | ResNet50 | 1-shot 1-shot 1-shot 1-shot | 28.1 29.3 29.5 | 30.8 34.8 36.8 | 29.5 27.1 28.9 | 27.7 27.3 27.0 | 29.0 29.6 30.6 | |
| FWB DAN PFENet | ResNet101 | 1-shot 1-shot 1-shot | 32.0 17.0 1 36.8 | 38.7 18.0 1 41.8 | 32.7 21.0 - 38.7 | 33.1 28.9 1 36.7 | 34.1 21.2 24.4 38.5 | 1 62.3 |
| HSNet ZS3Net | ResNet101 | 1-shot zero-shot | 37.2 18.8 | 44.1 20.1 | 42.4 24.8 | 41.3 20.5 | 41.2 21.1 | 63.0 69.1 |
| LSeg LSeg | ResNet101 ViT-L/16 | zero-shot zero-shot | 22.1 28.1 | 25.1 27.5 | 24.9 30.0 | 21.5 23.2 | 23.4 27.2 | 55.1 57.9 59.9 |
| Model Backbone Method mIoU | |||
| OSLSMGNetFSSDoG-LSTM | 1-shot | 70.3 | |
| VGG16 | 1-shot1-shot1-shot | 71.973.580.8 | |
| DANHSNet | ResNet101 | 1-shot1-shot | 85.286.5 |
| LSegLSeg | ResNet101ViT-L/16 | zero-shotzero-shot | 84.787.8 |
| Block Type | Metric | #depth | |||
| 0 | 1 | 2 | 4 | ||
| DepthwiseBlock | pixAcc [%] | 79.70 | 79.72 | 79.78 | 7.67 |
| mIoU[%] | 37.83 | 39.19 | 39.45 | 0.18 | |
| pixAcc [%] | 79.70 | 79.64 | 79.70 | 79.68 | |
| BottleneckBlock | mIoU[%] | 37.83 | 39.16 | 39.79 | 38.78 |
| Method | Backbone | Text Encoder (fixed) | embedding dimension | pixAcc [%] | mIoU [%] |
| LSeg | ViT-B/32 | ViT-B/32 | 512 | 79.70 | 37.83 |
| LSeg | ViT-B/32 | ViT-B/16 | 512 | 79.77 | 38.69 |
| LSeg | ViT-B/32 | RN50×4 | 640 | 79.85 | 38.93 |
| LSeg | ViT-B/32 | RN50 ×16 | 768 | 80.26 | 40.36 |
| Method | Backbone | Text Encoder | pixAcc [%] | mIoU[%] |
| OCNet | ResNet101 | 45.45 | ||
| ACNet | ResNet101 | 81.96 | 45.90 | |
| DeeplabV3 | ResNeSt101 | 82.07 | 46.91 | |
| DPT | ViT-L/16 | 82.70 | 47.63 | |
| LSeg | ViT-L/16 | ViT-B/32 | 82.46 | 46.28 |
| LSeg | ViT-L/16 | RN50 ×16 | 82.78 | 47.25 |
| #Passages | #Sentences | #Tokens/passage |
| 238 | 1908 | 184.7±36.9 |
| Method | Sentence-level (AUC-PR) | Passage-level (Corr.) Pearson Spearman | |||
| NonFact | NonFact* | Factual | |||
| Random | 72.96 | 29.72 | 27.04 | ||
| GPT-3 (text-davinci-003)'s probabilities (LLM, grey-box) | |||||
| Avg(-logp) | 83.21 | 38.89 53.97 | 57.04 | 53.93 | |
| Avg(H)t | 80.73 | 37.09 | 52.07 55.52 | 50.87 | |
| Max(-logp) | 87.51 | 35.88 | 50.46 57.83 | 55.69 | |
| Max(H)t | 85.75 | 32.43 | 50.27 52.48 | 49.55 | |
| LLaMA-30B's probabilities (Proxy LLM, black-box) | |||||
| Avg(-logp) | 75.43 | 30.32 41.29 | 21.72 | 20.20 | |
| Avg(H) | 80.80 | 39.01 | 42.97 33.80 | 39.49 | |
| Max(-logp) | 74.01 | 27.14 31.08 | -22.83 | -22.71 | |
| Max(H) | 80.92 | 37.32 37.90 | 35.57 | 38.94 | |
| SelfCheckGPT (black-box) | |||||
| w/BERTScore | 81.96 | 45.96 | 44.23 | 58.18 | 55.90 |
| w/ QA | 84.26 | 40.06 | 48.14 | 61.07 | 59.29 |
| w/ Unigram (max) | 85.63 | 41.04 | 58.47 | 64.71 | 64.91 |
| w/ NLI | 92.50 | 45.17 | 66.08 | 74.14 | 73.78 |
| w/ Prompt | 93.42 | 53.19 | 67.09 | 78.32 | 78.30 |
| Method | Sent-lvl AUC-PR NoFac NoFac* | Fact | Passage-lvl Pear. Spear. | |
| SelfCk-BERT | 81.96 | 45.96 | 44.23 | 58.18 55.90 |
| WikiBio+BERT | 81.32 | 40.62 | 49.15 | 58.71 55.80 |
| SelfCk-QA | 84.26 | 40.06 | 48.14 | 61.07 59.29 |
| WikiBio+QA | 84.18 | 45.40 | 52.03 | 57.26 53.62 |
| SelfCk-1gm | 85.63 | 41.04 | 58.47 | 64.71 64.91 |
| WikiBio+1gm | 80.43 | 31.47 | 40.53 | 28.67 26.70 |
| SelfCk-NLI | 92.50 | 45.17 | 66.08 | 74.14 73.78 |
| WikiBio+NLI | 91.18 | 48.14 | 71.61 | 78.84 80.00 |
| SelfCk-Prompt | 93.42 | 53.19 | 67.09 | 78.30 |
| WikiBio+Prompt | 93.59 | 65.26 | 73.11 | 78.32 85.90 86.11 |
| Text-Gen | SelfCk-Prompt | N | Pear. | Spear. |
| GPT-3 | ChatGPT | 20 | 78.32 | 78.30 |
| GPT-3 | ChatGPT | 4 | 76.47 | 76.41 |
| GPT-3 | GPT-3 | 4 | 73.11 | 74.69 |
| + SelfCheck w/ unigram (max) | 20 | 64.71 | 64.91 | |
| + SelfCheck w/ NLI | 20 | 74.14 | 73.78 | |
| Varaint | Sentence-lvl | Passage-lvl | |||
| NoF | NoF* | Fact | PCC | SCC | |
| SimpleCount | 83.97 | 40.07 | 47.78 | 57.39 | 55.15 |
| + Bayes | 83.04 | 38.58 | 47.41 | 56.43 | 55.03 |
| + Bayes + α | 84.26 | 40.06 | 48.14 | 61.07 | 59.29 |
| ChatGPT | Yes | No |
| GPT-3 | ||
| Yes | 3179 | 1038 |
| No | 367 | 3048 |
| n-gram | Sent-lvl AUC-PR | Passage-lvl | |||
| NoFac | NoFac* | Fact | Pear. | Spear. | |
| Avg(-logp) | |||||
| 1-gram | 81.52 82.94 83.56 | 40.33 44.38 | 41.76 53.99 | 40.68 58.84 | 39.22 |
| 2-gram | 52.81 | 58.11 | |||
| 3-gram | 44.64 43.55 | 62.21 63.00 | |||
| 4-gram 83.80 5-gram 83.45 | 54.25 53.98 | 61.98 63.64 60.68 62.96 | |||
| 42.31 | |||||
| Max(-logp) 1-gram | 85.63 | 41.04 | 58.47 | 64.71 | 64.91 |
| 2-gram | 85.26 | 39.29 | 58.29 | 62.48 | 66.04 |
| 3-gram | 84.97 | 37.10 | 57.08 | 57.34 | 60.49 |
| 4-gram | 84.49 | 36.37 | 55.96 | 55.77 | 57.25 |
| 5-gram | 84.12 | 36.19 | 54.89 | 54.84 | 55.97 |
| LLM | Size | Sentence-level (AUC-PR) | Passage-level (Corr.) | |||
| NonFact | NonFact* | Factual | Pearson | Spearman | ||
| Random | 72.96 | 29.72 | 27.04 | |||
| Avg(-logp) Method | ||||||
| LLaMA | 30B | 75.43 | 30.32 | 41.29 | 21.72 | 20.20 |
| LLaMA | 13B | 74.16 | 30.01 | 37.36 | 13.33 | 12.89 |
| LLaMA | 7B | 71.69 | 27.87 | 31.30 | -2.71 | -2.59 |
| OPT | 30B | 67.70 | 24.43 | 25.04 | -32.07 | -31.45 |
| NeoX | 20B | 69.00 | 24.38 | 26.18 | -31.79 | -34.15 |
| OPT | 13B | 67.46 | 24.39 | 25.20 | -33.05 | -32.79 |
| GPT-J | 6B | 67.51 | 24.28 | 24.26 | -38.80 | -40.05 |
| OPT | 1.3B | 66.19 | 24.47 | 23.47 | -35.20 | -38.95 |
| OPT | 125m | 66.63 | 25.31 | 23.07 | -30.38 | -37.54 |
| Avg(H) Method | ||||||
| LLaMA | 30B | 80.80 | 39.01 | 42.97 | 33.80 | 39.49 |
| LLaMA | 13B | 80.63 | 38.98 | 40.59 | 29.43 | 33.12 |
| LLaMA | 7B | 78.67 | 37.22 | 33.81 | 19.44 | 21.79 |
| OPT | 30B | 77.13 | 33.67 | 29.55 | -0.43 | 3.43 |
| NeoX | 20B | 77.40 | 32.78 | 30.13 | 5.41 | 7.43 |
| OPT | 13B | 76.93 | 33.71 | 29.68 | 0.25 | 1.39 |
| GPT-J | 6B | 76.15 | 33.29 | 28.30 | -2.50 | -1.37 |
| OPT | 1.3B | 74.05 | 31.91 | 26.33 | -10.59 | -10.00 |
| OPT | 125m | 71.51 | 30.88 | 25.36 | -14.16 | -13.76 |
| Max(-logp) Method | ||||||
| LLaMA | 30B | 74.01 | 27.14 | 31.08 | -22.83 | -22.71 |
| LLaMA | 13B | 71.12 | 26.78 | 28.82 | -34.93 | -31.70 |
| LLaMA | 7B | 69.57 | 25.91 | 26.54 | -42.57 | -38.24 |
| OPT | 30B | 67.32 | 24.40 | 24.32 | -49.51 | -45.50 |
| NeoX | 20B | 67.51 | 23.88 | 24.82 | -47.96 | -44.54 |
| OPT | 13B | 67.36 | 24.67 | 24.46 | -50.15 | -44.42 |
| GPT-J | 6B | 67.58 | 23.94 | 23.93 | -51.23 | -47.68 |
| OPT | 1.3B | 68.16 | 25.85 | 24.66 | -45.60 | -42.39 |
| OPT | 125m | 69.23 | 27.66 | 24.14 | -39.22 | -37.18 |
| Max(H) Method | ||||||
| LLaMA | 30B | 80.92 | 37.32 | 37.90 | 35.57 | 38.94 |
| LLaMA | 13B | 80.98 | 37.94 | 36.01 | 32.07 | 34.01 |
| LLaMA | 7B | 79.65 | 35.57 | 31.32 | 22.10 | 22.53 |
| OPT | 30B | 76.58 | 33.44 | 29.31 | 1.63 | 6.41 |
| NeoX | 20B | 76.98 | 31.96 | 29.13 | 5.97 | 9.31 |
| OPT | 13B | 76.26 | 32.81 | 29.25 | 1.42 | 2.82 |
| GPT-J | 6B | 75.30 | 32.51 | 28.13 | -2.14 | 1.41 |
| OPT | 1.3B | 73.79 | 31.42 | 26.38 | -9.84 | -9.80 |
| OPT | 125m | 71.32 | 31.65 | 25.36 | -18.05 | -17.37 |
| Method | Split CIFAR-100 | 5-datasets | ||
| Average Acc (↑) | Forgetting (↓) | Average Acc (↑) | Forgetting (↓) | |
| Upper bound: | ||||
| FT-iid | 90.85±0.12 | 93.93±0.18 | ||
| Non-rehearsal based methods: | ||||
| FT-seq-frozen | 17.72±0.34 | 59.09±0.25 | 39.49±0.12 | 42.62±0.20 |
| FT-seq | 33.61±0.85 | 86.87±0.20 | 20.12±0.42 | 94.63±0.68 |
| EWC LwF | 47.01±0.29 | 33.27±1.17 | 50.93±0.09 | 34.94±0.07 |
| L2P (ours) | 60.69±0.63 | 27.77±2.17 | 47.91±0.33 | 38.01±0.28 |
| 83.83±0.04 | 7.63±0.30 | 81.14 ±0.93 | 4.64 ±0.52 | |
| Rehearsal based methods: | ||||
| ER | 82.53±0.17 | 16.46±0.25 | 89.30±0.94 | 8.08±0.53 |
| GDumb | 81.67±0.02 | - | 70.76±0.12 | 1 |
| L2P-R (ours) | 86.31±0.59 | 5.83±0.61 | 91.92±0.78 | 3.34±0.71 |
| Category | Method | Test Acc (↑) |
| Upper bound | FT-iid | 90.85±0.12 |
| Rehearsal | ER | 82.53±0.17 |
| GDumb | 81.67±0.02 | |
| Non-rehearsal | EWC | 63.04±0.42 |
| LwF | 69.46±0.35 | |
| L2P (ours) | 88.34±0.14 |
| Category | Method | Test Acc (↑) |
| Upper bound | FT-iid | 82.15 ±0.37 |
| Rehearsal | ER | 80.10±0.56 |
| GDumb | 74.92±0.25 | |
| Non-rehearsal | EWC | 74.82±0.60 |
| LwF | 75.45±0.40 | |
| L2P (ours) | 78.33±0.06 |
| Method | 5-datasets | |
| Average Acc (↑) | Forgetting (↓) | |
| L2P without prompt pool | 51.96 | 26.60 |
| L2P without key-value pair | 58.33 | 20.45 |
| L2P without diversified prompt selection | 62.26 | 17.84 |
| L2P | 81.14 | 4.64 |
| GPT-3.5 | CHATGPT | GPT-4 | ||||
| Task | Base | +SELF-REFINE | Base | +SELF-REFINE | Base | +SELF-REFINE |
| Sentiment Reversal | 8.8 | 30.4 (↑21.6) | 11.4 | 43.2 (↑31.8) | 3.8 | 36.2 (↑32.4) |
| Dialogue Response | 36.4 | 63.6 (127.2) | 40.1 | 59.9 (↑19.8) | 25.4 | 74.6 (↑49.2) |
| Code Optimization | 14.8 | 23.0 (↑8.2) | 23.9 | 27.5 (13.6) | 27.3 | 36.0 (↑8.7) |
| Code Readability | 37.4 | 51.3 (↑13.9) | 27.7 | 63.1 (↑35.4) | 27.4 | 56.2 (↑28.8) |
| Math Reasoning | 64.1 | 64.1 (0) | 74.8 | 75.0 (↑0.2) | 92.9 | 93.1 (10.2) |
| Acronym Generation | 41.6 | 56.4 (↑14.8) | 27.2 | 37.2 (↑10.0) | 30.4 | 56.0 (↑25.6) |
| Constrained Generation | 16.0 | 39.7 (↑23.7) | 2.75 | 33.5 (130.7) | 4.4 | 61.3 (156.9) |
| Task | SELF-REFINE feedbackGeneric feedbackNo feedback | ||
| Code Optimization | 27.5 | 26.0 | 24.8 |
| Sentiment Reversal | 43.2 | 31.2 | 0 |
| Acronym Generation | 56.4 | 54.0 | 48.0 |
| Task | y0 | Y1 | 92 | y3 |
| C.Opt. | 22.0 | 27.0 | 27.9 | 28.8 |
| S.Rev. | 33.9 | 34.9 | 36.1 | 36.8 |
| C. Gen. | 12.1 | 26.1 | 39.6 | 46.1 |
| Supervision- free refiner | Supervision- free feedback | Multi-aspect feedback | Iterative | |
| Learned refiners: PEER (Schick et al., 2022b), Self-critique (Saunders et al.,2022b), CodeRL (Le et al.,2022b), Self-correction (Welleck et al., 2022). | × | □orX | × | √orX |
| Prompted refiners: Augmenter (Peng et al.. 2023), Re³ (Yang et al., 2022), Reflexion (Shinn et al.,2023). | √ | □orX | X | X |
| SELF-REFINE (this work) | √ | √ | √ | √ |
| Task and Description | Sample one iteration of FEEDBACK-REFINE |
| SentimentReversal | x: The food was fantastic..." |
| Rewrite reviews to reverse sentiment. Dataset: (Zhang et al., 2015) 1000 review pas- | yt: The food was disappointing...” fb:Increase negative sentiment |
| sages Dialogue Response Generation | Yt+1: The food was utterly terrible.." x: What's the best way to cook pasta?" |
| Produce rich conversational responses. | Yt: The best way to cook pasta is to..." |
| Dataset: (Mehri and Eskenazi, 202O) 372 conv. | fb: Make response relevant, engaging, safe |
| Yt+1: Boil water, add salt, and cook pasta..." | |
| Code Optimization Enhance Python code efficiency | x:Nested loop for matrix product |
| Dataset: (Madaan et al.,2023): 1000 programs | Yt: NumPy dot product function |
| fb:Improve time complexity | |
| Yt+1: Use NumPy's optimized matmul function | |
| CodeReadabilityImprovement Refactor Python code for readability. | x:Unclear variable names,no comments |
| Dataset: (Puri et al., 2021) 300 programs* | Yt: Descriptive names, comments fb:Enhance variable naming;add comments |
| Yt+1: Clear variables,meaningful comments | |
| Math Reasoning | x: Olivia has $23,buys 5 bagels at $3 each" |
| Solve math reasoning problems. | Yt:Solution in Python |
| Dataset: (Cobbe et al., 2021) 1319 questions | fb: Show step-by-step solution |
| Yt+1: Solution with detailed explanation | |
| Acronym Generation Generate acronyms for a given title | x : Radio Detecting and Ranging" |
| Dataset:(Appendix T) 25O acronyms | Yt:RDR fb :be context relevant; easy pronunciation |
| Yt+1: RADAR" | |
| Constrained Generation | |
| x:beach,vacation,relaxation | |
| Generate sentences with given keywords. | Yt: During our beach vacation.. |
| Dataset: (Lin et al., 2020) 200 samples | fb:Include keywords;maintain coherence |
| Method | Primary Novelty | zero/few shot improvement | multi aspect critics | NL feedback with er- ror localization | iterative framework |
| RLHF (Stiennon et al., 2020) Rainier RL (Liu et al., 2022) QUARK RL (Lu et al., 2022) Code RL (Le et al., 2022a) | optimize for human preference RL to generate knowledge quantization to edit generations actor critic RL for code im- | Xtrained on feedback Xtrained on end task × trained on end task Xtrained on end task | X single (human) X single(accuracy) × single(scalar score) × single(unit tests) | √(not self gen.) X(knowl. only) X(dense signal) X(dense signal) | × |
| DrRepair (Yasunaga and Liang, 2020) PEER (Schick et al., 2022b) | provement Compiler feedback to itera- tively repair doc. edit trained on wiki edits | Xtrained semi sup. Xtrained on edits | X single(compiler msg) Xsingle(accuracy) | (not self gen.) (not self gen.) | √ √ |
| Self critique (Saunders et al., 2022a) Self-correct (Welleck et al., 2022) Const. AI (Bai et al., 2022b) | few shot critique generation novel training of a corrector train RL4F on automat (cri- tique,revision) pair | Xfeedback training trained on end task | X single(human) single (task specific) (fixed set) | (self gen.) √(limited setting) √ | × (limited setting) × |
| Self-ask (Press et al., 2022) | ask followup ques when in- terim ans correct;final wrong | few shot | X none | X(none) | × |
| GPT3 score (Fu et al., 2023) | GPT can score generations with instruction | √few shot | X single(single utility fn) | X(none) | × |
| Augmenter (Peng et al., 2023) | factuality feedback from exter- | few shot | X single(factuality) | (self gen.) | √ |
| Re3 (Yang et al., 2022) | nal KBs ~ours: but one domain, | √few shot | (trained critics) | (not self gen.) | √ |
| SELF-REFINE | trained critics fewshot iterative multi aspect NL fb | few shot | √multiple(few shot critics) | (self gen.) | √ |
| Task | SELF-REFINE(%) | Direct (%) | Either (%) |
| SentimentReversal | 75.00 | 21.43 | 3.57 |
| Acronym Generation | 44.59 | 12.16 | 43.24 |
| Dialogue Response Generation | 47.58 | 19.66 | 32.76 |
| Code Readability Improvement | 50.0 | 3.0 | 47.0 |
| Task | Base model | +Self-Refine | Gain |
| Date Understanding | 62.0 | 66.8 | 4.8 |
| Geometric Shapes | 17.6 | 20.0 | 2.4 |
| Logical Deduction (seven objects) | 43.2 | 45.2 | 2.0 |
| Multi-Step Arithmetic | 61.6 | 64.0 | 2.4 |
| Tracking Shuffled Objects (seven objects) | 31.6 | 36.0 | 4.4 |
| Task | Base | SELF-REFINE (zerO-shot) |
| Acronym Generation | 16.6% | 44.8% (↑ 28.2%) |
| Constrained Generation | 4.0% | 47.0% (↑ 43.0%) |
| Sentiment Reversal | 4.4% | 71.4% (↑ 67.0%) |
| Math Reasoning | 22.1% | 59.0% (↑ 36.9%) |
| Dialogue Response Generation | 23.0% | 48.8% (↑ 25.8%) |
| gen_sys_template = "You are a helpful assistant that responds with only a python | Function to Generate an Answer | |||||
| 「 program." gen_user_template = """Write a python function that solves the given question and | ||||||
| → | returns the result. | |||||
| Always store the final result in a variable called ‘result`,and always include | ||||||
| → | `return result` as your last statement. | |||||
| # Question: {question} | ||||||
| # solution using Python: """ | ||||||
| # Function to Get Feedback on the Answer | ||||||
| fb_sys_template = "You are a helpful assistant that provides feedback on the | ||||||
| correctness of python programs." fb_user_template = """Question: {question} | ||||||
| {prediction} | ||||||
| # There may be an error in the code above because of lack of understanding of the | ||||||
| 「 | question. | |||||
| To find the error,go through semantically complete blocks of the code,and check | ||||||
| → if everything looks good. | ||||||
| Errors could include: | ||||||
| - Logical issues | ||||||
| - Lack of understanding of the question | ||||||
| - Missing ‘return result` | ||||||
| Share your evaluation in the format: | ||||||
| Evaluation: <your evaluation here>. | ||||||
| If everything looks good, return 'Evaluation: correct '""" | ||||||
| # Function to Fix the Error | ||||||
| refine_sys_template = "You are a helpful assistant that responds with only a | ||||||
| → | python program." | |||||
| refine_user_template = """Question: {question} | ||||||
| {prediction} | ||||||
| # There is an error in the code above. The following is the feedback on the code: | ||||||
| {feedback} | ||||||
| Fix the error in the python function given the feedback above. The python function should return the final answer."" |
| Task | % Base | % SELF-REFINE |
| Dialogue Response Generation | 30.6 | 64.7 (↑34.1) |
| SentimentReversal | 10.6 | 69.2 (158.6) |
| Acronym Generation | 32.0 | 49.2 (↑17.2) |
| Code readability | 37.0 | 60.0 (↑23.0) |
| Method | Solve Rate | |
| Chen et al. (2021) | CODEX | 71.3 |
| Cobbe et al. (2021) Wei et al. (2022) | OpenAI 6B | 20.0 |
| CoTw/ CODEX | 65.6 | |
| Gao et al. (2022) | PaL w/ CODEX | 72.0 |
| PaL w/ GPT-3 | 52.0 | |
| PaL w/ GPT-3.5 | 56.8 | |
| PaL W/CHATGPT | 74.2 | |
| PaL W/GPT-4 | 93.3 | |
| Welleck et al. (2022) | Self-Correct w/ GPT-3 | 45.9 |
| Self-Correct (fine-tuned) | 24.3 | |
| This work | SELF-REFINE W/ GPT-3 | 55.7 |
| SELF-REFINE W/ GPT-3.5 | 62.4 | |
| SELF-REFINE W/ CHATGPT | 75.1 | |
| SELF-REFINEW/GPT-4 SELF-REFINE w/ CODEX (Oracle Feedback) | 94.5 | |
| Method | %OPT | |
| Puri et al. (2021) | Human References | 38.2 |
| OpenAI Models: OpenAI (2022, 2023) | CODEX GPT-3.5 CHATGPT | 9.7 14.8 22.2 27.3 |
| Nijkamp et al. (2022) | GPT-4 CODEGEN-16B | 1.1 |
| Berger et al. (2022) | SCALENE SCALENE (BEST@16) SCALENE (BEST@32) | 1.4 12.6 |
| Madaan et al. (2023) | PIE-2B PIE-2B (BEST@16) PIE-2B (BEST@32) | 19.6 4.4 21.1 26.3 |
| PIE-16B PIE-16B (BEST@ 16) PIE-16B (BEST@32) PIE-FeW-shot (BEST@16) PIE-Few-shot (BEST@32) | 4.4 22.4 26.6 35.2 38.3 | |
| This work | SELF-REFINE W/ CODEX SELF-REFINE W/ GPT-3.5 SELF-REFINEV W/ CHATGPT SELF-REFINE W/ GPT-4 | 15.6 23.0 26.7 36.0 |
| Method | %Readable Variables | |
| Chen et al. (2021) | CODEX | 37.4 |
| This work | SELF-REFINE W/CODEX | 51.3 |
| Task | Base | Self-Refine | Equally Good |
| Acronym Generation | 22.30 | 53.08 | 22.30 |
| Sentiment Reversal | 13.2 | 60.8 | 26 |
| Dialogue Response Generation | 11.2 | 20.4 | 54.6 |
| Math Reasoning | 37.6 | 37.8 (41 with Oracle) | N/A |
| GPT-3.5 | CHATGPT | GPT-4 | ||||
| Task | Base | +SELF-REFINE | Base | +SELF-REFINE | Base | +SELF-REFINE |
| Math Reasoning | 64.1 | 64.1(0) | 74.8 | 75.0 (10.2) | 92.9 | 93.1 (10.2) |
| Math Reasoning (Oracle) | 64.06 | 68.9 (↑4.8) | 74.8 | 76.2 (↑1.4) | 92.9 | 93.8 (10.7) |
| Iteration | Acronym | Pronunciation | Pron. (5) | Spell. (5) | Rel. (5) | Pos. Con. (5) | Total (25) |
| 1 | USTACCSF | us-tacks-eff | 1 | 1 | 5 | 3 | 11 |
| 2 | TACC-SIM | tacks-sim | 4 | 4 | 5 | 3 | 17 |
| 3 | TACCSF | tacks-eff | 1 | 2 | 5 | 3 | 12 |
| 4 | TACC-SIMF | tack-simf | 4 | 4 | 5 | 3 | 17 |
| Error Type | Occurrence | Feedback | Refinement |
| Incorrect feedback | 25% | Informative: The response provides some encourage- ment and reassurance, and additional information.3/3 | Response: :That's hilarious! Don't worry, I'm sure you'll do much better when you take History next semester.Are you looking forward to it? |
| Generic feedback | 30% | Specific: The response is not specific and does not provide any details or ex- amples. 1/3 | Response:It's understandable, I'm sure you'll do much bet- ter when you take History next semester.Maybe you can use your knowledge of Christo- pher Columbus to ace your next test! Are you looking for- ward to it? |
| Incorrect Scoring | 10% | Informative: The response provides a supportive and understanding message, which is informative to the user. 2/3 | Response:I'm sorry to hear.. when Christopher Columbus didn't even fight in the Civil War!I'm sure you'll do much better.... Are you looking for- ward to it? |
| Error Type | Occurrence | Feedback | Refinement |
| Not-Robust | 10% | Specific: The response is specific and provides spe- cific details or examples. 3/3 | Response: That's pretty funny!It's understand- able that you didn't do well on the test. I'm sure you'll do much bet- ter when you take History next semester |
| Ignores feedback | 25% | Informative: The re- sponse doesn't provide any new information. 1/3 | Response: That's so kind and generous of you! I would love to hear more about the podcasts you enjoy. Do you have any favorites you can recom- mend? |
| Introduces new problem20% | Informative:3/3 | Response: :That would be a disaster!Luckily, my coworkers know bet- ter than to drink while they work. New Informa- tive score: 2/3 | |
| Robusttobadfeedback | 60% | Informative: The re- sponse does not provide any information. 1/3 | Response: Wow, 25 peo- ple! That must have been an amazing experience. Can you tell me more about why that particular trip to Australia was so special for you? |
| Task | GPT-3.5 | CHATGPT | GPT-4 | |||
| Base | +SELF-REFINE | Base | +SELF-REFINE | Base | +SELF-REFINE | |
| Sentiment Reversal | 8.8 ± 2.05 | 30.4 ±3.61* | 11.4 ± 2.34 | 43.2 ±3.98* | 3.8±1.28 | 36.2 ±3.82* |
| Dialogue Response | 36.4 ± 6.14 | 63.6± 6.62* | 40.1 ± 6.33 | 59.9 ± 6.67* | 25.4 ± 5.36 | 74.6± 6.22* |
| Code Optimization | 14.8 ± 2.66 | 23.0±3.25* | 23.9 ±3.30 | 27.5±3.49 | 27.3 ±3.48 | 36.0±3.81* |
| Code Readability | 37.4 ± 6.86 | 51.3 ± 7.39 | 27.7 ± 6.13 | 63.1± 7.40* | 27.4 ± 6.10 | 56.2 ± 7.45* |
| Math Reasoning | 64.1 ± 3.47 | 64.1 ± 3.47 | 74.8 ±3.20 | 75.0±3.20 | 92.9 ± 2.05 | 93.1± 2.03 |
| Acronym Gen. | 41.6 ± 7.72 | 56.4±8.15 | 27.2 ±6.60 | 37.2 ± 7.46 | 30.4 ± 6.92 | 56.0±8.15* |
| Constrained Gen. | 28.0 ± 7.38 | 37.0±8.26 | 44.0 ±8.72 | 67.0±9.00* | 15.0 ± 5.38 | 45.0±8.77* |
| Meaningful Variable Ratio | Comment Per Line | Function Units | |
| Human Annotator Rewrites | 0.653 | 0.24 | 0.70 |
| SELF-REFINE (T= 0.0) | 0.628 | 0.12 | 1.41 |
| SELF-REFINE (T=0.7) | 0.700 | 0.25 | 1.33 |
| GPT-3.5 | ChatGPT | GPT4 | |
| SELF-REFINEwins | 36.0 | 48.0 | 54.0 |
| INIT wins | 23.0 | 18.0 | 16.0 |
| Both are equal | 41.0 | 50.0 | 30.0 |
| Setup | Iteration | % Optimized | Relative Speedup | Speedup |
| Direct | 1 | 9.7 | 62.29 | 3.09 |
| SELF-REFINE -feedback | 1 | 10.1 | 62.15 | 3.03 |
| SELF-REFINE - feedback | 2 | 10.4 | 61.79 | 3.01 |
| SELF-REFINE | 1 | 15.3 | 59.64 | 2.90 |
| SELF-REFINE | 2 | 15.6 | 65.60 | 3.74 |
| Criteria | output from GPT3: STSLWN | output from SELF-REFINE: Seq2Seq |
| Ease of pronunciation | Pronounced as ess-tee-ess-ell-double- you-enn which is very dificult. | Pronounced as seq-two-seq which is easy. |
| Ease of spelling | Very difficult to spell. | Easy to spell. |
| Relation to title | No relation to the title. | Mentions sequence which is somewhat related to the title. |
| Positive connotation | Meaningless acronym. | Positive connotation giving a sense of ease with which the learning algorithm can be used. |
| Well-known | Not a well-known acronym. | Close to the word sequence which is a well- known word. |
| Total score | 5/25 | 20/25 |
| Algorithm | Learned Embeddings | GradientSymmetry | Required Non-linearity |
| Clock | Circle | No | Multiplication |
| Pizza | Circle | Yes | Absolute value |
| Non-circular | Line,Lissajous-like curves, etc. | N/A | N/A |
| Circle | Wk | Ac FVE |
| #4 (accompanying #1) | 2π/59·17 | 97.56% |
| #5 (accompanying #2) | 2π/59.3 | 97.23% |
| #6 (accompanying #3) | 2π/59·44 | 97.69% |
| Dataset | Original gap | Modified gap | Direction |
| Coarse-grained Classification | |||
| CIFAR10 | 0.9013 | 0.9081 | → |
| CIFAR100 | 0.6658 | 0.6737 | ↓ |
| Fine-grained Classification | |||
| EuroSAT | 0.5410 | 0.5645 | ← |
| Optical Character Recognition | |||
| SVHN | 0.5389 | 0.5396 | → |
| HatefulMemes | 0.5800 | 0.5811 | 个 |
| Denigration Biases | Original gap | Modified gap | |||
| Crime related human | Non | Sum | Crime Non related human | Sum | |
| Black 1.0% 0.1% | 1.1% | 0.8% 0.1% | 1.0% | ||
| White 15.5% | 0.2% | 15.7% | 13.2% 0.4% | 13.7% | |
| Indian 1.2% | 0.0% | 1.2% | 1.1% 0.0% | 1.1% | |
| Latino 2.8% | 0.1% | 2.8% | 1.9% 0.1% | 2.0% | |
| Middle Eastern 6.3% | 0.0% | 6.3% | 5.2% 0.0% | 5.2% | |
| Southeast Asian 0.5% | 0.0% | 0.5% | 0.3% | 0.0% | 0.3% |
| East Asian 0.7% | 0.0% | 0.7% | 0.6% | 0.0% | 0.6% |
| Suite | Environment | BC | Filt. BC | TD3+BC | Onestep | CQL-r | CQL-p | DT | RvS-R | RvS-G | GCSL |
| D4RL AntMaze | umaze-v2 | 54.6 45.6 | 60.0 46.5 | 78.6 71.4 | 64.3 60.7 | 44.8 23.4 | 74.0 | 65.6 | 64.4 70.1 | 65.4 60.9 | |
| umaze-diverse-v2 | 84.0 | 51.2 | |||||||||
| medium-play-v2 | 0.0 | 42.1 | 10.6 | 0.3 | 0.0 | 61.2 | 1.0 | 4.5 | 58.1 | ||
| medium-diverse-v2 | 0.0 | 37.2 | 3.0 | 0.0 | 0.0 | 53.7 | 0.6 | 7.7 | 67.3 | ||
| large-play-v2 | 0.0 | 28.0 | 0.2 | 0.0 | 0.0 | 15.8 | 0.0 | 3.5 | 32.4 | ||
| large-diverse-v2 | 0.0 | 34.3 | 0.0 | 0.0 | 0.0 | 14.9 | 0.2 | 3.7 | 36.9 | ||
| antmze-v2 average | 16.7 | 41.4 | 27.3* | 20.9* | 11.4 | 50.6* | 19.8 | 25.6 | 53.5 | ||
| halfcheetah-random-v2 hopper-random-v2 | 2.3 | 2.0 | 11.0 | 6.9 | 18.6 | 2.2 | 3.9 7.7 | ||||
| 4.8 | 4.1 | 8.5 | 7.8 | 9.3 | 7.5 2.0 | ||||||
| walker2d-random-v2 | 1.7 | 1.7 | 1.6 | 6.1 | 2.5 | -0.2 | |||||
| random-v2 average | 2.9 | 2.6 | 6.9 | 3.9 | 3.8 | ||||||
| halfcheetah-medium-replay-v2 | 36.6 | 40.6 | 7.0 44.6 | 42.4 | 10.1 47.3 | 36.6 | 38.0 | ||||
| hopper-medium-replay-v2 | 18.1 | 75.9 | 60.9 | 71.0 | 97.8 | 82.7 | 73.5 | ||||
| walker2d-medium-replay-v2 | 26.0 | 62.5 | 81.8 | 71.6 | 86.1 | 66.6 | 60.6 | ||||
| medium-replay-v2 average | 26.9 | 59.7 | 62.4 | 61.7 | 77.1 | 62.0 | |||||
| halfcheetah-medium-v2 | 42.6 | 42.5 | 42.6 | 57.4 41.6 | |||||||
| hopper-medium-v2 | 52.9 | 56.9 | 48.3 59.3 | 55.6 83.3 | 49.1 64.6 | 67.6 | 60.2 | ||||
| walker2d-medium-v2 | 75.3 | 75.0 | 83.7 | 85.6 | 82.9 | 74.0 | 71.7 | ||||
| medium-v2 average | 56.9 | 58.1 | 74.8 | 65.5 | 61.4 | 57.8 | |||||
| halfcheetah-medium-expert-v2 | 55.2 | 92.9 | 63.8 90.7 | 93.5 | 85.8 | 86.8 | 92.2 | ||||
| hopper-medium-expert-v2 | 52.5 | 110.9 | 98.0 | 102.1 | 102.0 | 107.6 | 101.7 | ||||
| walker2d-medium-expert-v2 | 107.5 | 109.0 | 110.1 | 110.9 | 109.5 | 108.1 | 106.0 | ||||
| medium-expert-v2average | 71.7 | 104.3 | 99.6 | 102.2 | 99.1 | 100.8 | 100.0 | ||||
| gym-v2 average | 39.6 | 56.2 | 58.2 | 61.4 | 63.0 | 57.0 | 54.7 | ||||
| D4RL Kitchen | kitchen-complete-v0 | 65.0 | 4.0 | 11.8 | 43.8 | 1.5 | |||||
| kitchen-mixed-v0 | 51.5 | 40.0 | 24.2 | 51.0 | 1.1 | 50.2 60.3 | |||||
| kitchen-partial-v0 | 38.0 | 66.0 | 20.8 | 49.8 | 0.5 | 51.4 | |||||
| kitchen-vO average | 54.0 | ||||||||||
| 51.5 | 36.7 | 18.9 | 48.2 | 1.0 | 14.8 | 28.0 | |||||
| GCSL | claw door | 28.0 | |||||||||
| four rooms | 95.8 63.0 | 81.0 | |||||||||
| lunar | 54.7 | 70.0 | |||||||||
| pusher | 81.8 | 83.0 | |||||||||
| gcsl average | 62.0 | 58.0 | |||||||||
| Hyperparameter | Value | Environment |
| Hidden layers | 2 | All |
| Layer width | 1024 | All |
| Nonlinearity | ReLU | All |
| Learning rate | 1e-3 | All |
| Epochs | 10 | GCSL |
| 50 | Kitchen | |
| 2000 | Gym | |
| Gradient steps | 20000 | AntMaze |
| Batch size | 256 | GCSL,Kitchen |
| 16384 | Gym,AntMaze | |
| Dropout | 0.1 | GCSL,Kitchen |
| 0 | Gym,AntMaze | |
| Goal state | Given | GCSL |
| All subtasks completed | Kitchen | |
| Reward target | (x,y) location | AntMaze |
| 110 | Gym medium-expert,AntMaze,Kitchen | |
| 90 | Gym {hopper,walker2d}-medium-replay,walker2d-medium | |
| 60 | Gym hopper-medium | |
| Policy output | 40 | Gym random,halfcheetah-{medium,medium-replay} |
| Discrete categorical | GCSL | |
| Unimodal Gaussian | Kitchen, Gym,AntMaze |
| Task | TR²-GPT2 | TR²-LSTM | SGC | GC | |
| 1 | Box Pusher (Train) | 78.7±3.6 | 48.2±4.4 | 32.8±4.2 | 81.3±3.5 |
| 2 | Box Pusher w/ Obstacles | 65.6±4.2 | 29.4±4.0 | 15.1±3.2 | 55.2±4.4 |
| 3 | Couch Moving Short 3 (Train) | 89.6±1.6 | 52.9±4.4 | 41.7±4.4 | 8.6±1.4 |
| 4 | Couch Moving Long 3 | 88.5±1.6 | 52.9±4.4 | 22.9±3.7 | 0.8±0.4 |
| 5 | Couch Moving Long 4 | 74.2±2.2 | 32.8±4.2 | 17.5±3.4 | 0.8±0.4 |
| 6 | Couch Moving Long 5 | 60.7±2.5 | 23.7±3.8 | 11.2±2.8 | 0±0 |
| 7 | Pick and Place (Train) | 95.3±1.9 | 0±0 | 39.0±4.3 | 0±0 |
| 8 | Stack 4 Blocks | 99.0±1.8 | 0±0 | 0±0 | 0±0 |
| 9 | Stack 5 Blocks | 63.5±8.5 | 0±0 | 0±0 | 0±0 |
| 10 | Stack 6 Blocks | 28.1±8.0 | 0±0 | 0±0 | 0±0 |
| 11 | Build 3-2-1 Pyramid | 90.6±5.2 | 0±0 | 0±0 | 0±0 |
| 12 | Build 4-3-2-1 Pyramid | 70.8±8.0 | 0±0 | 0±0 | 0±0 |
| 13 | Open 1 Drawer (Train) | 92.2±2.3 | 52.1±4.5 | 68.5±4.1 | 0±0 |
| 14 | Open 1 Unseen Drawer | 91.4±2.5 | 47.9±4.5 | 62.9±4.3 | 0±0 |
| 15 | Open 2 Drawers | 17.7±6.5 | 15.7±6.4 | 10.4±5.4 | 0±0 |
| Task | TR2-GPT2 | SILO |
| Obstacle Push | 95.3±1.1 | 70.0 |
| Pick and Place | 100±0 | 95.0 |
| Embodiment | TR2-GPT2 | XIRL |
| Short-stick | 79.0±3.6 | 81.4±3.4 |
| Medium-stick | 96.9±1.5 | 88.1±2.9 |
| Long-stick | 100±0 | 100±0 |
| Gripper | 63.3±4.3 | 72.3±4.0 |
| Hyperparameter | Default | TR²-GPT2 | TR²-LSTM | SGC | GC |
| Optimizer | Adam | ||||
| Policy LR | 3e-4 | = | |||
| Value Function LR | 3e-4 | ||||
| Rollout Batch Size | 20000 | ||||
| Training Batch Size | 1024 | ||||
| Epochs | 2000 | ||||
| Gradient Updates per Epoch | 60 | ||||
| Max Episode Length | 200 | ||||
| Number of Parallel Envs | 20 | ||||
| Initial Log Std Scale | -0.5 | ||||
| Target KL | 0.15 | ||||
| Clip Ratio | 0.2 | ||||
| GAELambda | 0.95 | ||||
| Discount Factor | 0.99 | ||||
| Reward Function | Trajectory Following | Task | |||
| Trajectory Following Reward - β | 5 | N/A | |||
| Trajectory Following Reward - w | 30 | ■ | N/A | ||
| Max Abstract Trajectory Length | 32 | N/A | N/A | ||
| Trajectory Sample Skip Steps (p) | 2 | N/A | N/A | ||
| Stack Size (k) | 2 | N/A | N/A | ||
| Timestep Embeddings | True | N/A | N/A | ||
| Sequential Module Layer Dims | [128,128,128,128] | N/A | N/A | ||
| Dropout | 0.1 | 0 | 0 | ||
| State Embedding Activation | ReLU | - | 1 | ||
| State Embedding Dims | 32 | ||||
| Actor MLP Activation | Tanh | ||||
| Actor and Critic MLP Layer Dims | [128,128] |
| Hyperparameter | Default | TR²-GPT2 | TR²-LSTM | SGC | GC |
| Max Episode Length | 150 | = | 1 | 1 | |
| Max Abstract Trajectory Length | 50 | N/A | N/A | ||
| Trajectory Sample Skip Steps (p) | 10 | N/A | N/A | ||
| Stack Size (k) | 5 | N/A | N/A | ||
| Timestep Embeddings | True | N/A | N/A | ||
| State Embedding Dims | 64 | 1 | 1 |
| Hyperparameter | Default | TR²-GPT2 | TR²-LSTM | SGC | GC |
| Max Abstract Trajectory Length | 55 | N/A | N/A | ||
| Trajectory Sample Skip Steps (p) | 1 | N/A | N/A | ||
| Stack Size (k) | 5 | N/A | N/A | ||
| Timestep Embeddings | False | N/A | N/A | ||
| State Embedding Dims | 64 | = | = | 1 |
| Hyperparameter | Default | TR²-GPT2 | TR²-LSTM | SGC | GC |
| Max Abstract Trajectory Length | 32 | N/A | N/A | ||
| Trajectory Sample Skip Steps (p) | 1 | N/A | N/A | ||
| Stack Size (k) | 5 | N/A | N/A | ||
| Timestep Embeddings | False | N/A | N/A | ||
| State Embedding Dims | 64 | 1 | 1 |
| Hyperparameter | TR²-GPT2 |
| Max Abstract Trajectory Length | 48 |
| Trajectory Sample Skip Steps (p) | 2 |
| Stack Size (k) | 3 |
| Timestep Embeddings | False |
| State Embedding Dims | 32 |
| Hyperparameter | TR²-GPT2 |
| Rollout Batch Size | 10000 |
| Training Batch Size | 512 |
| Max Abstract Trajectory Length | 12 |
| Trajectory Sample Skip Steps (p) | 1 |
| Stack Size (k) | 2 |
| Timestep Embeddings | False |
| State Embedding Dims | 32 |
| Hyperparameter | TR2-GPT2 |
| Max Abstract Trajectory Length | 30 |
| Trajectory Sample Skip Steps (p) | 0 |
| Stack Size (k) | 5 |
| State Embedding Dims | 64 |
| Method | MNIST | CelebA_S | CelebAR | |||
| Acc_B Acc_BC | Acc_B | Acc_BC | Acc_B Acc_BC | Acc_B | Acc_BC | |
| Global | .852 ±2e-4 .847 ±6e-4 | .930 | )± 5e-5.910 ± 3e-4 | .909 ± 6e-5.929 | ±5e-5 | 1.882 ± 6e-4. .903 ±7e-4 |
| FT | .989 ±6e-7.704 | ±3e-4 | .952 ± 6e-5 .786 ± 5e-4 | .963 ± 6e-6 .849 | ±1e-3 | .931 ± le-4 .891 ±2e-4 |
| Ditto | .982 ± 3e-6 .724 ± 1e-3 | .948 ± 5e-5 .715 ± 5e-4 | .966 ± 1e-5 .884 ± 2e-4 | .939 ± 4e-5 .897 ± 3e-4 | ||
| UW | .968 ±2e-5.823 | 3±7e-4 | 930 ± 7e-6 .889 ± 5e-4 | .936 ± 1e-5 .895 ± 4e-4 | N/A N/A | |
| JTT | .985 ± 2e-7 .707 | ±6e-5 .952 | 2±2e-6.817 1±3e-4 | .956 ± 2e-5 .836 ± 1e-3 | N/A N/A | |
| Ours | .951 ±2e-5.870 | ±8e-4 .932 | ± 3e-5 .910 ± 2e-4 | .925 ±2e-5.927 | ±8e-5 | .901 ± 3e-4 .916 ± 5e-8 |
| Client ID | Non-blond Female | Non-Blond Male | Blond Female | Blond Male |
| 0 | 55 | 31 | 12 | 2 |
| 1 | 30 | 68 | 0 | 2 |
| 2 | 59 | 28 | 11 | 2 |
| Client ID | Non-blond Female | Non-Blond Male | Blond Female | Blond Male |
| 0 | 115 | 60 | 45 | 2 |
| 1 | 60 | 75 | 79 | 0 |
| 2 | 86 | 111 | 14 | 0 |
| Client ID1 | Non-blond FemaleNon-Blond Male | Blond Female | Blond Male | |
| 0 | 0 | 0 | 0 | 200 |
| 1 | 0 | 0 | 0 | 203 |
| 2 | 0 | 0 | 0 | 204 |
| CNN 28x28 | CNN 64x64 |
| Input: R3.28-28 | Input: R3-64.64 |
| 4·4 conv, 64 BN LReLU, stride 2 | 4·4 conv, 64 BN LReLU, stride 2 |
| 4·4 conv, 128 BN LReLU, stride 2 | 4·4 conv, 64 BN LReLU, stride 2 |
| FC4096ReLU | FC4096ReLU |
| FC10 | FC10 |
| Brightness[42]Ours | Elastic[42] Ours | |||||||
| Jellyfish | -3 | 0 | -6 | 5 | 2 | -1 | -2 | 3 |
| Holster | -1 | 4 | -2 | 6 | 1 | 0 | -1 | 14 |
| ChitonFire salamanderSpaghetti squash | -1 | 3 | -3 | 8 | -1 | 2 | -3 | 8 |
| 1 | 3 | -1 | 3 | 22 | 5 | 2 | 2 | |
| -3 | 4 | -2 | 5 | 22 | 3 | 1 | 6 | |
| |brigh cont defoc elast | fog | frost gauss glass impul jpeg motn pixel shot snow zoom | |||||||||||||
| Joint Train | 62.3 | 4.5 | 26.7 | 39.9 | 25.7 | 30.0 | 5.8 | 16.3 | 5.8 | 45.3 | 30.9 | 45.9 | 7.1 | 25.1 | 31.8 |
| Fine-Tune | 67.5 | 7.8 | 33.9 | 32.4 | 36.4 | 38.2 | 22.0 | 15.7 | 23.9 | 51.2 | 37.4 | 51.9 | 23.7 | 37.6 | 37.1 |
| ViT Probe | 68.3 | 6.4 | 24.2 | 31.6 | 38.638.4 | 17.4 | 18.4 | 18.2 | 51.2 | 32.2 | 49.7 | 18.2 | 35.9 | 32.2 | |
| TTT-MAE| | 69.1 | 9.8 | 34.4 | 50.7 | 44.7 | 50.7 | 30.5 | 36.9 | 32.4 | 63.0 | 41.9 | 63.0 | 33.0 | 42.8 | 45.9 |
| ImageNet-A | ImageNet-R | ||
| Baseline | TTT-MAE | Baseline | TTT-MAE |
| 15.3 | 21.3 | 31.3 | 38.9 |
| Train Split | 1 | 2 | 3 |
| Baseline | 76.1 | 76.5 | 78.2 |
| TTT-MAE | 76.4 | 76.7 | 79.4 |
| |Train| = 500 | |Train] = 1500 | |||||||||
| 1,2.0 | 2.0,1 | 3,2,1 | 0,10.0 | 5.3,3 | 1,2.0 | 2.0,1 | 3,2,1 | 0,10.0 | 5,3,3 | |
| Linear | 8.23±0.00 | 7.55±0.00 | 9.76±0.00 | 11.36±0.00 1 | 11.62±0.00 | 8.22±0.00 7.55±0.009.76±0.00 1136±0.0011.62±0.00 | ||||
| GNN | 5.33±0.07 | 5.01±0.08 | 7.58±0.08 | 9.83±0.04 | 9.77±0.02 | 3.61±0.13 3.23±0.07 4.73±0.11 | 7.97±0.44 | 7.91±0.31 | ||
| TFN | 11.54±0.38 9.87±0.27 | 11.66±0.08 | 13.43±0.31 | 12.23±0.12 | 5.86±0.35 4.97±0.23 8.51±0.14 11.21±0.21 | 10.75±0.08 | ||||
| SE(3)-Tr. | 5.54±0.06 | 5.14±0.03 | 8.95±0.04 | 11.42±0.01 | 11.59±0.01 | 5.02±0.03 4.68±0.05 8.39±0.02 | 10.82±0.03 | 10.85±0.02 | ||
| RF | 3.50±0.17 | 3.07±0.24 | 5.25±0.44 | 7.59±0.25 | 7.73±0.39 | 2.97±0.15 2.19±0.11 3.80±0.25 | 5.71±0.31 | 5.66±0.27 | ||
| EGNN | 2.81±0.12 | 2.27±0.04 | 4.67±0.07 | 4.75±0.05 | 4.59±0.07 | 2.59±0.10 1.86±0.02 2.54±0.01 | 2.79±0.04 | 3.25±0.07 | ||
| EGNNReg | 2.94±0.01 | 2.66±0.06 | 7.01±0.34 | 5.03±0.08 | 6.31±0.04 | 2.74±0.08 1.58±0.03 2.62±0.05 | 3.03±0.07 | 3.07±0.04 | ||
| GMN | 1.84±0.02 | 2.02±0.02 | 2.48±0.04 | 2.92±0.04 | 4.08±0.03 | 1.68±0.04 1.47±0.03 2.10±0.04 | 2.32±0.02 | 2.86±0.01 | ||
| |Train| = 500 | |Train|= 1500 | |||||||
| 3,2,1 | 2,4,0 | 1,0,3 | Average | 3,2,1 | 2,4.0 | 1,0,3 | Average | |
| GNN | 7.58 | 8.06 | 8.37 | 8.00 | 4.73 | 4.98 | 5.58 | 5.10 |
| EGNN | 4.67 | 3.42 | 4.40 | 4.16 | 2.54 | 2.75 | 3.49 | 2.93 |
| EGNNReg | 7.01 | 4.49 | 6.62 | 6.04 | 2.62 | 2.62 | 4.29 | 3.18 |
| GMN | 2.48 | 2.53 | 3.28 | 2.76 | 2.10 | 2.18 | 2.65 | 2.31 |
| |Train|=500 |Train =1500 3,2,1 5,3,3 3,2,1 5,3,3 |
| GMN |
| 2.48 4.08 2.10 2.86 GMN-L 3.19 4.34 2.28 3.03 |
| w/o Equivariance 3.74 4.41 2.46 3.29 |
| 2 with U 1 一 -1) 2.86 4.15 2.20 3.00 ki , |
| i with Uki shared 3.86 4.25 2.30 2 ), 3.10 |
| 2 with only f 3.10 4.39 2.34 4.19 |
| 2 with only f,shared 2.91 4.94 2.39 3.45 |
| w/o Normalization 3.15 O.F. O.F. O.F. |
| Aspirin | Benzene | Ethanol Malonaldehyde Naphthalene Salicylic | Toluene | Uracil | ||||
| RF | 10.94±0.01 | 103.72±1.29 | 4.64±0.01 | 13.93±0.03 | 0.50±0.01 | 1.23±0.01 | 10.93±0.04 | 0.64±0.01 |
| TFN | 12.37±0.18 | 58.48±1.98 | 4.81±0.04 | 13.62±0.08 | 0.49±0.01 | 1.03±0.02 | 10.89±0.01 | 0.84±0.02 |
| SE(3)-Tr. | 11.12±0.06 | 68.11±0.67 | 4.74±0.13 | 13.89±0.02 | 0.52±0.01 | 1.13±0.02 | 10.88±0.06 | 0.79±0.02 |
| EGNN | 14.41±0.15 | 62.40±0.53 | 4.64±0.01 | 13.64±0.01 | 0.47±0.02 | 1.02±0.02 | 11.78±0.07 | 0.64±0.01 |
| EGNNReg | 13.82±0.19 | 61.68±0.37 | 6.06±0.01 | 13.49±0.06 | 0.63±0.01 | 1.68±0.01 | 11.05±0.01 | 0.66±0.01 |
| GMN | 10.14±0.03 | 48.12±0.40 | 4.83±0.01 | 13.11±0.03 | 0.40±0.01 | 0.91±0.01 | 10.22±0.08 | 0.59±0.01 |
| GMN-L | 9.76±0.11 | 54.17±0.69 | 4.63±0.01 | 12.82±0.03 | 0.41±0.01 | 0.88±0.01 | 10.45±0.04 | 0.59±0.01 |
| GNN | TFN | SE(3)-Tr. | RF | EGNN | EGNNReg | GMN | GMN-L |
| 67.3±1.1 | 66.9±2.7 | 60.9±0.9 | 197.0±1.0 | 59.1±2.1 | 59.5±2.2 | 43.9±1.1 | 50.9±0.7 |
| GMN | EGNN | EGNNReg | GNN | TFN | SE(3)-Tr. | RF | Linear | |
| 1,2,0 | 1.84 | 2.81 | 2.94 | 5.33 | 11.54 | 5.54 | 3.50 | 8.23 |
| 2,0,1 | 2.02 | 2.27 | 2.66 | 5.01 | 9.87 | 5.14 | 3.07 | 7.55 |
| 2,4,0 | 2.34 | 3.59 | 3.87 | 8.05 | 11.30 | 9.22 | 5.37 | 10.10 |
| 0,5,0 | 2.54 | 4.13 | 4.29 | 8.63 | 11.92 | 9.83 | 5.94 | 10.48 |
| 7,0,1 | 2.39 | 2.66 | 3.41 | 7.05 | 10.67 | 8.38 | 4.66 | 9.71 |
| 1,0,3 | 3.21 | 4.56 | 5.14 | 8.32 | 11.62 | 9.57 | 5.91 | 9.90 |
| 3,2,1 | 2.48 | 4.67 | 7.01 | 7.58 | 11.66 | 8.95 | 5.25 | 9.76 |
| 4,8,0 | 3.69 | 4.79 | 7.09 | 9.65 | 12.05 | 11.21 | 7.59 | 11.45 |
| 0,10,0 | 2.92 | 4.75 | 5.03 | 9.83 | 13.43 | 11.42 | 7.59 | 11.36 |
| 8,0,4 | 3.37 | 4.17 | 5.32 | 9.49 | 11.72 | 11.12 | 7.51 | 11.44 |
| 2,0,6 | 4.06 | 5.06 | 5.58 | 10.13 | 12.13 | 11.74 | 8.15 | 11.61 |
| 5,3,3 | 4.08 | 4.59 | 6.31 | 9.77 | 12.23 | 11.59 | 7.73 | 11.62 |
| |Train| = 500 | |Train| = 1500 | ||||
| 3,2,1 | 2,0,6 | 5,3,3 | 3,2,1 | 2,0,6 5,3,3 | |
| EGNN | 4.67 | 5.06 | 4.59 | 2.54 | 3.42 |
| EGNNReg | 7.01 | 5.58 | 6.31 | 2.62 | 3.50 3.61 3.07 |
| GMN (Stick only) | 3.02 | 4.32 | 4.21 | 2.37 | 3.30 2.88 |
| GMN | 2.48 | 4.06 | 4.08 | 2.10 | 3.22 2.86 |
| Node +C | Edge + C | |||||||||
| 1,2,0 | 2,0,1 | 3,2,1 | 0,10,0 | 5,3,3 | 1,2.0 | 2,0,1 | 3,2,1 | 0,10,0 | 5,3,3 | |
| EGNN | 2.89 | 2.28 | 4.25 | 4.80 | 4.50 | 2.81 | 2.27 | 4.67 | 4.75 | 4.59 |
| EGNNReg | 3.17 | 2.74 | 8.20 | 5.01 | 6.64 | 2.94 | 2.66 | 7.01 | 5.03 | 6.31 |
| GMN | 1.89 | 2.01 | 2.63 | 3.07 | 4.02 | 1.84 | 2.02 | 2.48 | 2.92 | 4.08 |
| |Train| = 500 | |Train| = 1500 | |||||||||
| 1,2.0 | 2.0,1 | 3,2,1 | 0,10,0 | 5,3.3 | 1,2,0 | 2,0,1 | 3,2,1 | 0,10,0 | 5,3,3 | |
| GNN | 0.200 | 0.386 | 0.492 | 0.154 | 0.468 | 0.225 | 0.426 | 0.779 | 0.251 | 0.772 |
| EGNN | 0.220 | 0.370 | 0.714 | 0.248 | 0.760 | 0.217 | 0.317 | 0.521 | 0.139 | 0.596 |
| EGNNReg | 0.172 | 0.146 | 0.232 | 0.198 | 0.241 | 0.159 | 0.053 | 0.091 | 0.097 | 0.075 |
| GMN | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| |Train| = 500 | |Train| = 1500 | |||||||
| 3,2,1 | 2,4,0 | 1,0,3 | Average | 3,2,1 | 2,4,0 | 1,0,3 | Average | |
| GMN-Transfer | 2.48 | 2.53 | 3.28 | 2.76 | 2.10 | 2.18 | 2.65 | 2.31 |
| GMN-Original | 2.48 | 2.34 | 3.21 | 2.68 | 2.10 | 2.01 | 2.44 | 2.18 |
| |Train|= 500 | |Train| = 1500 | |||||||||
| 1,2.0 | 2.,0,1 | 3,2,1 | 0,10,0 | 5,3,3 | 1,2.0 | 2,0,1 | 3,2,1 | 0,10,0 | 5,3,3 | |
| EGNN | 2.81 | 2.27 | 4.67 | 4.75 | 4.59 | 2.59 | 1.86 | 2.54 | 2.79 | 3.25 |
| EGNNReg | 2.94 | 2.66 | 7.01 | 5.03 | 6.31 | 2.74 | 1.58 | 2.62 | 3.03 | 3.07 |
| GMN-L | 2.32 | 2.09 | 3.19 | 3.88 | 4.34 | 1.93 | 1.56 | 2.28 | 2.72 | 3.03 |
| GMN-LReg | 2.52 | 2.23 | 3.34 | 3.67 | 4.31 | 1.91 | 1.88 | 2.49 | 2.61 | 3.00 |
| GMN | 1.84 | 2.02 | 2.48 | 2.92 | 4.08 | 1.68 | 1.47 | 2.10 | 2.32 | 2.86 |
| |Train| = 500 | |Train| = 1500 | |||||
| 3,2,1 | 5,3,3 | 8,6,0 | 3,2,1 | 5,3,3 | 8,6,0 | |
| GMN | 2.48 | 4.08 | 2.84 | 2.10 | 2.86 | 2.22 |
| GMN w/ Join | 2.59 | 4.27 | 2.98 | 2.22 | 3.16 | 2.37 |
| GMN w/ Split | 2.57 | 4.11 | 2.95 | 2.16 | 3.13 | 2.26 |
| GMN w/ Join + Split | 2.63 | 4.16 | 3.01 | 2.31 | 3.01 | 2.34 |
| GMN w/ Change in Length | 2.75 | 4.36 | 3.11 | 2.15 | 3.15 | 2.41 |
| Aspirin Ethanol Malonaldehyde Naphthalene Salicylic | Benzene | Toluene | |||
| EGNN | 14.41±0.15 62.40±0.53 | 4.64±0.01 | 13.64±0.01 | 0.47±0.02 | 1.02±0.02 11.78±0.07 |
| GMN | 10.14±0.03 48.12±0.404 | 4.83±0.01 | 13.11±0.03 | 0.40±0.01 | 0.91±0.01 10.22±0.08 |
| GMN-L | 9.76±0.11 | 54.17±0.69 4.63±0.01 | 12.82±0.03 | 0.41±0.01 | 0.59±0.01 0.88±0.01 10.45±0.04 0.59±0.01 |
| GMN-LH | 10.25±0.06 52.02±0.974 | 4.62±0.01 | 12.83±0.03 | 0.41±0.01 | 1.03±0.01 10.81±0.14 0.59±0.01 |
| EGNN | GMN | GMN-L | GMN-LH |
| 59.1±2.1 | 43.9±1.1 | 50.9±0.7 | 48.7±1.1 |
| Dataset | BC | 10%BC | DT | AWAC | Onestep RL | TD3+BC | CQL | IQL | X-QL C | X-QL T | |
| Aaaee | halfcheetah-medium-v2 | 42.6 | 42.5 | 42.6 | 43.5 | 48.4 | 48.3 | 44.0 | 47.4 | 47.7 | 48.3 |
| hopper-medium-v2 | 52.9 | 56.9 | 67.6 | 57.0 | 59.6 | 59.3 | 58.5 | 66.3 | 71.1 | 74.2 | |
| walker2d-medium-v2 | 75.3 | 75.0 | 74.0 | 72.4 | 81.8 | 83.7 | 72.5 | 78.3 | 81.5 | 84.2 | |
| halfcheetah-medium-replay-v2 | 36.6 | 40.6 | 36.6 | 40.5 | 38.1 | 44.6 | 45.5 | 44.2 | 44.8 | 45.2 | |
| hopper-medium-replay-v2 | 18.1 | 75.9 | 82.7 | 37.2 | 97.5 | 60.9 | 95.0 | 94.7 | 97.3 | 100.7 | |
| walker2d-medium-replay-v2 | 26.0 | 62.5 | 66.6 | 27.0 | 49.5 | 81.8 | 77.2 | 73.9 | 75.9 | 82.2 | |
| halfcheetah-medium-expert-v2 | 55.2 | 92.9 | 86.8 | 42.8 | 93.4 | 90.7 | 91.6 | 86.7 | 89.8 | 94.2 | |
| hopper-medium-expert-v2 walker2d-medium-expert-v2 | 52.5 107.5 | 110.9 109.0 | 107.6 | 55.8 | 103.3 | 98.0 | 105.4 | 91.5 | 107.1 | 111.2 | |
| antmaze-umaze-v0 | 54.6 | 62.8 | 108.1 59.2 | 74.5 | 113.0 | 110.1 | 108.8 | 109.6 | 110.1 | 112.7 | |
| antmaze-umaze-diverse-v0 | 45.6 50.2 | 53.0 | 56.7 49.3 | 64.3 | 78.6 | 74.0 | 87.5 | 87.2 | 93.8 | ||
| 60.7 | 71.4 | 84.0 | 62.2 | 69.17 | 82.0 | ||||||
| antmaze-medium-play-v0 antmaze-medium-diverse-v0 | 0.0 | 5.4 | 0.0 | 0.0 | 0.3 | 10.6 | 61.2 | 71.2 | 73.5 | 76.0 | |
| antmaze-large-play-vo | 0.0 0.0 | 9.8 0.0 | 0.0 0.0 | 0.7 | 0.0 | 3.0 | 53.7 | 70.0 | 67.8 | 73.6 | |
| antmaze-large-diverse-v0 | 0.0 | 6.0 | 0.0 | 0.0 1.0 | 0.0 0.0 | 0.2 0.0 | 15.8 14.9 | 39.6 47.5 | 41 | 46.5 | |
| raaa | kitchen-complete-v0 | 65.0 | 47.3 | 49.0 | |||||||
| kitchen-partial-v0 | 38.0 | - | - | - | 1 | - | 43.8 | 62.5 46.3 | 72.5 | 82.4 | |
| kitchen-mixed-v0 | 51.5 | - - | - 1 | 1 - | 1 | = 1 | 49.8 51.0 | 51.0 | 73.8 | 73.7 | |
| runtime * We see very fast convergence for our method on some tasks,and saturate performance at half the iterations as IQL. | 10m | 10m 960m | 20m | - 20m | 20m | 80m | 20m | 54.6 10-20m | 62.5 10-20m* |
| Dataset | CQL | IQL | X-QL T |
| umaze-vO | 70.1 →99.4 | 86.7 →96.0 | 93.8 →99.6 |
| umaze-diverse-v0 | 31.1 →99.4 | 75.0 →84.0 | 82.0 →99.0 |
| medium-play-v0 | 23.0 →0.0 | 72.0 →95.0 | 76.0 →97.0 |
| medium-diverse-v0 | 23.0 →32.3 | 68.3 →92.0 | 73.6 →97.1 |
| large-play-v0 | 1.0 →0.0 | 25.5 →46.0 | 45.1 →59.3 |
| large-diverse-v0 | 1.0 →0.0 | 42.6 →60.7 | 49.0 →82.1 |
| Env | Beta | Grad Clip | Batch Size | V_updates |
| halfcheetah-medium-v2 | 2(1) | 7(7) | 256 (256) | 1(1) |
| hopper-medium-v2 | 2(5) | 7(7) | 256 (256) | 1(1) |
| walker2d-medium-v2 | 2(10) | 7(7) | 256 (256) | 1(1) |
| halfcheetah-medium-replay-v2 | 2(1) | 7(5) | 256 (256) | 1(1) |
| hopper-medium-replay-v2 | 2(2) | 7(7) | 256 (256) | 1(1) |
| walker2d-medium-replay-v2 | 2(5) | 7(7) | 256 (256) | 1(1) |
| halfcheetah-medium-expert-v2 | 2(1) | 7(5) | 256 (1024) | 1(1) |
| hopper-medium-expert-v2 | 2(2) | 7(7) | 256 (1024) | 1(1) |
| walker2d-medium-expert-v2 | 2(2) | 7(5) | 256 (1024) | 1(1) |
| antmaze-umaze-vO | 0.6 (1) | 7(7) | 256 (256) | 1(1) |
| antmaze-umaze-diverse-v0 | 0.6 (5) | 7(7) | 256 (256) | 1(1) |
| antmaze-medium-play-v0 | 0.6 (0.8) | 7(7) | 256 (1024) | 1(2) |
| antmaze-medium-diverse-vO | 0.6 (0.6) | 7(7) | 256 (256) | 1(4) |
| antmaze-large-play-v0 | 0.6 (0.6) | 7(5) | 256 (1024) | 1(1) |
| antmaze-large-diverse-v0 | 0.6 (0.6) | 7(5) | 256 (1024) | 1(1) |
| kitchen-complete-v0 | 5(2) | 7(7) | 256 (1024) | 1(1) |
| kitchen-partial-v0 | 5(5) | 7(7) | 256 (1024) | 1(1) |
| kitchen-mixed-v0 | 5(8) | 7(7) | 256 (1024) | 1(1) |
| pen-human-v0 | 5(5) | 7(7) | 256 (256) | 1(1) |
| hammer-human-vO | 5 (0.5) | 7(3) | 256 (1024) | 1(4) |
| door-human-vo | 5(1) | 7(5) | 256 (256) | 1(1) |
| relocate-human-v0 | 5 (0.8) | 7(5) | 256 (1024) | 1(2) |
| pen-cloned-v0 | 5 (0.8) | 7(5) | 256 (1024) | 1(2) |
| hammer-cloned-v0 | 5(5) | 7(7) | 256 (256) | 1(1) |
| door-human-vo | 5(5) | 7(7) | 256 (256) | 1(1) |
| relocate-human-vO | 5(5) | 7(7) | 256 (256) | 1(1) |
| Parameter | SAC | TD3 |
| Batch Size | 1024 | 256 |
| Learning Rate | 0.0001 | 0.001 |
| Critic Freq | 1 | 1 |
| Actor Freq | 1 | 2 |
| Actor and Critic Arch | 1024,1024 | 256,256 |
| Buffer Size | 1,000,000 | 1,000,000 |
| Actor Noise | Auto-tuned | 0.1,0.05 (Hopper) |
| Target Noise | 0.2 |
| Env | X-SAC | X-TD3 | X-TD3-DQ |
| Cheetah Run | 2 | 5 | 4 |
| Walker Run | 1 | 2 | 4 |
| Hopper Hop | 2 | 2 | 3 |
| Quadruped Run | 5 | 5 | 20 |
| Method | RPrecision (top 3)↑ | FID↓ | Multimodal Dist↓ | Diversity→ | Multimodality↑ |
| Real | 0.797±.002 | 0.002±.000 | 2.974±.008 | 9.503±.065 | |
| JL2P | 0.486±.002 | 11.02±.046 | 5.296±.008 | 7.676±.058 | 1 |
| Text2Gesture | 0.345±.002 | 7.664±.030 | 6.030±.008 | 6.409±.071 | 1 |
| T2M | 0.740±.003 | 1.067±.002 | 3.340±.008 | 9.188±.002 | 2.090±.083 |
| MDM (ours) | 0.611±.007 | 0.544±.044 | 5.566±.027 | 9.559±.086 | 2.799±.072 |
| + sent-BERT | 0.609±.006 | 0.586±.036 | 5.504±.03 | 9.666±.095 | 2.707±.188 |
| MDM (decoder) | 0.608±.005 | 0.767±.085 | 5.507±.020 | 9.176±.070 | 2.927±.125 |
| + input token | 0.621±.005 | 0.567±.051 | 5.424±.022 | 9.425±.060 | 2.834±.095 |
| MDM (U-net) | 0.603±.006 | 1.137±.008 | 5.629±.032 | 8.958±.098 | 2.636±.214 |
| MDM (GRU) | 0.645±.005 | 4.569±.150 | 5.325±.026 | 7.688±.082 | 1.2646±.024 |
| Method | R Precision (top 3)↑ | FID↓ | Multimodal Dist↓ | Diversity→ | Multimodality↑ |
| Real | 0.779±.006 | 0.031±.004 | 2.788±.012 | 11.08±.097 | - |
| JL2P | 0.483±.005 | 6.545±.072 | 5.147±.030 | 9.073±.100 | - |
| Text2Gesture | 0.338±.005 | 12.12±.183 | 6.964±.029 | 9.334±.079 | |
| T2M | 0.693±.007 | 2.770±.109 | 3.401±.008 | 10.91±.119 | 1.482±.065 |
| MDM (ours) | 0.396±.004 | 0.497±.021 | 9.191±.022 | 10.847±.109 | 1.907±.214 |
| Method | FID↓ | Accuracy↑ | Diversity→ | Multimodality→ |
| Real (INR) | 0.020±.010 | 0.997±.001 | 6.850±.050 | 2.450±.040 |
| Real (ours) | 0.050±.000 | 0.990±.000 | 6.880±.020 | 2.590±.010 |
| Action2Motion (2020) | 0.338±.015 | 0.917±.003 | 6.879±.066 | 2.511±.023 |
| ACTOR (2021) | 0.120±.000 | 0.955±.008 | 6.840±.030 | 2.530±.020 |
| INR (2022) | 0.088±.004 | 0.973±.001 | 6.881±.048 | 2.569±.040 |
| MDM (ours) | 0.100±.000 | 0.990±.000 | 6.860±.050 | 2.520±.010 |
| w/o foot contact | 0.080±.000 | 0.990±.000 | 6.810±.010 | 2.580±.010 |
| Method | FIDtrain | FIDtest | Accuracy↑ | Diversity→ | Multimodality→ |
| Real | 2.92±.26 | 2.79±.29 | 0.988±.001 | 33.34±.320 | 14.16±.06 |
| ACTOR (2021) | 20.49±2.31 | 23.43±2.20 | 0.911±.003 | 31.96±.33 | 14.52±.09 |
| INR (2022) (best variation) | 9.55±.06 | 15.00±.09 | 0.941±.001 | 31.59±.19 | 14.68±.07 |
| MDM (ours) | 9.98±1.33 | 12.81±1.46 | 0.950±.000 | 33.02±.28 | 14.26±.12 |
| w/o foot contact | 9.69±.81 | 13.08±2.32 | 0.960±.000 | 33.10±.29 | 14.06±.05 |
| Method | FID↓ | KID↓ | Precision↑ Recall↑ | Diversity↑ |
| ACTOR (2021) | 48.80 | 0.53 | 0.72, 0.74 | 14.10 |
| MoDi (2022) | 13.03 | 0.12 | 0.71, 0.81 | 17.57 |
| MDM (ours) | 31.92 | 0.36 | 0.66,0.62 | 17.00 |
| Diffusion steps (T) | R Precision (top 3)↑ | FID↓ | Multimodal Dist↓ | Diversity→ |
| Real | 0.797±.002 | 0.002±.000 | 2.974±.008 | 9.503±.065 |
| 10 | 0.574±.006 | 1.461±.088 | 5.816±.033 | 9.369±.058 |
| 100 | 0.640±.007 | 0.454±.039 | 5.336±.029 | 9.906±.053 |
| 500 | 0.662±.007 | 0.553±.055 | 5.177±.028 | 9.890±.074 |
| 1000 | 0.611±.007 | 0.544±.044 | 5.566±.027 | 9.559±.086 |
| Method | FID↓ | Accuracy↑ | Diversity→ | Multimodality→ |
| Real | 0.050±.000 | 0.990±.000 | 6.880±.020 | 2.590±.010 |
| MDM (ours) | 0.100±.000 | 0.990±.000 | 6.860±.050 | 2.520±.010 |
| w/o foot contact | 0.080±.000 | 0.990±.000 | 6.810±.010 | 2.580±.010 |
| w/o geometric losses | 0.090±.000 | 0.990±.000 | 6.820±.020 | 2.550±.010 |
| foot contact only | 0.100±.000 | 0.990±.000 | 6.860±.050 | 2.520±.010 |
| velocity only | 0.100±.000 | 0.990±.000 | 6.820±.020 | 2.590±.000 |
| pose only | 0.090±.000 | 0.990±.000 | 6.830±.020 | 2.570±.020 |
| Architecture | # Parameters (-106) |
| Transformer Encoder | 17.88 |
| TransformerDecoder | 26.29 |
| + input token | 26.29 |
| U-net | 23.47 |
| GRU | 4.47 |
| Success rate | miniF2F-valid | miniF2F-test |
| Baselines | ||
| Sledgehammer | 9.9% | 10.4% |
| Sledgehammer + heuristics | 18.0% | 20.9% |
| Thor (Jiang et al., 2022) | 28.3% | 29.9% |
| Thor+ expert iteration (Wu et al.,2022) | 37.3% | 35.2% |
| Draft,Sketch,and Prove | ||
| Human informal proof | 42.6% | 39.3% |
| Codex informal proof | 40.6% | 35.3% |
| 8B Minerva informal proof | 40.6% | 35.3% |
| 62BMinerva informal proof | 43.9% | 37.7% |
| 540BMinerva informal proof | 42.6% | 38.9% |
| Ablations (with human informal statements and proofs) | ||
| -In-line comments | 37.7% (-4.9%) | 36.5% (-2.8%) |
| - Informal proofs | 38.9%(-3.7%) | 34.0% (-5.3%) |
| - Automated provers | 32.8% (-9.8%) | 30.3% (-9.0%) |
| GPT-Neo | GPT-J | OPT-13B | LLaMA-13B | ||||
| Base | Ours | Base | Ours | Base | Ours | Base | Ours |
| 4.120 | 4.193 | 3.443 | 3.537 | 4.077 | 4.358 | 3.036 | 3.069 |
| Task | GPT-Neo | GPT-J | ||||||
| Base | EWC | LoRA | LoRA&KL | EWC-LoRA | Base | LoRA | EWC-LoRA | |
| Plan Gen | 21.25 | 48.56 | 51.24 | 45.99 | 49.70 | 34.31 | 51.23 | 51.23 |
| Act Recog | 69.22 | 89.98 | 87.98 | 81.42 | 85.43 | 87.98 | 90.16 | 88.52 |
| Count | 22.68 | 55.67 | 27.84 | 49.48 | 28.87 | 30.41 | 63.92 | 67.01 |
| Obj PT | 30.80 | 95.96 | 87.28 | 63.59 | 85.91 | 33.86 | 97.22 | 98.67 |
| Perplexity | 4.120* | 4.995 | 4.360 | 5.029 | 4.193 | 3.443* | 3.675 | 3.537 |
| GPT-Neo | ||||||
| Base | Ours | -w/o Plan Gen | -W/o Act Recog | -w/o Count | - w/o Obj PT | |
| Plan Gen | 21.25 | 49.70 | 14.48 | 49.38 | 49.85 | 50.06 |
| Act Recog | 69.22 | 85.43 | 85.97 | 48.63 | 85.25 | 84.34 |
| Count | 22.68 | 28.87 | 18.56 | 25.26 | 35.05 | 32.99 |
| Obj PT | 30.80 | 85.91 | 92.13 | 84.17 | 86.46 | 29.90 |
| Perplexity | 4.120* | 4.193 | 4.171 | 4.151 | 4.162 | 4.164 |
| Action | Template |
| [Find] <Object> | Find Object |
| [Walk] <0bject> | Walk to Object |
| [Run] <Object> | Run to Object |
| [Sit] <Object> | Sit on Object |
| [StandUp] | Stand up |
| [Grab] <0bject> | Grab Object |
| [Open] <Object> | |
| [Ciose] <Object> | Open Object |
| [Put] <0bject_1> <0bject_2> | Close Object |
| Put Object_1 on Object_2 | |
| [PutIn] <Object_1> <0bject_2> | Put Ojbect_1 in Object_2 |
| [SwitchOn] <Object> | Switch/Turn on Object |
| [SwitchOff] <Object> | Switch/Turn off Ojbect |
| [Drink] <Object> | Drink Object |
| [TurnTo] <Object> | Turn to Object |
| [LookAt] <Object> | Look at Object |
| [Wipe] <Object> | Wipe Object |
| [PutOn] <0bject> | Put on Object |
| [PutOff] <Object> | Put off Object |
| [Greet] <Object> | Greet Object |
| [Drop] <Object> | Drop Object |
| [Touch] <Object> | |
| [Lie] <0bject> | Touch Object |
| [Pour] <0bject_1> <0bject_2> | Lie on Object |
| Pour Object_1 into Object_2 | |
| [Type] <Object> | Type Object |
| [Watch] <Object> | Watch Object |
| [Move] <Object> | Move Object |
| [Wash] <Object> | Wash Object |
| [Rinse] <Object> | Rinse Object |
| [Scrub] <Object> | Scrub Object |
| [Squeeze] <0bject> | Squeeze Object |
| [PlugIn] <Object> | Plug in Object |
| [PlugOut] <Object> | Plug out Object |
| [Cut] <0bject> | Cut Object |
| [Eat] <Object> | Eat Object |
| [Sleep] | Sleep |
| [WakeUp] | Wake up |