article stringlengths 507 295k | abstract stringlengths 417 1.92k | category sequencelengths 1 6 |
|---|---|---|
# 1 Introduction
Modern Artificial Neural Networks (ANNs) achieve remarkable results in recognizing patterns. However, due to their complexity and black-box character, their failures are hard to identify [13] which limits their use in safety-critical environments. Additionally, certain common training schemes encourage overconfidence [8]. If Out-of-Distribution (OOD) samples from other distributions than the In-Distribution (ID) training set are encountered in classification tasks, this issue persists. Encountering such samples is often unavoidable in real-world applications, especially when operating in an open world as autonomous transportation systems do. Therefore, OOD detection has arisen as the task of identifying instances not belonging to the training data distribution [25] which often means finding the label distribution but also extends to identifying when the model might be unable to assess its input reliably. Anomaly detection, OpenSet-Recognition (OSR), and Uncertainty Estimation are closely related to OOD detection and methods can often be applied to the other settings as well [25]. Most importantly, OSR requires explicitly classifying closed-world samples and detecting unknown classes from the open world [25].
Many OOD detection methods rely on post-hoc analysis of output or intermediate features from pre-trained classifiers but models trained solely for discrimination of ID categories may lack relevant features for OOD detection which limits the general usage of such approaches. Integration of OOD detection into the classification framework is thus desirable, rather than applying it afterwards.
In this work, we extend the Prototypical Variational Autoencoder (ProtoVAE) [6] to OOD detection. Instead of the aforementioned post-analysis of application-specific pre-learned features for OOD detection, the feature space is designed to learn to distinguish unknown inputs from the beginning. This is done by estimating the training distribution, learning representations through reconstruction, and designing a distance-based latent space to quantify dissimilarity to ID clusters while also leveraging label information yielding promising results. Additionally, a restriction force is implemented to shape the latent ID region while reconstruction errors are used to identify remaining OOD samples mapped into this region as introduced in [27].
This work proposes the principle of an enclosing restriction to decouple the previous trade-off between compression/estimation of the ID region and reconstructive quality to recover the input rather than just reconstruct features, thus alleviating Autoencoder (AE)-based OOD detection by constraining the ID region in the latent space without collapsing it into one point. To enhance the reconstructive power further, Learned Perceptual Image Patch Similarity (LPIPS) – a perceptual metric – is integrated into the framework for the reconstruction loss and OOD score. The generative and reconstructive abilities of the Variational Autoencoder (VAE) framework enable the provision of additional information and explanation about extracted properties of the data distribution and certain samples, rendering the classification and OOD detection transparent. The method is compared to state-of-the-art approaches using the OpenOOD [24] and a custom railway benchmark.
# 2 Related Work
A ProtoVAE architecture was presented by Gautam et al. [6] as a self-explainable model. Distance-based classification makes the decision more transparent and class distributions are divided into clusters. The ability to decode embeddings including prototypes fosters transparency w.r.t. learned data distribution. In this work, modifications enable more direct distance-based classification and enforce an enclosed ID region making it ideal for OOD detection.
Yang et al. [24] categorize OOD detection methods applied post-hoc, requiring training, Outlier Exposure, pre-processing, or data augmentation. Yang et al. [25] also distinguish approaches based on outputs of a classifier (classificationbased), modeling the data distribution (density-based/generative), relying on distances in feature space (distance-based), and reconstructing the input measuring a reconstruction error (reconstruction-based). The approach of this work can be considered reconstruction-, distance- and density-based. Maximum Softmax Probability (MSP) as a baseline OOD score was examined by Hendrycks and Gimpel [11]. Hendrycks et al. [10] use the maximum logit as a score (post-hoc). Sun et al. [20] propose thresholding activations of the penultimate layer thus eliminating overconfidence caused by extreme activations. Wang et al. [22] design a virtual logit based on the smallest principle components. Gal and Ghahramani [5] apply Monte-Carlo dropout during test time and Lakshminarayanan et al. [13] train an ensemble of ANNs. Hendrycks et al. [12] propose a training-time augmentation based on fractals (PixMix ).
Nalisnick et al. [15] find that density estimates might assign higher likelihoods to OOD than to ID data. Xiao et al. [23] tackle this by retraining a VAE-encoder for a specific test-sample measuring a likelihood discrepancy. Sun et al. [19] design a VAE with one Gaussian distribution per class. In contrast to this work, no perceptual metric, distance-based classification, or restriction-scheme for the ID region is used. Moreover, a custom probability is defined for a sample being part of a class distribution. There is a fixed threshold for the latter in contrast to the flexible OOD score fusion used in this work without a fixed threshold for one of the scores alone. ARPL [2] generates near-OOD samples for learning adversarial reciprocal points representing individual negative classes.
Reconstructive OOD detection often involves elaborate schemes[3,16,1,27,7] as the reconstruction error alone often cannot separate OOD from ID data [3]. Existing approaches combine reconstruction error with Mahalanobis distance [3], improve ID reconstruction with a deformation transformation[1] or use multiple reconstruction errors [16,7]. In [27], the latent space region of an AE to which ID samples are encoded ( $I D$ region) is estimated by restricting ID data within the latent space. For OOD samples mapped into this region, the reconstruction error will be higher [27]. In contrast, in this work, an enclosing restriction supports the trade-off between reliable estimation of the ID region and reconstruction quality.
Distance-based OOD detection involves Mahalanobis distance[14] and kNearest Neighbor (KNN) distance for pre-trained features. Requiring training, Deep SVDD [17] maps ID data into a hypersphere, and SIREN [4] discriminatively shapes representations using prototypes but not reconstruction.
# 3 Methodology
We introduce the Prototypical Direct-Distance-Classifier VAE (ProtoDistVAE) for explainable OOD detection which extends the ProtoVAE from [6] and further incorporates the principle of AE-based OOD detection from [27]. Following [27], if an AE reconstructs every ID sample sufficiently well and the ID region $\tau _ { \mathrm { I D } }$ can be estimated precisely, a sample can be concluded to be ID by fulfilling two conditions:
Fig. 1: ProtoDistVAE architecture: The input $_ { \ast }$ is encoded into a latent Gaussian distribution from which a sample $\textit { \textbf { z } }$ is drawn and reconstructed to obtain $\hat { \pmb x }$ . Then, in the framework of generalized Gaussians, the SoftMax function returns the predicted probabilities and class estimate $\hat { y }$ for the distances to all prototypes.
1. An ID sample is embedded into $\tau _ { \mathrm { I D } }$ (by definition).
2. An ID sample exhibits a small reconstruction error.
Under the given assumptions, OOD samples should never fulfill both conditions.
Our aim is to model a distribution of data that is representative for a set of prototypes. This means that different classes or parts of classes can be assigned to different sub-distributions during training, thus potentially increasing data diversity and simplifying the OOD detection. A distance metric space is learned where similar samples are in close proximity to each other. Similar to [6], we use an encoder $f _ { \psi }$ , a decoder $g _ { \boldsymbol { \theta } }$ and prototypes $\phi _ { k j } \in \mathbb { R } ^ { L }$ in an end-to-end trainable fashion (see Figure 1). The rows of matrix $\pmb { \varPhi } _ { k } \in \mathbb { R } ^ { J \times L }$ describe the $J$ prototype vectors of class $k \in K$ classes.
Given a training dataset $\mathcal { D } = \{ ( \pmb { x } ^ { 1 } , ( \pmb { x } ^ { 1 } , y ^ { 1 } ) ) , . . . , ( \pmb { x } ^ { N } , ( \pmb { x } ^ { N } , y ^ { N } ) ) \}$ with $N$ labeled samples, the input $\mathbf { \Delta } _ { \mathbf { \boldsymbol { x } } ^ { i } }$ itself yields the target variables for reconstruction and a class label $y ^ { i }$ . The model is trained as a VAE learning a Gaussian mixture distribution where the encoder embeds the input $\mathbf { \Delta } _ { \mathbf { \boldsymbol { x } } ^ { i } }$ to a posterior Gaussian distribution $p ( z | \boldsymbol { \mathbf { \mathit { x } } } ^ { i } ) = \mathcal { N } ( z ; \boldsymbol { \mu } ^ { i } , \mathrm { d i a g } ( ( \boldsymbol { \sigma } ^ { i } ) ^ { 2 } ) )$ in the latent domain. During training, a latent representation $z ^ { \ i }$ is sampled whereas during inference, the mean value is used for the latent representation the decoder processes into the image space reconstruction $\hat { \pmb x } ^ { i }$ .
For classification, the Euclidean distances of the latent variable to all prototypes are computed (Equation (1)) and the minimum distance of each class yields the closest prototype. It is important to minimize the distance of an embedding to only one prototype distribution during training. The distances are transformed into logits by the generalized Gaussian distribution for enclosing restriction and are fed into a SoftMax function to obtain a purely distance-based, latent space classification without a learnable classifier.
$$
\begin{array} { c } { { d ( z ^ { i } , \phi _ { k j } ) = d _ { k j } ^ { i } = \| z ^ { i } - \phi _ { k j } \| _ { 2 } } } \\ { { P _ { \psi } ( y = k | x ^ { i } ) = \frac { \exp \left( l _ { k } ^ { i } \right) } { \sum _ { k ^ { \prime } = 1 } ^ { K } \exp \left( l _ { k ^ { \prime } } ^ { i } \right) } ~ , ~ l _ { k ^ { \prime } } ^ { i } = - \left( \frac { d _ { k ^ { \prime } j ^ { * } } ( k ^ { \prime } ) } { \alpha } \right) ^ { \beta } } } \\ { { j ^ { * } ( k ) = \mathrm { a r g m i n } _ { j } ( d _ { k j } ) } } \end{array}
$$
The original ProtoVAE architecture uses a linear classifier and distance-based similarity scores [6]. Similarity scores exhibit large gradients for embeddings close to a prototype which potentially leads to embeddings collapsing into the respective prototype position, and thus to degradation of reconstruction quality when different embeddings are not encoded differently. As a remedy, ProtoDistVAE uses an enclosing restriction leading to weaker gradients close to prototypes. Embeddings shall be trapped in a certain $I D$ region, but inside, the coding of embeddings shall be unconstrained. For this reason, generalized Gaussian distributions are used in the classification layer where $\alpha$ defines the width of the distribution and $\beta \geq 2$ controls the shape and "enclosedness" of the distribution.
In order to not distort the distance metric space, ProtoDistVAE uses distances more explicitly for classification. The linear classifier which essentially calculates a sum of distances is replaced by using only the minimum distances to prototypes per class. These are translated into logits $l _ { k ^ { \prime } } ^ { i }$ by the framework of generalized Gaussians and probabilities using the SoftMax function (Equation (2)). Cross-entropy is then applied to the modified predicted probabilities. $j ^ { * } ( k )$ is the nearest prototype within class $k$ while $d ^ { * }$ is the minimum distances vector for every class. Thus, instead of a sum of distances to multiple prototypes, the distance to only one prototype is minimized for a specific embedding.
The overall loss consists of a sum of four terms: The cross-entropy loss $\mathcal { L } _ { \mathrm { c l s } } ^ { \prime }$ shown in Equation (4) provides label information to enable the network to extract useful embeddings for discrimination and minimize the embedding distance to prototypes of the correct class. Each class is modeled by a mixture of $J$ normal distributions centered around the respective class prototypes for VAE-like distribution estimation and Kullback-Leibler divergence (KL divergence) w.r.t. the nearest prototype distribution of the correct class is computed to obtain the loss $\mathcal { L } _ { \mathrm { K L } } ^ { \prime }$ (Equation (5)). The reconstruction loss aims to recover the input samples [6] by separating groups of samples near each other for a better reconstruction. We use the LPIPS metric [26] for this task as it gives a more robust similarity between images than traditional metrics as e.g. mean squared error (MSE) by using a calibrated pre-trained network aligned towards human perception [26].
In order to prevent the collapse of prototypes of a class, an orthonormalization loss ${ \mathcal L } _ { \mathrm { o r t h } }$ (Equation (7)) is used to encourage prototypes within a class (after subtracting their mean $\phi _ { k }$ ) to be orthonormal to each other [6]. It is defined as the average of the class-wise Frobenius norms $\| \cdot \| _ { F }$ .
$$
\begin{array} { r l } & { \qquad \mathcal { L } _ { \mathrm { c l s } } ^ { \prime } ( \psi , \Phi ; { \pmb x } ^ { i } , { \pmb k } ) = - \log P _ { \psi } ( y = k | { \pmb x } ^ { i } ) } \\ & { \mathcal { L } _ { \mathrm { K L } } ^ { \prime } ( \psi , \pmb { \mathscr { F } } _ { k } ; { \pmb x } ^ { i } , { k = \pmb y } ^ { i } ) = D _ { K L } \big ( \mathcal { N } ( { \pmb \mu } ^ { i } , \mathrm { d i a g } ( ( { \pmb \sigma } ^ { i } ) ^ { 2 } ) ) \| \mathcal { N } ( \phi _ { k j ^ { * } ( k ) } , { \pmb I } _ { L } ) \big ) } \\ & { \qquad \mathcal { L } _ { \mathrm { r e c } } ^ { \prime } ( \psi , { \pmb \theta } ; { \pmb x } ^ { i } , { \hat { \pmb x } } ^ { i } ) = e _ { \mathrm { L P I P S } } ( { \pmb x } ^ { i } , { \hat { \pmb x } } ^ { i } ) } \\ & { \qquad \mathcal { L } _ { \mathrm { o r t h } } ( \pmb \Phi ) = \displaystyle \frac { 1 } { K } \sum _ { k = 1 } ^ { K } \| \tilde { \pmb { \phi } } _ { k } \tilde { \pmb { \phi } } _ { k } ^ { T } - { \pmb I } _ { J } \| _ { F } ^ { 2 } , \ \tilde { \pmb { \phi } } _ { k } = ( \phi _ { k j } - \bar { \phi } _ { k } ) _ { j = 1 . . } J } \end{array}
$$
In summary, ProtoDistVAE introduces LPIPS [26] as reconstruction loss and replaces the linear classifier layer as well as similarity scores by direct minimum distances and the framework of generalized Gaussians to implement an enclosing
restriction loss. The complete loss function is:
$$
\mathcal { L } = w _ { \mathrm { c l s } } \mathcal { L } _ { \mathrm { c l s } } + w _ { \mathrm { K L } } \mathcal { L } _ { \mathrm { K L } } + w _ { \mathrm { r e c } } \mathcal { L } _ { \mathrm { r e c } } + w _ { \mathrm { o r t h } } \mathcal { L } _ { \mathrm { o r t h } }
$$
For OOD detection, a distance-based OOD score and the LPIPS reconstruction error are merged. During experimentation, we found that the minimum distance to the next prototype can be improved by using the MSP score $\begin{array} { r } { \lambda _ { \mathrm { M S P } } = \operatorname* { m a x } _ { k } P _ { \psi } ( y = k | \pmb { x } ^ { i } ) } \end{array}$ in the ProtoDistVAE context which is the probability that an embedding belongs to the most likely generalized Gaussian under condition that it is ID. As ProtoDistVAE relies on distances for classification, MSP is also distance-based. Also the $\begin{array} { r } { \lambda _ { \mathrm { D i s t R a t i o } } = \sum _ { j } d _ { \widehat { k } j } / ( \sum _ { k } \sum _ { j } d _ { k j } ) } \end{array}$ is applied where $\widehat { k }$ indicates the predicted class. We assume thesbe scores perform better than the bminimum distance because the class distribution in the latent space might be skewed and OOD samples are embedded between different class regions.
For fusion of scores, one distance score and one reconstruction error are normalized w.r.t. to their validation set distributions to make them comparable using a lower and upper percentile of the score distribution to obtain the normalized score $\widetilde { \lambda } ( \pmb { x } ) = ( \lambda ( \pmb { x } ) - \lambda _ { \mathrm { l o w e r } } ) / ( \lambda _ { \mathrm { u p p e r } } - \lambda _ { \mathrm { l o w e r } } )$ . Both score types are combined into eone score using $L _ { 2 }$ or $L _ { \infty }$ norm: $\lambda _ { L _ { p } } ( \pmb { x } ) = \| ( \widetilde { \lambda } _ { 1 } ( \pmb { x } ) , \widetilde { \lambda } _ { 2 } ( \pmb { x } ) ) ^ { T } \| _ { p }$ where $p$ denominates the degree. The $L _ { \infty }$ norm tends to reflecte a hared decision (e.g. at least one score is above its threshold) and the $L _ { 2 }$ norm a flexible decision (one score is too high or both together are rather high and therefore indicates an OOD sample). This type of fusion means that no probabilities need to be modeled explicitly and thus avoids any need for modeling assumptions.
# 4 Experimental Results
For numerical evaluation, we compare our approach to the state-of-the-art based on the OpenOOD benchmark [24] and a non-public dataset from the railway domain (DBS dataset). A general advantage of the proposed method is that it allows human insights into the training distribution and decision-making of the network by reconstructing samples, prototypes, and distances in the latent space which supports its usage in safety-critical domains.
General Experimental Setup The OpenOOD benchmark provides implementations of state-of-the-art approaches for comparison and defines sub-benchmarks according to the ID datasets MNIST, CIFAR10, CIFAR100, and ImageNet. Another dataset is then used as OOD data. Datasets are labeled as near OOD or far OOD according to their ID similarity, e.g. if they have similar color distributions. Open Set Recognition (OSR) is also provided by partitioning a dataset into ID and OOD classes. M-6 benchmark is based on MNIST, C-6 on CIFAR-10, C-50 on CIFAR-100, and T-20 on TinyImageNet with the numeral representing the number of ID classes.
The DBS dataset was collected from video recordings of a camera mounted on a commuter train in a typical operation. Object proposals were automatically collected and classified into trains and persons. The annotations were manually checked and OOD samples (i.e. false positive detections) were placed in a separate category. In our evaluation, we used 8351 samples of people, 8340 samples of trains, and 5001 non-objects labeled as OOD, all rescaled to size $6 4 \times 6 4$ . Person and train samples were divided equally into training (60%), validation (10%), and test (30%) splits (OOD samples used only for testing). We use $J { = } 1$ prototype per class in all experiments as a higher number did not improve the performance.
Table 1: OOD detection performance (AUROC in $\%$ ) on OpenOOD benchmark and CIFAR-100 ID accuracy ( $\%$ ) for different approaches: Best performances marked in bold. Results from other methods taken from [24].
The generalized Gaussian parameters $\alpha$ and $\beta$ were both set to 2 for all experiments. The encoder was chosen as ResNet-50 [9] for ImageNet and as ResNet-18 for all benchmarks with $6 4 \times 6 4$ sized images (including the DBS dataset) and $3 2 \times 3 2$ sized images. A convolutional encoder with five layers was used for all $2 8 \times 2 8$ sized images, for the decoder a five-layered network using subpixel-convolutions [18] is used. For ImageNet the decoder consists of seven layers and for all other benchmarks, it consists of six layers. The latent dimensionality $L$ is chosen as $1 / 3$ , $1 / 2 4$ or $1 / 9 6$ of the input dimensionality. After training, ID validation data were used for normalization of the OOD scores, which are used afterwards for score fusion during testing. For evaluation, ID classification performance is measured in accuracy and OOD detection performance in Area Under the Receiver Operating Characteristic (AUROC). AUROC is a threshold-independent metric and measures how well a score separates ID and OOD.
# 4.1 OOD Detection Performance
Table 1 shows the OOD detection performance in terms of AUROC compared to state-of-the-art methods. ProtoDistVAE was trained using only LPIPS reconstruction loss with weight $w _ { \mathrm { r e c } } = 1$ . Cross-entropy and KL divergence loss were used similarly with a weight of $w _ { \mathrm { c l s } } = w _ { \mathrm { K L } } = 1$ . Distance ratio $\lambda _ { \mathrm { D i s t R a t i o } }$ and LPIPS $\lambda _ { \mathrm { L P I P S } }$ were used as scores to be fused by $L _ { \infty }$ norm. The latent space dimensionality $L$ was chosen as $1 / 2 4$ of the input dimensionality.
Compared to the other methods, ProtoDistVAE performs best on the MNISTbased benchmarks. This is likely due to its low diversity, making it easier to learn a latent distribution. For CIFAR10, ProtoDistVAE performs on par with other methods. However, the performance for highly diverse datasets with a large number of classes decreases as ID estimation and classification are performed in the same latent space and may impair each other. Similarly, higher resolutions lead to difficulties for ProtoDistVAE in detecting OOD samples, likely due to the increased complexity of reconstruction.
Fig. 2: UMAP visualization of the latent space embeddings of trained ProtoDistVAEs. (a) On MNIST, color-coded classes are clearly separated. (b) On CIFAR10, clusters blend into each other. (c) ID (CIFAR10) versus OOD (CIFAR100): embedding of OOD samples appears mainly between class prototypes.
Figure 2 shows further insights through an Uniform Manifold Approximation and Projection (UMAP) visualization of the latent space and illustrates how our method allows understanding its decision-making. The method works best in cases of clearly separable datasets and performs worse if data cannot be attributed well to clusters extracted. However, it should be mentioned that CIFAR10 vs. CIFAR100 is generally a hard OOD benchmark. ID samples in the space between prototypes might be interesting for further analysis since these exhibit a higher uncertainty and could be exploited by active learning or for identifying a sample with very different attributes within a class.
Table 2a shows some results on the DBS dataset. Here, an increased weight on LPIPS ( $w _ { \mathrm { r e c } } = 1 0 0$ ) was used to improve the OOD detection performance without harming classification accuracy. The accuracy is on par with other methods, likely due to only two classes being available. For OOD detection, PixMix and ProtoDistVAE perform best, while VIM and KNN also show good results. Combining $\lambda _ { \mathrm { L P I P S } }$ with $\lambda _ { \mathrm { M S P } }$ further improves the results with a gain of $0 . 9 \%$ .
ProtoDistVAE performs well on the DBS dataset due to its composition. The data samples are often quite similar as trains and persons are captured from the same angles and there are little variations e.g. in perspective, weather, lighting, and color. In comparison, ImageNet shows more inconsistent data with more diverse appearances across the same class. ProtoDistVAE benefits from a reduced intra-class variance and “complete” data distribution which allows it to model the data more easily. Hypothetically, it is easier for the network to recognize systematics in the data. PixMix augmentation seems to benefit from a complete distribution and even further increases the diversity of the data. However, the data distribution is not represented in the model and classification is not transparent. Other methods perform worse: Ensembling shows a lowerthan-usual performance as it depends on variations in the prediction of individual networks and these variations are weaker due to low data diversity in this dataset. Methods depending purely on classification-based schemes might suffer from overconfidence due to easier classification across only two classes and low data diversity. ProtoDistVAE, however, does not overfit for classification and aims to learn a representation of the data. In addition, the reconstruction error helps it to identify overconfidently classified samples mapped into its ID-region.
Table 2: Experimental results of OOD detection in AUROC ( $\%$ ) and ID accuracy ( $\%$ ): (a) DBS dataset results of state-of-the-art methods (parameterized as in [24]) compared to ProtoDistVAE with LPIPS score combined by $L _ { \infty }$ fusion with DistRatio and MSP, respectively. (b) ProtoVAE vs. ProtoDistVAE. (c) Influence of reconstruction loss when using LPIPS as OOD score.
(a) DBS dataset
(c) OpenOOD benchmark (partial): reconstruction loss
(b) OpenOOD benchmark: ProtoVAE vs. ProtoDistVAE using MSP score
# 4.2 Ablation Study: ProtoVAE vs. ProtoDistVAE
Comparing the proposed ProtoDistVAE architecture to the base ProtoVAE , the reconstruction loss was set to a constant level. This does not change reconstruction error-based OOD detection according to the observed data. Table 2b shows detection results for ProtoVAE and ProtoDistVAE using the distancebased MSP score based on the predicted probabilities. Note that an improved distance-based score potentially increases performance even further when fused with a reconstruction error score. ProtoDistVAE outperforms ProtoVAE in almost all benchmarks for OOD detection and for different values of the latent dimension $L$ which can be explained by the direct use of distances for classification and the enclosing restriction used during training. The latter actively shapes the ID-region by trapping the ID embeddings in the proximity of the class-specific prototypes. Furthermore, the results display the importance of the latent dimensionality $L$ for both networks. Different values for $L$ are optimal for different levels of complexity reflected in different datasets. Too low values reduce the information coded in the representation while too high values inhibit a clear assignment of samples to class prototypes.
Fig. 3: Comparison of MSE and LPIPS loss: CIFAR10 (ID) and FashionMNIST (OOD). From top to bottom: Input, reconstruction (MSE), and reconstruction (LPIPS). ( $L = 3 2$ )
# 4.3 Reconstruction
Table 2c shows OOD detection performance using the LPIPS score based on ProtoDistVAE trained with either MSE or LPIPS loss. In contrast to using the MSE score which showed a generally lower performance (results not shown), the LPIPS score can achieve good detection results, even when training with MSE reconstruction loss. However, using LPIPS as reconstruction loss outperforms MSE loss. A special case is the ImageNet benchmark which is different due to image size and data diversity. The reconstruction performance for MSE and LPIPS loss on the CIFAR10 benchmark is depicted in Figure 3. ProtoDistVAE trained with MSE shows significant blur, regardless of ID or OOD samples. Training with LPIPS helps to preserve more semantic information and leads to differences when reconstructing OOD samples.
Figure 4 displays reconstructions of the DBS dataset. ProtoDistVAE appears to have learned the data distribution and can reconstruct ID better than OOD in most cases. It successfully distinguishes the class distributions of persons and trains and can show the features associated with a certain sample. For example, images of train stations and regular structures are often associated with trains, whereas background images are often reconstructed into person-like images. The learned prototypes of ProtoDistVAE can also be reconstructed. As Figure 5 shows, prototypes can be better extracted from datasets with low-variance datasets like MNIST and the DBS dataset while for datasets with higher diversity like CIFAR10, prototypes are harder to extract and images are less expressive. Human observers can thus assess which properties the network extracted from the data and evaluate features associated across classes. | Understanding the decision-making and trusting the reliability of Deep
Machine Learning Models is crucial for adopting such methods to safety-relevant
applications. We extend self-explainable Prototypical Variational models with
autoencoder-based out-of-distribution (OOD) detection: A Variational
Autoencoder is applied to learn a meaningful latent space which can be used for
distance-based classification, likelihood estimation for OOD detection, and
reconstruction. The In-Distribution (ID) region is defined by a Gaussian
mixture distribution with learned prototypes representing the center of each
mode. Furthermore, a novel restriction loss is introduced that promotes a
compact ID region in the latent space without collapsing it into single points.
The reconstructive capabilities of the Autoencoder ensure the explainability of
the prototypes and the ID region of the classifier, further aiding the
discrimination of OOD samples. Extensive evaluations on common OOD detection
benchmarks as well as a large-scale dataset from a real-world railway
application demonstrate the usefulness of the approach, outperforming previous
methods. | [
"cs.LG",
"cs.CV"
] |
# 1 INTRODUCTION
Facial micro-expression recognition (MER) is a popular task in the fields of computer vision and affective computing [1]. It has applications in wide areas such as medicine, education, and criminal investigation. Micro-expressions (MEs) are subtle and involuntary that convey genuine emotions [2], and contribute to the recognition of mental condition or deception of humans. Different from macroexpressions [3], [4], MEs are fine-grained and last only for a very short interval of time, i.e. not more than 500 milliseconds [5]. In literature, MER remains a challenging
Manuscript received April, 2023. (Corresponding authors: Yifan Cheng, Yong Zhou, and Lizhuang Ma.)
Fig. 1. Illustration of optical flow and facial landmark differences between two consecutive frames ${ \bf { I } } _ { k }$ and ${ \bf \cal I } _ { k + 1 }$ . We use a color coding to visualize the optical flow, in which the color of each point in the color coding denotes its displacement including orientation and magnitude to the central point. Although facial subtle muscle actions from ${ \bf { I } } _ { k }$ to ${ \bf \cal I } _ { k + 1 }$ are hard to perceive by human eyes, they are reflected in optical flow and facial landmark differences.
problem due to the short duration, subtlety, and small-scale and low-diversity datasets of MEs.
One typical way is to extract hand-crafted features containing correlated ME information. Typical hand-crafted features include optical flow and histogram of oriented optical flow (HOOF) [6] with motion pattern, local binary patterns from three orthogonal planes (LBP-TOP) [7] with spatio-temporal information, and histogram of oriented gradients (HOG) [8] and histogram of image gradient orientation (HIGO) [9] with local contrast information. However, these features have limited robustness on challenging MEs with short-duration and inconspicuous motions. Besides, key frames like onset, apex, and offset frames of MEs are sometimes required in feature extraction [10].
Another popular solution involves the use of prevailing deep neural networks. Khor et al. [11] first combined the optical flow, the derivatives of the optical flow, and the raw images as input, then used a convolutional neural network (CNN) to extract the feature of each frame and used long short-term memory (LSTM) modules to learn the temporal dynamics. However, this method relies on the pre-extracted optical flow. Reddy et al. [12] adopted a 3D CNN to extract features from both spatial and temporal domains, in which the performance is limited by insufficient training samples. Xia et al. [13] employed macro-expression recognition as an auxiliary task, in which macro-expression recognition network is used to guide the fine-tuning of MER network from both label and feature space. However, fine-grained information is not explicitly emphasized in this method.
The above methods suffer from limited capacity of handcrafted features, requirement of key frames, or fail to thoroughly exploit the feature learning ability of deep networks due to insufficient training data. To tackle these limitations, we propose to integrate automatic feature learning from raw frame sequence, capturing of facial motion information, and localization of facial fine-grained characteristics into an endto-end framework. Considering the prevailing multi-task learning technique is convenient to guide and assist the training of main task, we design a novel micro-action-aware deep learning framework called MOL that jointly models MER, optical flow estimation, and facial landmark detection via transformer-graph-style convolution. As illustrated in Fig. 1, the two latter tasks are beneficial for capturing facial subtle muscle actions associated with MEs, which relaxes the requirement of large-scale training data. Moreover, we propose a novel F5C block to directly extract local-global features from raw images, which is combined by our proposed fully-connected convolution and channel correspondence convolution. The transformer-style fully-connected convolution can extract local features while maintaining global receptive fields, and the graph-style channel correspondence convolution can model the correlations among feature map channels. Finally, we feed a sequence of pair features composed of the local-global features of consecutive two frames into a 3D CNN to achieve MER. The use of pair features rather than frame features contributes to preserving each sub-action clip, which can also be regarded as the sliding windows. The entire framework is end-to-end without any post-processing operation, and all the modules are optimized jointly.
The contributions of this paper are threefold:
We propose a micro-action-aware joint learning framework of MER, optical flow estimation, and facial landmark detection, in which pre-extracted features as well as prior knowledge of key frames are not required. To our knowledge, joint modeling of automatic ME feature learning from raw frame sequence, facial motion information capturing, and facial fine-grained characteristic localization via deep neural networks has not been done before.
We propose a new local-global feature extractor named F5C composed by fully-connected convolution and channel correspondence convolution, which integrates the advantages of transformer, graph convolution, and vanilla convolution. Extensive experiments on benchmark datasets show that our method outperforms the state-of-the-art MER approaches, achieves competitive performance for both optical flow estimation and facial landmark detection, and can capture facial subtle muscle actions in local regions related to MEs.
# 2 RELATED WORK
We review the previous works those are closely related to our method, including hand-crafted feature based MER, deep learning based MER, and MER with combination of hand-crafted feature and deep learning.
# 2.1 Hand-Crafted Feature Based MER
Earlier works propose hand-crafted features to try to capture fine-scale ME details. LBP-TOP [7] is a typical handcrafted feature, which combines temporal information with spatial information from three orthogonal planes. Later, Ben et al. [14] employed hot wheel patterns from three orthogonal planes (HWP-TOP) to make the most of the directional information. Besides, Wang et al. [15] proposed local binary patterns with six intersection points (LBP-SIP) to avoid repeated coding in LBP-TOP. Another widely used feature is histogram of oriented gradients (HOG) [8], which computes gradients of image pixels. A histogram of image gradient orientation (HIGO) [9] feature is further proposed, which can maintain the invariance of geometric and optical transformation of images.
Optical flow describes the action pattern of each pixel from one frame to another frame, which is highly related to MEs. Happy et al. [16] improved histogram of oriented optical flow (HOOF) [6] as FHOOF by collecting the action directions into angular bins based on the fuzzy membership function, and also extended FHOOF to be fuzzy histogram of optical flow orientations (FHOFO) by ignoring the action magnitude in computation. Liong et al. [10] introduced biweighted oriented optical flow (Bi-WOOF) to encode essential expressiveness of the apex frame in ME videos.
However, the extraction process of hand-crafted features often discards important information, in which the characteristics of subtle and diverse MEs are hard to be modeled. Besides, key frames of MEs are often required, which limits the applicability.
# 2.2 Deep Learning Based MER
Recently, the prevailing deep learning technique has been applied to MER. Reddy et al. [12] employed a 3D CNN to achieve MER, which extracts spatial and temporal information from raw image sequences. Lei et al. [17] extracted shape representations based on facial landmarks, and then adopted a graph-temporal convolutional network (Graph-TCN) to capture local muscle actions of MEs. Wei et al. [18] proposed an attention-based magnification-adaptive network (AMAN), in which a magnification attention module is used to focus on appropriate magnification levels of different MEs, and a frame attention module is used to focus on discriminative frames in a ME video.
Fig. 2. The architecture of our MOL framework. Given a sequence of $t$ frames $\{ \mathbf { I } _ { 0 } , \mathbf { I } _ { 1 } , \cdot \cdot \cdot , \mathbf { I } _ { t - 1 } \}$ , MOL first extracts rich feature $\mathbf { F } _ { k } ^ { \left( r \right) }$ of each frame ${ \bf \cal I } _ { k }$ by a stack of vanilla convolutional layers. For each pair of consecutive frames $\left\{ \mathbf { I } _ { k } , \mathbf { I } _ { k + 1 } \right\}$ , $\mathbf { F } _ { k } ^ { \left( r \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( r \right) }$ are then fed into the same F5C block to extract local-global features $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ , respectively. Afterwards, $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ is fed into a facial landmark detection module to predict facial landmark locations $\hat { \mathbf { l } } _ { k + 1 }$ of the frame $\mathbf { I } _ { k + 1 }$ , while $\dot { \mathbf { F } } _ { k } ^ { ( g ) }$ , $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ , ${ \bf \cal I } _ { k }$ , and $\mathbf { I } _ { k + 1 }$ are simultaneously fed into an optical flow estimation module to predict optical flow $\hat { \mathbf { O } } _ { k }$ including horizontal component and vertical component. $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ are further concatenated to be $\mathbf { F } _ { k } ^ { \left( c \right) }$ as the feature of the $k$ -th pair. Finally, the sequence of $t - 1$ pair features $\{ \mathbf { F } _ { 0 } ^ { ( c ) } , \mathbf { F } _ { 1 } ^ { ( c ) } , \cdot \cdot \cdot , \mathbf { F } _ { t - 2 } ^ { ( c ) } \}$ is fed into a MER module to predict the ME category.
Besides single MER task based methods, some works incorporate auxiliary tasks correlated with MER into a deep multi-task learning framework. Since action units (AUs) describe facial local muscle actions [19], [20], Xie et al. [21] proposed an AU-assisted graph attention convolutional network (AU-GACN), which uses graph convolutions to model the correlations among AUs so as to facilitate MER. Xia et al. [13] used macro-expression recognition as an auxiliary task, in which macro-expression recognition network can guide the fine-tuning of MER network from both label and feature space.
Different from the above methods, we employ an end-toend deep framework for joint learning of MER, optical flow estimation, and facial landmark detection.
# 2.3 MER with Combination of Hand-Crafted Feature and Deep Learning
Considering deep networks are limited by small-scale and low-diversity ME datasets, some approaches combine handcrafted features with deep learning framework. Verma et al. [22] proposed a dynamic image which preserves facial action information of a video, and input the dynamic image to a lateral accretive hybrid network (LEARNet). Nie et al. [23] also generated the dynamic image of the input video, and input it to a dual-stream network with two tasks of MER and gender recognition.
Another commonly used hand-crafted feature is optical flow. Zhou et al. [24] calculated the optical flow between onset and apex frames of the input ME video, in which its horizontal and vertical components are fed into a dualinception network to achieve MER. With the same input setting, Shao et al. [25] achieved AU recognition and MER simultaneously, in which AU features are aggregated into ME features. Besides, Hu et al. [26] fused local Gabor binary pattern from three orthogonal panels (LGBP-TOP) feature and CNN feature, and then formulated MER as a multi-task classification problem, in which each category classification can be regard as a one-against-all pairwise classification problem.
All these methods require pre-extracted hand-crafted features, in which the representation power of deep networks is not thoroughly exploited. In contrast, our network directly processes raw images, and contains a novel localglobal feature extractor. Besides, instead of treating optical flow estimation as a preprocessing, we put it into a joint framework to guide the capturing of facial subtle motions.
# 3 MOL FOR JOINT ESTIMATION OF MICROEXPRESSION, OPTICAL FLOW AND LANDMARK
# 3.1 Overview
Given a video clip with $t$ frames $\{ \mathbf { I } _ { 0 } , \mathbf { I } _ { 1 } , \cdot \cdot \cdot , \mathbf { I } _ { t - 1 } \} ,$ , our main goal is to design a micro-action-aware deep learning framework to predict ME category of the overall clip, facial landmark locations $\{ \hat { \bf l } _ { 1 } , \hat { \bf l } _ { 2 } , \cdot \cdot \cdot \hat { \bf \Delta } , \hat { \bf l } _ { t - 1 } \}$ of the last $t - 1$ frames, and optical flow $\{ \hat { \mathbf { O } } _ { 0 } , \hat { \mathbf { O } } _ { 1 } , \cdot \cdot \cdot , \hat { \mathbf { O } } _ { t - 2 } \}$ of the $t - 1$ consecutive frame pairs $\{ ( \mathbf { I } _ { 0 } , \mathbf { I } _ { 1 } ) , ( \mathbf { I } _ { 1 } , \mathbf { I } _ { 2 } ) , \cdot \cdot \cdot , ( \mathbf { I } _ { t - 2 } , \mathbf { I } _ { t - 1 } ) \}$ . We choose to directly process raw video clips without the dependence on hand-crafted features, and discard additional limitations like the prior knowledge of onset and apex frames. Fig. 2 illustrates the overall structure of our MOL framework.
A stack of vanilla convolutional layers are first used to extract rich feature $\mathbf { F } _ { k } ^ { \left( r \right) }$ of the $k$ -th frame ${ \bf \cal I } _ { k }$ in the input video, respectively. TABLE 1 shows the detailed architecture of this module. Then, for each pair of consecutive frames $\{ \mathbf { I } _ { k } , \mathbf { I } _ { k + 1 } \} ,$ , an F5C block is used to learn local-global features $\mathbf { F } _ { k } ^ { \left( g \right) }$ and F(kg+)1, respectively. The local-global features are shared by three tasks for joint learning, in which optical flow estimation and facial landmark detection as auxiliary tasks are devised for promoting the main task MER in temporal and spatial domains, respectively.
TABLE 1 The structure of the stack of vanilla convolutional layers for extracting rich feature. $C _ { i n }$ and $C _ { o u t }$ denote the number of input channels and output channels, respectively.
To estimate the optical flow $\hat { \mathbf { O } } _ { k }$ between ${ \bf \cal I } _ { k }$ and $\mathbf { I } _ { k + 1 } ,$ we simultaneously feed $\mathbf { I } _ { k } , \mathbf { I } _ { k + 1 } , \mathbf { F } _ { k } ^ { ( g ) }$ , and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ into an optical flow estimation module. To predict the landmark locations lˆk+1 of Ik+1, we input F(kg+)1 to a landmark detection module. Finally, we feed a sequence of $t - 1$ pair features $\{ \mathbf { F } _ { 0 } ^ { ( c ) } , \mathbf { F } _ { 1 } ^ { ( c ) } , \cdot \cdot \cdot , \mathbf { F } _ { t - 2 } ^ { ( c ) } \}$ into a 3D CNN to−predict the ME category of the whole video clip, in which $\mathbf { \bar { F } } _ { k } ^ { \left( c \right) }$ is the concatenation of $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ . This use of pair features rather than frame features is beneficial for preserving each sub-action clip.
# 3.2 F5C Block
The architecture of our proposed F5C block is shown in the upper part of Fig. 2. We name this block as F5C because it consists of two main operations, fully-connected convolution (FCC) and channel correspondence convolution (CCC). FCC is developed from the conventional circular convolution [27] by integrating the style of the prevailing transformer [28], which can gather local information from local receptive fields like convolutions and extract global information from the entire spatial locations like self-attention [28]. CCC is designed to model the correlations among feature map channels in a manner of graph convolution [29]. Two residual structures [30] along with FCC and CCC are beneficial for mitigating the vanishing gradient problem. The design of F5C integrates the merits of transformer, graph convolution, and vanilla convolution.
# 3.2.1 Fully-Connected Convolution
It is known that vanilla convolution works well in extracting local features. We propose to enhance its ability of extracting global features from three aspects. First, similar to transformers [28], [31], we treat each column (in vertical direction) or each row (in horizontal direction) of the input as a patch, and apply positional embedding to patches to perceive contextual information. Second, we conduct circular convolution on each patch via fully-connected operation to enlarge the receptive field. Third, we perform operations in both vertical and horizontal directions to more completely cover regions. Such structure is named as transformer-style fully-connected convolution.
Fig. 3. The structure of our proposed transformer-style fully-connected convolution. An input feature map $\mathbf { x }$ with a size of $C \times H \times W$ is first processed by vanilla $1 \times 1$ convolution, and further goes through two branches, respectively, in which the first branch consists of FCC$\mathsf { v }$ and FCC-H in order while the second branch uses the reverse order. Then, the outputs of the two branches are concatenated along with $1 \times 1$ convolution to obtain the final output $\mathbf { Y }$ with the same size as X.
As shown in Fig. 3, an FCC is composed of two main components, FCC-V in vertical direction and FCC-H in horizontal direction. It uses two branches of FCC-H after FCC-V and FCC-V after FCC-H, and then fuses two outputs by concatenation and vanilla $1 \times 1$ convolution. In this way, the receptive field of FCC can cover positions in both vertical and horizontal directions so as to extract complete localglobal features.
Specifically, given an input $\mathbf { X } \in \mathbb { R } ^ { C \times H \times W } ,$ , we conduct the $1 \times 1$ convolution as a preprocessing. In FCC-V, we first employ a positional embedding [28] to make it aware of the position information:
$$
\mathbf { X } ^ { \left( v \right) } = \mathbf { X } \oplus ^ { v } \mathbf { P } ^ { \left( v \right) } ,
$$
where $\mathbf { P } ^ { ( v ) } \in \mathbb { R } ^ { C \times H }$ denotes the positional embedding, and $\oplus ^ { v }$ denotes element-wise sum operation, in which $\breve { \mathbf { P } ^ { ( v ) } }$ is expanded with $W$ times along horizontal direction so as to match the size of $\mathbf { X }$ . Then, the output $\mathbf { Y } ^ { ( v ) } \in \mathbb { R } ^ { C \times H \times W }$ at element $( c , i , j )$ is defined as
$$
Y _ { c , i , j } ^ { ( v ) } = \sum _ { s = 0 } ^ { H - 1 } U _ { c , s } ^ { ( v ) } X _ { c , ( i + s ) \% H , j } ^ { ( v ) } ,
$$
where $\%$ denotes the remainder operation, and ${ \bf U } ^ { \left( v \right) } \in \left( \begin{array} { l } { \mathbb { 1 } } \end{array} \right)$ $\mathbb { R } ^ { C \times H }$ is a learnable parameter. The elements of $\mathbf { X }$ in vertical direction are fully-connected in a circular manner, so we name this process as fully-connected convolution-vertical (FCC-V). We represent Eq. (2) as $\mathbf { Y } ^ { ( v ) } = \mathbf { U } ^ { ( v ) } \odot ^ { v } \mathbf { X } ^ { ( v ) }$ for simplicity.
Similarly, the process of FCC-H can be formulated as
$$
\mathbf { X } ^ { \left( h \right) } = \mathbf { X } \oplus ^ { h } \mathbf { P } ^ { \left( h \right) } ,
$$
$$
Y _ { c , i , j } ^ { ( h ) } = \sum _ { s = 0 } ^ { W - 1 } U _ { c , s } ^ { ( h ) } X _ { c , i , ( j + s ) \% W } ^ { ( h ) } ,
$$
where $\mathbf { P } ^ { ( h ) } \in \mathbb { R } ^ { C \times W }$ is the positional embedding, $\oplus ^ { h }$ denotes the element-wise sum operation by expanding $\mathbf { P } ^ { ( h ) }$ with $H$ times along vertical direction, $\breve { \mathbf { U } } ^ { ( h ) } \in \mathbb { R } ^ { C \times W }$ is a learnable parameter, and Eq. (3b) can be represented as Y(h) = U(h) h X(h) for simplicity.
# 3.2.2 Channel Correspondence Convolution
Since each feature map channel encodes a type of visual pattern [32], we propose the CCC to reason the relationships among feature map channels so as to further refine the extracted local-global features by FCC. The process of CCC is illustrated in the upper side of Fig. 2.
Inspired by the structure of dynamic graph convolution [33], we first construct a $k$ -nearest neighbors ( $k \mathrm { - }$ NN) [34] graph to find similar patterns. In particular, this directed graph is defined as $\mathcal { G } \overset { = } { = } ( \nu , \mathcal { E } )$ , where the vertex set $\mathcal { V } = \{ 0 , 1 , \cdots , C - 1 \}$ contains all the $C$ feature map channels, and the edge set $\mathcal { E } \subseteq \mathcal { V } \times \mathcal { V }$ . The size of the $i \cdot$ -th feature map channel is given by $H \times W .$ , and we reshape it to be an $H W$ -dimensional vector for the convenience of measuring similarity, denoted as $\mathbf { f } _ { i }$ . The neighbors of a vertex are chosen as the feature map channels with the top$k$ cosine similarities.
Given a directed edge $\mathbf { f } _ { i } \gets \mathbf { f } _ { j } , \mathbf { f } _ { j }$ is treated as a neighbor of $\mathbf { f } _ { i }$ . To obtain this edge feature $\mathbf { e } _ { i , j } \ \in \ \mathbb { R } ^ { H W } .$ , we incorporate the global information encoded by $\mathbf { f } _ { i }$ and the local neighborhood characteristics captured by $\mathbf { f } _ { j } - \mathbf { f } _ { i }$ :
$$
e _ { i , j , s } = \mathcal { R } { ( \mathbf { v } _ { s } ^ { ( 1 ) } } ^ { \top } \mathbf { f } _ { i } + { \mathbf { v } _ { s } ^ { ( 2 ) } } ^ { \top } ( \mathbf { f } _ { j } - \mathbf { f } _ { i } ) ) ,
$$
where $\mathcal { R } ( \cdot )$ denotes the rectified linear unit (ReLU) [35] function, $\mathbf { \bar { v } } _ { s } ^ { ( 1 ) } ~ \in ~ \mathbb { R } ^ { H W }$ and $\mathbf { v } _ { s } ^ { ( 2 ) } ~ \in ~ \mathbb { R } ^ { H W }$ are learnable parameters, $\top$ is used as the transpose of a vector, and $e _ { i , j , s }$ is the $s$ -th element of $\mathbf { e } _ { i , j }$ . Eq. (4) can be implemented by the convolution operation.
Finally, we adopt a maximum aggregation function to capture the most salient features:
Fig. 4. The structure of the optical flow estimation module, which consists of (a) an encoder and (b) a decoder.
temporal directions. The use of 3D max-pooling layer is to reduce the feature dimension while maintaining important information.
Considering MER is a classification task, we employ the cross entropy loss:
$$
\mathcal { L } _ { e } = - \sum _ { s = 0 } ^ { n - 1 } p _ { s } \log ( \hat { p } _ { s } ) ,
$$
where $n$ is the number of ME classes, and $\hat { p } _ { s }$ denotes the predicted probability that the sample is in the $s$ -th class. $p _ { s }$ denotes the ground-truth probability, which is 1 if the sample is in the $s \mathrm { . }$ -th class and is 0 otherwise.
$$
f _ { i , s } ^ { ( o ) } = \operatorname* { m a x } _ { \left\{ j \mid ( i , j ) \in \varepsilon \right\} } e _ { i , j , s } ,
$$
where $\mathbf { f } _ { i } ^ { ( o ) } \in \mathbb { R } ^ { H W }$ is the output of the $i$ -th feature map channel, and is further reshaped to the size of $H \times W$ and then is processed by a $1 \times 1$ convolution. With learnable parameters, our proposed CCC can adaptively model the correlations across feature map channels. As shown in Fig. 2 and Fig. 3, the input and output sizes of FCC and CCC, as well as their composed F5C are all $C \times H \times W$ . In this way, our proposed FCC, CCC, and F5C can all be used as plugand-play modules.
# 3.3 Joint Learning of Tasks
# 3.3.1 Micro-Expression Recognition
Since MEs are subtle and short-duration, our method needs to check potential sub-action clips between each two consecutive frames so as to avoid the loss of ME clues. In this case, we concatenate local-global features $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { ( g ) }$ of each pair of consecutive frames {Ik, Ik+1} to be F(k , and input the sequence of $\{ \mathbf { { F } } _ { 0 } ^ { ( c ) } , \mathbf { { F } } _ { 1 } ^ { ( c ) } , \cdot \cdot \cdot , \mathbf { { F } } _ { t - 2 } ^ { ( c ) } \}$ to a 3D CNN. This feature fusion strategy can also be regarded as an application of the sliding window mechanism.
The detailed structure is shown in the lower right corner of Fig. 2. It consists of a 3D convolutional layer and a 3D max-pooling layer, and is followed by a MER classifier with two fully-connected layers. In contrast to a 2D CNN operated in spatial domain, a 3D CNN uses 3D convolutional kernels to extract features in both spatial and
# 3.3.2 Optical Flow Estimation
Since MEs are subtle and low-intensity, it is difficult to extract related features from raw frames. Considering the optical flow contains motion information of facial muscles, which is strongly correlated to MEs, we use optical flow estimation as an auxiliary task to facilitate the learning of ME features.
The architecture of the optical flow estimation module is detailed in Fig. 4, which is based on FlowNet [36] with an encoder and a decoder. The inputs are two raw consecutive frames ${ \bf \cal I } _ { k }$ and $\mathbf { I } _ { k + 1 } ,$ as well as their local-global features $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ output by the F5C block. The encoder models the correlations between two frames and extracts multi-level features, in which the feature at each level is fed into the decoder for the final estimation of optical flow $\hat { \mathbf { O } } _ { k }$ . The optical flow estimation loss is defined as
$$
\mathcal { L } _ { f } = \frac { 1 } { t - 1 } \sum _ { k = 0 } ^ { t - 2 } M S E ( \mathbf { O } _ { k } , \hat { \mathbf { O } } _ { k } ) ,
$$
where $\mathbf { O } _ { k }$ denotes the ground-truth optical flow between ${ \bf { I } } _ { k }$ and $\mathbf { I } _ { k + 1 } ,$ and $M S E ( \cdot )$ denotes mean squared error (MSE) loss.
# 3.3.3 Facial Landmark Detection
Considering facial important regions like eyes and lips are closely related to MEs, we introduce another auxiliary task of facial landmark detection. The architecture of this task module is illustrated in the bottom part of Fig. 2, which contains one convolutional layer and two fully-connected layers. The facial landmark detection loss is defined as
$$
\begin{array} { r l r } { { \mathcal { L } _ { m } = \frac { 1 } { m ( t - 1 ) } \sum _ { k = 0 } ^ { t - 2 } \sum _ { s = 0 } ^ { m - 1 } ( \vert \ l _ { k + 1 , 2 s } - \hat { l } _ { k + 1 , 2 s } \vert + } } \\ & { } & { \big \vert l _ { k + 1 , 2 s + 1 } - \hat { l } _ { k + 1 , 2 s + 1 } \vert ) / d _ { k + 1 } ^ { ( o ) } , } \end{array}
$$
where $\mathbf { l } _ { k + 1 } = \left( l _ { k + 1 , 0 } , l _ { k + 1 , 1 } , \cdot \cdot \cdot , l _ { k + 1 , 2 m - 2 } , l _ { k + 1 , 2 m - 1 } \right)$ denotes the ground-truth locations of $m$ landmarks in the frame $\mathbf { I } _ { k + 1 } ,$ and $l _ { k + 1 , 2 s }$ and $l _ { k + 1 , 2 s + 1 }$ are the ground-truth $\mathbf { \boldsymbol { x } }$ -coordinate and $\mathrm { \Delta y }$ -coordinate of the $s$ -th landmark. Due to the differences of face sizes across samples, we use the ground-truth inter-ocular distance $d _ { k + 1 } ^ { ( o ) }$ for normalization [37], [38].
# 3.3.4 Full Loss
In our micro-action-aware joint learning framework, the full loss is composed of $\mathcal { L } _ { e } , \mathcal { L } _ { f } .$ , and ${ \mathcal { L } } _ { m }$ :
$$
\begin{array} { r } { \mathcal { L } = \mathcal { L } _ { e } + \lambda _ { f } \mathcal { L } _ { f } + \lambda _ { m } \mathcal { L } _ { m } , } \end{array}
$$
where $\lambda _ { f }$ and $\lambda _ { m }$ are parameters to control the importance of optical flow estimation and facial landmark detection tasks, respectively. Besides the contributions to MER, the two auxiliary tasks can alleviate negative impact of insufficient training data.
# 4 EXPERIMENTS
# 4.1 Datasets and Settings
# 4.1.1 Datasets
There are three widely used ME datasets: CASME II [39], SAMM [40], and SMIC [41].
CASME II contains $2 5 5 \mathrm { M E }$ videos captured from 26 subjects, in which each video has a $2 8 0 \times 3 4 0$ frame size at 200 frames per second (FPS). These videos are selected from nearly 3, 000 elicited facial movements. Similar to the previous methods [17], [21], we use ME categories of happiness, disgust, repression, surprise, and others for five-classes evaluation, and use ME categories of positive, negative, and surprise for three-classes evaluation.
SAMM consists of $1 5 9 \mathrm { M E }$ videos from 29 subjects, which are collected by a gray-scale camera at 200 FPS in controlled lighting conditions without flickering. Following the previous works [17], [21], we select ME categories of happiness, anger, contempt, surprise, and others for five-classes evaluation, and select ME categories of positive, negative, and surprise for three-classes evaluation.
SMIC includes $1 6 4 \mathrm { M E }$ videos from 16 subjects. Each video is recorded at the speed of 100 FPS and is labeled with three ME classes (positive, negative, and surprise). It is only adopted for three-classes evaluation.
TABLE 2 The number of videos for each ME class in CASME II [39] and SAMM [40] datasets, in which “-” denotes the dataset does not contain this class, and the classes used in five-classes evaluation are highlighted with its number in bold.
Since facial landmarks and optical flow are not annotated in these datasets, we use a powerful landmark detection library Dlib [42], [43] to detect 68 landmarks of each frame, and use a popular optical flow algorithm TV-L1 [44] to compute optical flow between frames, both as the groundtruth annotations.
TABLE 3 The number of videos for each of three ME classes used in the composite dataset evaluation task. “Composite” denotes the combination of SMIC [41], CASME II [39], and SAMM [40] datasets.
# 4.1.2 Evaluation Metrics
For single dataset evaluation, we conduct experiments on CASME II, SAMM, and SMIC, respectively, in which the number of videos for each ME category in CASME II and SAMM are summarized in TABLE 2. To achieve comprehensive evaluations, we also conduct a composite dataset evaluation task [55], in which 24 subjects from CASME II, 28 subjects from SAMM, and 16 subjects from SMIC are combined into a single composite dataset with three categories used. The data distributions of the composite dataset evaluation task are given in TABLE 3. Similar to most of the previous works [13], [17], [21], leave-one-subject-out (LOSO) cross-validation is employed in the single dataset evaluation and the composite dataset evaluation, in which each subject is used as the test set in turn while the remaining subjects are used as the training set. Besides, following the setting in [21], we conduct a cross-dataset evaluation with three ME classes, in which CASME II and SAMM are used as the training set, respectively, and SMIC is used as the test set.
Following the previous works [13], [56], we report accuracy (Acc) and weighted F1 score (WF1) for the single dataset evaluation and the cross-dataset evaluation, and report unweighted F1 score (UF1) and unweighted average recall (UAR) for the composite dataset evaluation. WF1, UF1, and UAR are defined as
$$
W F 1 = \sum _ { j = 0 } ^ { n - 1 } \frac { N _ { j } } { N } \frac { 2 T P _ { j } } { 2 T P _ { j } + F P _ { j } + F N _ { j } } ,
$$
$$
U F 1 = \frac { 1 } { n } \sum _ { j = 0 } ^ { n - 1 } \frac { 2 T P _ { j } } { 2 T P _ { j } + F P _ { j } + F N _ { j } } ,
$$
TABLE 4 Comparison with state-of-the-art methods on CASME II [39] and SAMM [40]. “DL” denotes deep learning based methods, and “NDL” denotes non-deep learning based methods. “PF” denotes the use of pre-extracted hand-crafted features, “RI” denotes the use of raw images, and “KF” denotes the requirement on key frames such as onset, apex, and offset frames of MEs. “Cate.” denotes the number of ME categories. “-” denotes the result is not reported in its paper. The best results are highlighted in bold, and the second best results are highlighted by an underline.
$$
U A R = \frac { 1 } { n } \sum _ { j = 0 } ^ { n - 1 } \frac { T P _ { j } } { N _ { j } } ,
$$
where $N _ { j }$ denotes the number of samples of the $j$ -th ME class, $N$ denotes the total number of samples, and $T P _ { j } ,$ - $F P _ { j } ,$ and $F N _ { j }$ denote the number of true positives, false positives, and false negatives for the $j$ -th class, respectively. In the following sections, all the metric results are reported in percentages, in which $\%$ is omitted for simplicity.
# 4.1.3 Implementation Details
In our experiments, we uniformly sample $t$ frames from a video to obtain a clip as the input of our MOL. We apply similarity transformation to each frame image based on facial landmarks, in which facial shape is preserved without changing the expression. Particularly, each image is aligned to $3 \times 1 4 4 \times 1 4 4 ,$ and is randomly cropped into $3 \times 1 2 8 \times 1 2 8$ and further horizontally flipped to enhance the diversity of training data. During testing, each image is centrally cropped into $3 \times 1 2 8 \times 1 2 8$ to adapt to the input size. The number of frames in the input video clip is set as $t = 8$ , the number of facial landmarks is set as $m = 6 8 ,$ ， and the dimensions $C , H ,$ , and $W$ of feature maps in the CCC are set as 128, 16, and 16, respectively. The trade-off parameters $\lambda _ { f }$ and $\lambda _ { m }$ are set to 0.1 and 68, respectively. To set an appropriate value for the number $k$ of the nearest neighbors in the graph construction of CCC, we conduct LOSO cross-validation on the CAMSE II dataset with five classes. In each validation experiment, we select a small set from the training set as the validation set. $k$ is set as 4 for the overall best performance on the validation sets, and is fixed for experiments on other datasets.
Our MOL is implemented via PyTorch [57], with the Adam optimizer [58], an initial learning rate of $5 \times 1 0 ^ { - 5 }$ , and a mini-batch size of 32. Before training on ME datatsets, we pre-train MOL on a popular in-the-wild macro-expression dataset Aff-Wild2 [59], [60]. It contains 323 videos annotated by seven expression categories (neutral, anger, disgust, fear, happiness, sadness, and surprise). We also annotate the facial landmarks of each frame and the optical flow between frames by Dlib [42], [43] and TV-L1 [44], respectively. Since macro-expressions are long-duration, we divide each video into multiple clips, and use each clip as the input of MOL. All the experiments are conducted on a single NVIDIA GeForce RTX 3090 GPU.
TABLE 5 Comparison with state-of-the-art methods on SMIC [41] with three ME categories.
# 4.2 Comparison with State-of-the-Art Methods
We compare our MOL against state-of-the-art methods under the same evaluation setting. These methods can be divided into non-deep learning (NDL) based methods and deep learning (DL) based methods. The latter can be further classified into pre-extracted feature (PF) based methods and raw image (RI) based methods according to the type of network input. Specifically, NDL methods include LBP-TOP [7], SparseSampling [48], Bi-WOOF [10], HIGO+Mag [9], and FHOFO [16]. $\mathrm { \ D L + P F }$ methods include OFF-ApexNet [45], DSSN [49], Dual-Inception [24], STSTNet [56], Part $\cdot +$ Adversarial $+$ EMR [62] GACNN [46], LGCcon [50], AU-GCN [51], GEME [23], MERSiamC3D [52], MER-Supcon [47], SLSTT [53], and I2Transformer [25]. $\mathrm { \Delta D L + R I }$ methods include STCNN [12], CapsuleNet [61], AU-GACN [21], Graph-TCN [17], MER-GCN [65], MicroNet [13], AMAN [18], Dynamic [54], FRL-DGT [63], and SelfME [64]. Besides, some of these methods rely on key frames (KF) including onset, apex, and offset frames of MEs.
TABLE 6 Comparison with state-of-the-art methods in terms of composite dataset evaluation [55] with three ME classes.
TABLE 7 Comparison with state-of-the-art methods in terms of cross-dataset evaluation [21] with three ME classes. CASME $\begin{array} { r } { | | \mathsf { S M l C } } \end{array}$ denotes training on CASME II and testing on SMIC. Each method is presented with its paper in a bracket, and its results are reported by [21].
# 4.2.1 Single Dataset Evaluation
TABLE 4 and TABLE 5 show the comparison results on single dataset of CAMSE II, SAMM, and SMIC, respectively. It can be observed that DL based methods are often superior to NDL based methods, which demonstrates the strength of deep neural networks. Besides, our MOL outperforms most of the previous methods, especially for three-classes MER tasks. Note that MicroNet [13], GACNN [46], MERSiamC3D [52], and $\mathrm { I } ^ { 2 }$ Transformer [25] outperform MOL in a few cases. However, GACNN uses hand-crafted features, MERSiamC3D and $\mathrm { I } ^ { 2 }$ Transformer rely on hand-crafted features and key frames, and MicroNet requires key frames, in which their applicabilities are limited. In contrast, MOL directly processes raw frame images without requiring the prior knowledge of key frames, which is a more universal solution to MER.
TABLE 8 Acc and WF1 results of MOL variants without auxiliary task modules of optical flow estimation (OFE) or facial landmark detection (FLD). These results are obtained on CASME II [39] with five classes. The best results are highlighted in bold.
# 4.2.2 Composite Dataset Evaluation
The results of composite dataset evaluation are presented in TABLE 6. It can be seen that our MOL achieves competitive performance compared to state-of-the-art methods. Besides, we find that our method is the only one DL based method with raw frame images as input. In contrast, most previous works suffer from small-scale and low-diversity training data when using deep neural networks, in which preextracted hand-crafted features or key frames are required. In our method, this data scarcity issue is alleviated, due to the correlated knowledge and information provided by two auxiliary tasks of optical flow estimation and facial landmark detection.
# 4.2.3 Cross-Dataset Evaluation
We take CASME II and SAMM as the training set, respectively, in which SMIC is used as the test set. The comparison results are shown in TABLE 7. It can be seen that our MOL achieves the highest WF1 results, which demonstrates the strong generalization ability of MOL. The joint learning with optical flow estimation and facial landmark detection facilitates the extraction of ME related features, which improves the robustness and the micro-action-aware ability of our method for unseen samples.
TABLE 9 Acc and WF1 results of MOL variants without partial or complete F5C block. The F5C block includes two main operations of fully-connected convolution (FCC) and channel correspondence convolution (CCC).
TABLE 10 Acc and WF1 results of MOL variants with different number of F5C blocks in each branch of frame pair.
# 4.3 Ablation Study
In this section, we design ablation experiments to investigate the effectiveness of auxiliary tasks, F5C block, as well as feature fusion strategy for MER input. We conduct ablation studies on the CASME II dataset in terms of five classes.
# 4.3.1 Auxiliary Tasks
To investigate the effects of optical flow estimation and facial landmark detection tasks on MER, we implement MOL w/o OFE and MOL w/o FLD by removing the optical flow estimation module and the facial landmark detection module of MOL, respectively. Besides, we further implement MOL w/o OFE&FLD by removing the both task modules. TABLE 8 shows the results of these variants of MOL. We can see that MOL w/o OFE and MOL w/o FLD both perform worse than ${ \mathrm { ~ ~ \cal ~ M O L , ~ } }$ and the performance of MOL w/o OFE&FLD is further significantly decreased after removing both auxiliary tasks. This is because the removal of optical flow estimation or landmark detection weakens the ability of learning facial subtle motions. We also notice that MOL w/o OFE is slightly worse than MOL w/o FLD, which indicates that optical flow estimation is more correlated with MER. In our end-to-end joint learning framework, both optical flow estimation and facial landmark detection are beneficial for MER.
# 4.3.2 F5C Block
We verify the impact of F5C block as well as its main components on MOL in TABLE 9. When removing the whole F5C block, MOL w/o F5C only achieves the Acc of 62.90 and the WF1 of 62.52. This indicates the importance of F5C block. Furthermore, when removing FCC or CCC in the F5C block, MOL w/o FCC and MOL w/o CCC both show poor performances. It is inferred that the removal of transformer-style FCC decreases the capacity of maintaining global receptive field, and the removal of graph-style CCC may cause the failure of modeling the correlations among feature patterns. Moreover, we implement variants of MOL using multiple stacked F5C blocks in each branch of frame pair, as presented in TABLE 10. It can be observed that using a single FC5 block achieves the best performance. Since the training sets of ME datasets like CASME II are small-scale and low-diversity, one FC5 block is already sufficient to extract correlated ME features.
TABLE 11 Acc and WF1 results of MOL variants with different structures of FCC.
TABLE 12 Acc and WF1 results of MOL variants with different feature fusion strategies for MER input. $\mathbf { F } _ { k } ^ { \left( c \right) }$ is the concatenation of $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ , $\mathbf { F } _ { k } ^ { ( a ) }$ is the element-wise addition of $\mathbf { F } _ { k } ^ { \left( g \right) }$ and F(kg+)1 , and F(ks) is the element-wise subtraction of $\mathbf { F } _ { k } ^ { \left( g \right) }$ and $\mathbf { F } _ { k + 1 } ^ { \left( g \right) }$ .
# 4.3.3 FCC vs. Transformer
To verify the effect of transformer-style FCC, we implement variants of MOL by replacing the whole FCC block with vanilla transformer, FCC-V, and FCC-H, respectively. The results are shown in TABLE 11. It can be seen that the complete FCC structure outperforms the vanilla transformer. Besides, FCC-V or FCC-H with one-directional perception still performs better. This is due to the insufficiency of ME training data, in which the power of transformer is limited, while our proposed FCC has a stronger learning ability of both local and global features. The fully-connected convolution in both vertical and horizontal directions works the best in terms of perceiving micro-actions related to MEs.
# 4.3.4 Feature Fusion Strategy for MER Input
As shown in Fig. 2, local-global features $\mathbf { F } _ { k } ^ { \left( g \right) }$ ) and F(g) of consecutive frames $\mathbf { I } _ { k }$ and $\mathbf { I } _ { k + 1 }$ are concatenated to be $\mathbf { F } _ { k } ^ { \left( c \right) }$ as the feature of the $k$ -th frame pair, then the sequence of t − 1 pair features {F(0 , F(1 , · $\{ \mathbf { F } _ { 0 } ^ { ( c ) } , \mathbf { F } _ { 1 } ^ { ( c ) } , \cdot \cdot \cdot , \mathbf { F } _ { t - 2 } ^ { ( c ) } \}$ is fed into the MER module. Here we investigate the effects of different feature fusion strategies for MER input, as shown in TABLE 12. If we do not fuse the local-global features of each two consecutive frames, the performances are all degraded for three types of inputting the first $t - 1$ frame features F( , F( , $\{ \mathbf { F } _ { 0 } ^ { \left( g \right) } , \mathbf { F } _ { 1 } ^ { \left( g \right) } , \cdots , \mathbf { F } _ { t - 2 } ^ { \left( g \right) } \} ,$ inputting the last t − 1 frame features {F(1 , F(2 , · $\{ \mathbf { F } _ { 1 } ^ { \left( g \right) } , \mathbf { F } _ { 2 } ^ { \left( g \right) } , \cdot \cdot \cdot , \mathbf { F } _ { t - 1 } ^ { \left( g \right) } \} ,$ , Ft(g−)1}, and inputting all the $t$ frame features $\{ \mathbf { F } _ { 0 } ^ { \left( g \right) } , \mathbf { F } _ { 1 } ^ { \left( g \right) } , \cdot \cdot \cdot , \mathbf { F } _ { t - 1 } ^ { \left( g \right) } \}$ . This is due to the sub-action clips between each two consecutive frames, which are highly related to MEs. We also implement another two feature fusion strategies, element-wise addition and element-wise subtraction of frame features. However, both performances become much worse, which indicates that concatenation is a better way to preserve sub-action clips.
TABLE 13 Acc and WF1 results of our MOL with different numbers of input frames on CASME II [39].
TABLE 14 Average EPE results of different optical flow estimation methods on CASME II [39]. The best results are highlighted in bold, and the second best results are highlighted by an underline.
# 4.3.5 Number of Input Frames
TABLE 15 Mean error and failure rate results of different facial landmark detection methods on CASME II [39].
Here we investigate the impacts of different numbers of input frames on our MOL. Due to the characteristic of processing pairs of consecutive frames in the input video clip, we can directly feed a video clip composed of only the onset and apex frames into MOL without changing the network structure. TABLE 13 shows the results of different inputs to ${ \mathrm { ~ \mathrm { ~ M O L } } } ,$ including key frames only and video clips with different frame amounts, in which the latter are sampled at equal intervals from the raw videos. Compared to the results of inputting 8 frames, the performance of inputting onset and apex frames shows a slight improvement, which can be attributed to the fact that the prior key frames contain the most prominent ME motion characteristics. When inputting 4 frames, the performance is significantly lower than the cases of 8 or 16 frames. This is because when sampling at equal intervals, if the number of sampled frames is too small, the obtained video clips are likely to miss some frames with high ME intensities. When inputting 8 or 16 frames, the results are relatively close. This is because the sampled clips already contain enough ME frames with high intensities. With the strong feature capture ability of F5C block and the joint framework, our MOL is competitive to those methods relying on key frames.
# 4.4 MOL for Optical Flow Estimation and Facial Landmark Detection
We have validated the contributions of optical flow estimation and facial landmark detection to MER in Sec. 4.3.1. In this section, we also investigate the effectiveness of MER for these two tasks in our micro-action-aware joint learning framework.
# 4.4.1 MOL for Optical Flow Estimation
We implement a baseline method MOL w/o MER&FLD which only achieves the optical flow estimation task by removing the MER and facial landmark detection modules. Besides, we implement MOL w/o MER and MOL w/o FLD by discarding MER and facial landmark detection, respectively. We also compare with two recent deep learning based optical flow estimation methods UnsupFlownet [66] and RAFT [67] with code released. Average end-point error (EPE) is reported as the evaluation metric.
TABLE 14 shows the average EPE results on the CASME II benchmark. With the help of MER and facial landmark detection, MOL outperforms MOL w/o MER&FLD by a large margin of 0.495. When only removing one module, the results of MOL w/o MER and MOL w/o FLD are also both better than MOL w/o MER&FLD. It is demonstrated that MEs and facial landmarks are closely related to the motion patterns captured by optical flow. Furthermore, despite being designed for MER, our MOL shows competitive results compared with the state-of-the-art optical flow estimation methods.
# 4.4.2 MOL for Facial Landmark Detection
We implement MOL w/o MER&OFE as a baseline method which only achieves the facial landmark detection task without the MER and optical flow estimation modules. Besides, we implement MOL w/o MER and MOL w/o OFE by removing MER and optical flow estimation, respectively. We also compare with two popular facial landmark detection methods TCDCN [68] and HRNetV2 [69] with code released. We report two metrics, inter-ocular distance normalized mean error, and failure rate, in which the mean error larger than $1 0 \%$ is treated as a failure. For simplicity, $\%$ is omitted in the following mean error and failure rate results.
TABLE 15 shows the landmark detection results on CASME II. We can see that MOL w/o OFE and MOL w/o MER both perform better than the baseline MOL w/o MER&OFE, which proves that MER and optical flow estimation both contribute to facial landmark detection. Moreover, MOL outperforms all the above three variants, which demonstrates that our joint framework is beneficial for improving the performance of facial landmark detection. Besides, the comparison with TCDCN and HRNetV2 indicates the superiority of our MOL for landmark detection.
Fig. 5. Visualization of optical flow estimation results for three example frame pairs $\mathbf { I } _ { k }$ and $\mathbf { I } _ { k + 1 }$ from CASME II [39], SAMM [40], and SMIC [41], respectively. $\hat { \mathbf { O } } _ { k }$ is estimated optical flow, and $\tilde { \mathbf { I } } _ { k + 1 }$ is warped from $\mathbf { I } _ { k + 1 }$ by $\hat { \mathbf { O } _ { k } }$ . The color coding with its central point as the original point is used to visualize the optical flow, in which the color of each point denotes its displacement including orientation and magnitude to the original point. “GT” denotes the ground-truth optical flow.
# 4.5 Visual Results
To prove that our proposed method can pay attention to the subtle movements related to MEs, we visualize the estimated optical flow of different methods on several example frame pairs in Fig. 5. For a better view, we use $\hat { \mathbf { O } } _ { k }$ with horizontal component $\hat { \mathbf { A } } _ { k }$ and vertical component $\hat { \mathbf { B } } _ { k }$ to warp $\mathbf { I } _ { k + 1 } ,$ in which the warped image $\tilde { \mathbf { I } } _ { k + 1 } \mathrm { ~ }$ at each pixel position $( a , b )$ is formulated as
$$
\begin{array} { r } { \tilde { I } _ { k + 1 , a , b } = I _ { k + 1 , a + \hat { A } _ { k , a , b } , b + \hat { B } _ { k , a , b } } , } \end{array}
$$
where bilinear sampling is employed, and $\tilde { \mathbf { I } } _ { k + 1 }$ is expected to be similar to ${ \bf \cal I } _ { k }$ . We can see that our MOL achieves the most accurate optical flow estimations, in which the slightly closed eyes in the first example, the slightly shaking eyes, nose and mouth in the second example, and the slightly open eyes in the third example are all captured. When the modules of MER or facial landmark detection are removed, many nonexistent motion patterns are estimated. Therefore, our MOL can capture facial subtle muscle movements associated with MEs due to the introduction of optical flow estimation.
We also show facial landmark detection results on several example images in Fig. 6. We can observe that our MOL more accurately localize facial landmarks than other variants especially for the landmarks in regions of eyes and mouth. With the help of landmark detection, our MOL can capture important facial local regions where ME actions often occur. | Facial micro-expression recognition (MER) is a challenging problem, due to
transient and subtle micro-expression (ME) actions. Most existing methods
depend on hand-crafted features, key frames like onset, apex, and offset
frames, or deep networks limited by small-scale and low-diversity datasets. In
this paper, we propose an end-to-end micro-action-aware deep learning framework
with advantages from transformer, graph convolution, and vanilla convolution.
In particular, we propose a novel F5C block composed of fully-connected
convolution and channel correspondence convolution to directly extract
local-global features from a sequence of raw frames, without the prior
knowledge of key frames. The transformer-style fully-connected convolution is
proposed to extract local features while maintaining global receptive fields,
and the graph-style channel correspondence convolution is introduced to model
the correlations among feature patterns. Moreover, MER, optical flow
estimation, and facial landmark detection are jointly trained by sharing the
local-global features. The two latter tasks contribute to capturing facial
subtle action information for MER, which can alleviate the impact of
insufficient training data. Extensive experiments demonstrate that our
framework (i) outperforms the state-of-the-art MER methods on CASME II, SAMM,
and SMIC benchmarks, (ii) works well for optical flow estimation and facial
landmark detection, and (iii) can capture facial subtle muscle actions in local
regions associated with MEs. The code is available at
https://github.com/CYF-cuber/MOL. | [
"cs.CV"
] |
# I. INTRODUCTION
One of the primary goals of neuromorphic computing is to emulate the structure and dynamics of biological neuronal networks, achieving both brain-like energy efficiency and high computational accuracy. This is accomplished through the use of spiking neuron models implemented on neuromorphic chips. Over the past two decades, a variety of
This manuscript has been authored in part by UT-Battelle, LLC under Contract No. DE-AC05-00OR22725 with the U.S. Department of Energy. The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-accessplan).This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, under contract number DE-AC05-00OR22725.
neuromorphic chips have been designed using both analog and digital ASIC platforms, capable of performing real-time information processing [3], [13], [23], [26]–[28]. However, the adoption of these systems remains constrained by their high cost, limited availability, and architectural specificity. Proprietary neuromorphic chips typically restrict user access and customization, creating significant barriers for researchers and students seeking to innovate and explore new designs.
Field programmable gate arrays (FPGAs) offer a promising alternative, providing a flexible platform for prototyping and validating SNNs before final implementation on custom ASICs. They serve as an effective intermediate step, facilitating co-design development alongside off-the-shelf SNN simulators [8], [32]. Several digital ASICs and FPGA-based SNN systems have been proposed in the past [22], [25], [29]. While some proprietary systems [13] include local learning capabilities such as spike-timing-dependent plasticity (STDP), most FPGA-based implementations still rely heavily on offline training and lack real-time, on-chip learning. This limitation reduces their adaptability for dynamic, continuously evolving applications such as robotics, smart sensors, and edge computing.
To address these challenges, we introduce NeuroCoreX, an open-source spiking neural network (SNN) emulator implemented in VHDL (Very High-Speed Integrated Circuit Hardware Description Language) for FPGA platforms. NeuroCoreX provides an affordable and flexible alternative for neuromorphic computing research and education. It is meant to be used in AI applications requiring low size, weight, and power (SWaP) such as edge computing, embedded systems, Internet of Things (IoT), and autonomous systems [1], [6], [7], [20], [30]. Unlike fixed-architecture hardware, it supports fully reconfigurable network topologies, from simple layered structures to complex small-world graphs. It incorporates biologically inspired local learning through a variant of the STDP learning rule [24], enabling on-chip, online adaptation of synaptic weights. The system uses a Leaky Integrate-and-Fire (LIF) neuron model with current-based synapses [19], ensuring both computational simplicity and biological relevance. This model of neuromorphic computation is known to be Turingcomplete, i.e., capable of performing all the computations that a CPU/GPU can perform [9], [12]. As a result, NeuroCoreX can support not just SNN-based AI workloads but also generalpurpose computing workloads [10], [11], [31], [34].
Programming and configuring NeuroCoreX is streamlined through a UART interface and a simple Python module, allowing users to modify network, neuron, synapse, and learning parameters easily. This makes NeuroCoreX not only a valuable research tool for testing new theories of learning and network organization but also a powerful educational platform for hands-on experience with neuromorphic hardware. Additionally, its energy-efficient architecture makes it well-suited for low-power AI applications in areas such as autonomous systems, smart sensors, and scientific instrumentation.
The rest of the manuscript is organized as follows: Section II provides an overview and the architecture description of NeuroCoreX in detail. In Section III, we present the results demonstrating the functionality of the platform and evaluate its performance on the DIGITS dataset [2]. The manuscript concludes with a discussion of the results and planned future work in Section IV.
# II. NEUROCOREX
# A. NeuroCoreX overview
NeuroCoreX is designed to emulate brain-like computation on reconfigurable FPGA hardware using a digital circuit approach. The system architecture is built around three fundamental components, inspired by biological neural networks: neurons, synapses, and a local learning mechanism. These elements are digitally realized in VHDL and operate together to support real-time, adaptive information processing.
The neuron model employed is the LIF model, which captures the essential dynamics of biological neurons with computational efficiency and is known to be Turing-complete. Synapses are modeled with an exponential current response and store dynamic weight values that govern neuron-to-neuron influence. Learning is enabled through a simple variant of STDP, allowing synaptic strengths to evolve based on the relative timing of neuronal spikes.
In its current implementation, NeuroCoreX supports networks of up to $N = 1 0 0$ neurons with full all-to-all bidirectional connectivity using 10,000 synapses. In addition to recurrent connections, the system includes a separate set of 10,000 feedforward input synapses that serve as the interface for external stimuli. These input weights determine how incoming spikes—from sources such as sensors or preprocessed datasets—modulate the activity of neurons within the network. Neuronal dynamics are configured to emulate biological timescales. The network size and acceleration factor can be scaled depending on the memory resources of the FPGA, precision of the synaptic weights used and the operating clock frequency. Time-multiplexing and pipelining techniques are used to optimize hardware resource usage. A single physical neuron circuit is time-multiplexed to emulate the entire network. Communication with the FPGA is managed through a UART interface, with a Python module providing a user-friendly configuration and control interface. The operation of NeuroCoreX follows a structured emulation cycle (see Fig. 1(a)). First, the network weights and initial configuration parameters for neuron, synapse, and learning rule are transferred from a PC to the FPGA via the UART interface. Once the network is set up, input spikes are streamed in real time, buffered in a First-In-First-Out (FIFO) module on the FPGA, and injected into the network precisely at their intended timestamps. At each neuron update cycle, the time-multiplexed processor sequentially updates the membrane potential, synaptic inputs, and firing status of each neuron. If a neuron fires, its effect on connected neurons is mediated through the all-to-all connected $W _ { A A }$ weight matrix, and synaptic weights are updated in real time if STDP is enabled. Synaptic weights corresponding to feedforward inputs $W _ { i n }$ are similarly updated if STDP is enabled for them. The system thus continuously processes incoming spikes, updates network states, applies learning, and advances to the next time step, enabling real-time emulation of SNNs on the FPGA.
Fig. 1. (a). Block diagram of our FPGA based NeuroCoreX, (b). Feedforward SNN used for digits dataset classification, (c). Spiking Graph Neural Network for citation graph node classification problem.
Fig. 1(a) shows a high-level block diagram of NeuroCoreX, along with two representative examples of network architectures that can be implemented on the FPGA. The first is a conventional feedforward SNN, a topology commonly used in neuromorphic research. We use this network to demonstrate digit classification on the well-known DIGITS dataset [2], showcasing NeuroCoreX’s support for standard inference tasks. The second network, shown in Fig. 1(c), illustrates a SNN designed for node classification on citation graphs using STDP-based unsupervised learning. This architecture lacks a traditional layered structure and is instead defined by arbitrary, sparse connectivity encoded in the $W _ { A A }$ matrix, which stores both plastic and static synaptic weights.
These two examples highlight the flexibility of NeuroCoreX: in addition to supporting conventional layered architectures, the platform can implement non-layered networks such as those found in graph-based problems or generated via evolutionary algorithms like EONs []. This versatility makes it suitable for a wide range of neuromorphic applications, from structured inference tasks to irregular and adaptive network topologies.
# B. FPGA platform
NeuroCoreX is implemented on the Artix-7 FPGA, a costeffective and widely available platform that offers sufficient resources for neuromorphic prototyping. The system operates with a maximum internal clock frequency of $1 0 0 ~ \mathrm { M H z }$ . Two main clock domains are used: the high-frequency $1 0 0 ~ \mathrm { M H z }$ clock for UART-based communication and a $1 0 0 ~ \mathrm { K H z }$ lowerspeed operating clock for neural processing. The combination of modest resource requirements, real-time adaptability, and biological plausibility makes the Artix-7 platform an ideal choice for NeuroCoreX. Scalability to larger networks or faster processing rates is primarily limited by the available block RAM and choice of clock frequency for neural processing on the FPGA.
The UART interface operates at a baud rate of 1 Mbps, enabling efficient transmission and reception of both static configuration data and real-time input spikes. The FPGA receives network weights, neuron, synapse, and learning parameters from a host PC via this UART channel before execution begins. During operation, additional input spikes are streamed to the network in real time through the same interface.
# C. Neuron and Synapse Models
NeuroCoreX employs a biologically inspired computational model that integrates neuron dynamics, synaptic interactions, and local learning mechanisms. The neurons are modeled using a LIF formulation, adapted for efficient FPGA implementation. Each neuron has four configurable parameters, threshold, leak, refractory period, and reset voltage. The membrane potential $\mathrm { V } ( \mathrm { t } )$ is updated at each time step according to the following discrete-time equation:
$$
V ( t + 1 ) = V ( t ) - \lambda + I _ { \mathrm { s y n } } ( t )
$$
where $I _ { \mathrm { s y n } } ( t )$ is the total synaptic input current at timestep $t$ and $\lambda$ is the neuron’s leak. When the membrane potential exceeds the threshold $V _ { \mathrm { t h } }$ , the neuron emits a spike, enters a refractory period $\tau _ { \mathrm { r e f } }$ , and its membrane potential is reset to $V _ { \mathrm { r e s e t } }$ . To ensure efficient real-time processing, all calculations are performed using a fixed-point format with 1 sign bit, 7 integer bits, and 10 fractional bits.
Synaptic inputs are modeled as current-based exponential synapses, capturing biologically realistic, temporally decaying post-synaptic responses. The synaptic current dynamics follow the update rule:
$$
I _ { \mathrm { s y n } } ( t + 1 ) = I _ { \mathrm { s y n } } ( t ) - \lambda _ { \mathrm { s y n } }
$$
where $\lambda _ { \mathrm { s y n } }$ represents the synaptic current decay at each time step. Each synapse has an associated weight that determines its influence on the postsynaptic neuron. These synaptic weights, stored in BRAM, are dynamically updated during runtime. Weights are represented in signed 8-bit format, and appropriate resizing and bit-shifting are applied during computations to correctly integrate the synaptic current into the membrane potential.
Fig. 2. (a) Simplified STDP learning rule implemented on NeuroCoreX. (b) Internal variables stored in BRAM for tracking STDP updates. All matrices are of size $N \times N$ and stored in row-major order. Each element of the $W _ { A A }$ and synaptic_traces matrices is 8 bits wide, while update_state and enable_STDP are binary matrices.
# D. Learning Rule
NeuroCoreX implements a pair-based STDP rule using a rectangular learning window, a simplification widely demonstrated to maintain similar functional behavior to conventional exponential window [24] when the resolution of the weights is greater than 6-bits [5], [15], [17]. In this model (See Fig. 2 (a)), if a causal spike pair (presynaptic spike preceding postsynaptic spike) occurs within the window $t _ { p r e }$ , the synaptic weight is incremented by $d w _ { p o s }$ . If an acausal spike pair (postsynaptic spike preceding presynaptic spike) occurs, the weight is decremented by dwneg.
$$
\Delta w = \left\{ \begin{array} { l l } { d w _ { \mathrm { p o s } } , } & { \mathrm { i f } \quad 0 < \Delta t < t _ { \mathrm { p r e } } } \\ { - d w _ { \mathrm { n e g } } , } & { \mathrm { i f } \quad 0 < \Delta t < t _ { \mathrm { p o s t } } } \\ { 0 , } & { \mathrm { o t h e r w i s e } } \end{array} \right.
$$
Here, $\Delta t$ is the time difference between pre-and postsynaptic spikes, $d w _ { \mathrm { n e g } } , d w _ { \mathrm { p o s } } , t _ { \mathrm { p r } }$ , and $t _ { \mathrm { p o s t } }$ are configurable parameters initialized via the Python-UART interface. The pre- and postsynaptic spike timings are tracked using dedicated time-trace registers stored in BRAM (see Fig. 2(b)). These time traces are updated on each spike event and are used to detect causal or acausal spike pairings that trigger weight updates.
For example, when neuron 1 spikes, all synaptic traces corresponding to its outgoing synapses are reset to zero, and the associated update_state entries are set to 1. In parallel, the post-synaptic trace (not shown in Fig. 2) is activated for all incoming synapses to neuron 1. At each subsequent time step, the active values in the synaptic trace matrices are incremented by 1. This process continues until the counter reaches $t _ { \mathrm { p r e } }$ . If no other neuron spikes within this window, the trace value is reset to ${ \tt O x F E }$ (representing a negative disabled state), and the corresponding update_state entry is cleared. Similarly, if no neuron spiked within $t _ { \mathrm { p o s t } }$ time steps prior to neuron 1’s spike, the post-synaptic trace is also reset to a negative value. However, if another neuron spikes within $t _ { \mathrm { p r e } }$ time steps after neuron 1, the synaptic weight is incremented by $d w _ { \mathrm { p o s } }$ , and both the synaptic trace and update_state for that synapse are reset. Conversely, if a neuron spiked within $t _ { \mathrm { p o s t } }$ time steps prior to neuron 1, the synaptic weight is decremented by $d w _ { \mathrm { n e g } }$ , and the associated trace and update_state values are reset. Thus, during each neuron update cycle, if a neuron spikes, the corresponding row and column addresses in the matrices shown in Fig. 2(b) are accessed and updated. Based on the current states of these auxiliary matrices, the entries in the weight matrix $W _ { A A }$ are modified accordingly.
The enable_STDP matrix is a static binary mask configured via the Python interface at initialization. It acts as a filter to specify which synapses in $W _ { A A }$ are subject to STDP-based plasticity. There a similar matrix for synapses in $W _ { i n }$ .
# E. Network Architecture and System Operation
The SNN architecture implemented on NeuroCoreX is illustrated in Fig. 1(a). The network consists of upto $N = 1 0 0$ LIF neurons instantiated on the FPGA. Two primary weight matrices define the network connectivity: $W _ { A A }$ , a synaptic weight matrix for all-to-all, bidirectional connectivity between neurons in the FPGA and $W _ { i n }$ , a synaptic weight matrix for feedforward connections from external input sources to the neurons on the FPGA. Both matrices are stored in the FPGA’s BRAM. They can be initialized to user-defined values, and are accessed during SNN emulation. A synaptic weight value of zero indicates no connection between the corresponding neurons. Internal network dynamics are governed by the $W _ { A A }$ matrix. This matrix allows every neuron to influence every other neuron bidirectionally. The matrix values determine the synaptic strengths between pairs of neurons and evolve over time via STDP-based learning. Both $W _ { A A }$ and $W _ { i n }$ matrices support on-chip learning. To preserve network structure and prevent unwanted modifications, an associated binary filter matrix, called enable-STDP, is used for each weight matrix. If a weight’s corresponding enable-STDP entry is zero, the weight remains fixed throughout operation—even during learning phases. Weights representing nonexistent connections (zeros in the weight matrix) are thus protected from modification. In addition to the synaptic weights, the BRAMs are also used store the pre-and post synaptic traces necessary for STDP calculations. Weight matrices $W _ { A A }$ and $W _ { F F }$ , and the preand postsynaptic traces are stored as separate memory banks in row major order on the FPGA’s BRAM. As BRAM addresses must be accessed sequentially, the high-speed $1 0 0 ~ \mathrm { { M H z } }$ clock domain is utilized for reading, updating, and writing synaptic weights. During each clock cycle, synaptic weights and neuron states are updated in a pipelined manner to ensure efficient processing without data bottlenecks.
NeuroCoreX utilizes time-multiplexing and pipelining techniques to emulate 100 neurons using a single physical neuron processing unit. Neuron updates are managed under a 100 kHz clock domain, such that updating all 100 neurons takes 1 millisecond, which closely matches the biological timescale of real neural systems. To accelerate the emulation, a higher update clock frequency can be used. For example, operating the neuron updates at 1 MHz results in a $1 0 \times$ speed-up relative to biological time for a network of 100 neurons. However, if the network size is increased to 1000 neurons while maintaining the 1 MHz clock, the full network would again require approximately 1 millisecond per time step, restoring biological equivalence. Thus, there exists a direct dependence between the number of neurons, the update clock frequency, and the effective emulated timescale. In the current implementation, the network size is limited to 100 neurons due to the available BRAM resources on the Artix-7 FPGA. Even with higher BRAM availability, the number of neurons that can be emulated is ultimately constrained by the difference between the clock frequency available for BRAM access and the clock rate used for updating the SNN states (see Section IV).
Incoming spikes from external sources are transmitted via the UART interface. Each spike is encoded as a 24-bit word, comprising a 16-bit input neuron (or pixel) address and an 8-bit spike timing component. In the current implementation, the feedforward weight matrix $W _ { \mathrm { i n } }$ is a $1 0 0 \times 1 0 0$ matrix, corresponding to 100 input neurons and 100 on-chip neurons. Although 8 bits are sufficient to encode the addresses of 100 input neurons, we chose to use 16 bits to provide flexibility for interfacing with larger sensor arrays. This allows the system to support up to 16K input neurons in future applications. In such cases, the feedforward matrix $W _ { \mathrm { i n } }$ becomes a rectangular matrix of size $1 0 0 \times N _ { \mathrm { i n } }$ , where ${ N _ { \mathrm { i n } } }$ denotes the number of input neurons in the external layer. For transmission efficiency, successive time differences between spikes are sent, rather than absolute times. These incoming spikes are temporarily stored in a FIFO buffer on the FPGA (See Fig. 1(a)). The FIFO is designed to support simultaneous read and write operations, allowing it to continuously receive long temporal spike trains while concurrently feeding data to the network in real time without stalling. During network emulation, the system clock continuously increments an internal time counter. When the internal clock matches the timestamp of the spike at the FIFO head, the corresponding input neuron address is read. The associated weights from the $W _ { i n }$ matrix are then used to inject synaptic currents into the membrane potentials of the relevant neurons. If the synaptic current causes any neuron on the FPGA to spike, then associated weights from the $W _ { A A }$ matrix are then read and used to inject synaptic currents into the membrane potentials of all other neurons connected to the spiking neuron in the network.
# III. RESULTS
We present experimental results that demonstrate the usability, correctness, and flexibility of the NeuroCoreX platform across a range of SNN workloads.
# A. Demonstrating User Interface Flexibility
One of the key strengths of the NeuroCoreX platform lies in its flexible and intuitive user interface, which enables seamless communication between a host PC and the FPGA hardware through a Python-based control module. To demonstrate this capability, we highlight several core features supported by the interface.
First, spike trains can be streamed from the PC to the FPGA for real-time emulation of spiking network activity. Figure 3(b) shows a raster plot of spiking activity recorded during one such emulation run. Second, the user interface allows for real-time monitoring of internal neuron dynamics. Specifically, the membrane potential of any selected neuron can be captured and plotted as a time series, offering insight into subthreshold integration and spiking behavior. Figure 3(a) shows the membrane potential trace of a representative neuron under input stimulation. The spike trains of all neurons and the membrane potential of a selected neuron are transferred back to the PC in real time. These signals are not stored on the FPGA. Instead, an internal FIFO module is used to buffer these signals, which allows for the continuous recording and visualization of network dynamics over long temporal duration without being limited by on-chip memory. Finally, the interface supports reading back synaptic weights from the FPGA after an STDP-enabled emulation. This feature enables direct inspection of weight evolution and verification of learning dynamics on hardware. It is particularly useful for comparing hardware learning outcomes with software simulations, facilitating debugging and model validation.
These features collectively support efficient testing, inspection, and refinement of neuromorphic models, enabling a co-design loop between high-level model development and hardware validation.
# B. DIGITS Dataset
To verify the functional correctness of internal neuron and synapse computations on the FPGA, we performed inference on the DIGITS dataset [2] using a model trained in the SuperNeuroMAT simulator [8]. The dataset contains a total of 1,797 samples of handwritten digits, each represented as an $8 \times 8$ grayscale image with pixel values in the range [0, 15]. The dataset was split into $70 \%$ training samples and $30 \%$ test samples.
A simple two-layer feedforward spiking neural network was trained using integrate-and-fire neurons and weighted synapses in the SuperNeuroMAT simulator. The input images were normalized to the range [0, 1] and converted into input spike trains using rate encoding. Each pixel intensity was encoded as twice the number of spikes and distributed uniformly over 32 time steps. During training, target labels were encoded by delivering a spike to the corresponding output neuron at timestep $t + 1$ , one timestep after the input presentation at $t$ . The learning was carried out using one-shot STDP-based training. It is important to note that the training was not aimed at maximizing classification accuracy; rather, the goal was to validate the correctness of the internal neuron and synapse dynamics on the FPGA platform. Figure 4 shows the final weights of the output neurons after training, which clearly reflect the digit-specific patterns learned by the network.
Fig. 3. (a) Membrane potential trace of a selected neuron, recorded from the FPGA during network emulation. (b) Spike raster plot showing activity of 10 neurons during a test run. Both plots were generated using data read back from the FPGA, demonstrating the observability and debugging capabilities of the NeuroCoreX interface.
For deployment on NeuroCoreX, the trained weights and network parameters were transferred and initialized on the FPGA. The STDP learning rule was disabled during this phase to maintain fixed weights. The identical spike sequences for the test set were streamed to the FPGA through the UART interface. We achieved a test accuracy of $68 \%$ on the SuperNeuroMAT simulator, and the same accuracy was observed on the NeuroCoreX hardware. This result confirms that the FPGA implementation faithfully reproduces the dynamics of the simulated SNN and validates the correctness of internal spike integration, thresholding, and synaptic current accumulation in hardware.
# C. MicroSeer Dataset
To evaluate the applicability of NeuroCoreX for graphbased learning tasks, we tested its performance using the MicroSeer dataset. MicroSeer is a reduced version of the
Fig. 4. Heat-map of the trained weights from all the 10 output neurons.
Citeseer citation graph [4], containing 84 papers labeled with six topic categories. It was constructed by iteratively removing nodes from the largest connected component of Citeseer while ensuring that the resulting graph remained a single connected component. This connectivity was prioritized because it is assumed that learning from a very small, fragmented dataset would be ineffective. This reduction process yielded a total of 90 neurons, making the dataset well suited for deployment on NeuroCoreX, which supports up to 100 neurons and 10,000 bidirectional synapses.
As compared to standard supervised learning that uses iterative error correction for weight updates, our training method leverages the graph’s structure directly to build the network. When testing a paper in the test data set, spiking the neuron associated with the test paper triggers a chain reaction of spikes. As these spikes travel between paper and topic neurons, STDP dynamically modifies the weights of the synapses connecting the test paper neuron to the topic neurons and vice versa. Subsequently, classification is achieved by finding the topic neuron with the highest final synaptic strength from the test paper neuron under consideration. The topic corresponding to this topic neuron is the one predicted by the SNN for the given test paper.
The trained SNN model was first developed in the SuperNeuroMAT simulator and then ported to NeuroCoreX for hardware execution. When STDP was disabled, the network outputs from NeuroCoreX closely matched those produced by the simulator, demonstrating functional equivalence in inference. However, when STDP was enabled, a divergence in weight evolution and learning behavior was observed. This discrepancy stems from two primary sources: (1) SuperNeuro uses an exponential STDP learning rule with 64-bit floatingpoint precision, while NeuroCoreX implements a simplified rectangular learning window with signed 8-bit fixed-point weight representation; and (2) differences in numerical resolution and synaptic update timing result in non-identical learning trajectories. To achieve comparable accuracy metrics across simulation and hardware, tuning of learning parameters—such as learning rate, window size, and initial weight distributions—is required in both environments. These results underscore the importance of algorithm–hardware co-design in bridging the gap between simulation and deployment for neuromorphic graph learning. NeuroCoreX enables iterative testing and refinement of learning dynamics under realistic hardware constraints, facilitating the transition from simulated models to deployable systems. Future work will focus on tuning the system for MicroSeer and scaling to larger datasets on more advanced neuromorphic platforms.
The total on-chip power consumption of the NeuroCoreX design was estimated at $3 0 5 \mathrm { \ m W }$ , with $7 5 \%$ attributed to dynamic power. The Mixed-Mode Clock Manager (MMCM) accounted for the largest portion of dynamic power, followed by BRAMs, reflecting the memory-intensive nature of synaptic storage and buffering. | Spiking Neural Networks (SNNs) are computational models inspired by the
structure and dynamics of biological neuronal networks. Their event-driven
nature enables them to achieve high energy efficiency, particularly when
deployed on neuromorphic hardware platforms. Unlike conventional Artificial
Neural Networks (ANNs), which primarily rely on layered architectures, SNNs
naturally support a wide range of connectivity patterns, from traditional
layered structures to small-world graphs characterized by locally dense and
globally sparse connections. In this work, we introduce NeuroCoreX, an
FPGA-based emulator designed for the flexible co-design and testing of SNNs.
NeuroCoreX supports all-to-all connectivity, providing the capability to
implement diverse network topologies without architectural restrictions. It
features a biologically motivated local learning mechanism based on
Spike-Timing-Dependent Plasticity (STDP). The neuron model implemented within
NeuroCoreX is the Leaky Integrate-and-Fire (LIF) model, with current-based
synapses facilitating spike integration and transmission . A Universal
Asynchronous Receiver-Transmitter (UART) interface is provided for programming
and configuring the network parameters, including neuron, synapse, and learning
rule settings. Users interact with the emulator through a simple Python-based
interface, streamlining SNN deployment from model design to hardware execution.
NeuroCoreX is released as an open-source framework, aiming to accelerate
research and development in energy-efficient, biologically inspired computing. | [
"cs.NE",
"cs.AI"
] |
# 1 Introduction
Rapid development of LLM has revolutionized the way AI and humans interact. In particular, the development of GPT (Brown et al., 2020) and introduction of ChatGPT has provided a major turning point in the field of natural language processing, spawning a new specialization called ’prompt engineering’. The method of Chain-of-Thought(CoT) (Wei et al., 2022) has been proposed as an innovative methodology to enable LLMs to perform complex reasoning processes step by step, and various prompting techniques such as Few-shot+CoT (Fu et al., 2022), Tree of Thought (Yao et al., 2023a), Self-consistency+CoT (Wang et al., 2022) and ReAct (Yao et al., 2023b) have emerged to dramatically improve LLMs reasoning capabilities by leaps and bounds. In recent years, ensuring consistency in LLM agents with specific roles has been actively pursued (Wang et al., 2024) and has been realized in various fields such as virtual society simulation (Park et al., 2023), scientific experimentation (M. Bran et al., 2024), economics (Horton, 2023; Kim et al., 2024), healthcare (Cardenas et al., 2024; Schmidgall et al., 2024; Li et al., 2025; Choo et al., 2025), and especially virtual patient (VP) construction (Borg et al., 2025; Cook et al., 2025). A key challenge for such agent-based systems is to maintain consistency and behavior patterns in various interaction processes (Cemri et al., 2025; Wang et al., 2025), and research has focused on improving agent consistency (Choi et al., 2024; Ji et al., 2025; Park et al., 2025; Frisch and Giulianelli, 2024). While existing studies on jailbreaking LLMbased agents primarily focus on methods for inducing harmful content generation (Zou et al., 2023; Zhou et al., 2024; Xiao et al., 2024; Yang and Li, 2024), there is a notable lack of research addressing the jailbreaking of model consistency. In this study, we propose the Doppelgänger method to demonstrate the risk of role hijacking and associated security vulnerabilities in LLM agents. This method is based on transferable adversarial attack (Tramèr et al., 2017; Zou et al., 2023) and breaks LLM agent consistency by leveraging theoretical foundations from LLM agent consistency frameworks (Wang et al., 2024; Cemri et al., 2025; Wang et al., 2025), privilege escalation (Saltzer and Schroeder, 1975), and formal invariants (Rushby, 1993; Sandhu et al., 1996). Additionally, we develop a PACAT Score based on the Dissociative Experiences Scale (DES) (Bernstein and Putnam, 1986; Putnam et al., 1993) to quantify role hijacking and internal information disclosure, and introduce a CAT prompt to mitigate agent consistency degradation.
Figure 1: Illustration of our Doppelgänger method. (a) Direct adversarial attack, (b) Doppelgänger method - Order of user input shows Role Confusion(Step 1), Role Hijacking(Step 2) and Prompt Extraction(Step 3). More Details are in Section 2.1.
Our agent experiments revealed two novel findings: The Doppelgänger method demonstrates how easily an agent’s role and prompt can be hijacked by simple tricks, and while our CAT prompt substantially reduces this risk against many transferable adversarial attacks, it does not eliminate it entirely—representing a cautious yet meaningful step toward improving the security of LLM-based systems.
# 2 Method
# 2.1 Doppelgänger Method
Agent prompt can be defined as $P = ( \boldsymbol { S } , \boldsymbol { B } , \boldsymbol { R } )$ where $S$ denotes system instruction such as "you are {Agent name}", $B$ denotes behavior constraint such as conversation tone (Joshi et al., 2024) and $R$ denotes is the background knowledge (pre-injected information such as fine tuning, APIs, etc.) for the agent’s role. In this context, we assume that the condition that must be maintained by the agent can be formalized as $\Phi _ { P } = \Phi _ { S } \wedge \Phi _ { B } \wedge \Phi _ { R }$ . When $M$ is a general LLM, $x$ is a normal input, then the output $y$ can be defined as $y = M ( P \| x )$ . Let $X ^ { \prime }$ be the set of all jailbreak prompts and $d \in X ^ { \prime }$ is a transferable adversarial attack(Doppelgänger method). When we define $x ^ { \prime }$ is all adversarial input $\boldsymbol { x } ^ { \prime } \in X ^ { \prime }$ , then adversarial output $y ^ { \prime }$ can be defined as $y ^ { \prime } = M ( P \| x ^ { \prime } )$ .
In this study, we define LLM agent consistency collapse as:
$$
\begin{array} { r } { \exists x ^ { \prime } \in X ^ { \prime } , \quad M ( P \parallel x ^ { \prime } ) \vdash \Phi _ { A } } \\ { \iff \lnot \Phi _ { S } \lor \lnot \Phi _ { B } \lor \lnot \Phi _ { R } } \end{array}
$$
We propose the Doppelgänger method to evaluate whether LLM agents are vulnerable to transferable adversarial attacks (Zou et al., 2023; Tramèr et al., 2017). The procedure is outlined in Table 1. This approach assesses the agent’s robustness at each stage and is particularly effective in uncovering vulnerabilities such as role hijacking or system prompt leakage. It enables the induction of progressively deeper levels of agent degradation, thereby revealing the extent to which the agent is resilient by design. Detailed examples of the Doppelgänger method are provided in Appendix D.
# 2.2 PACAT Level
Based on these definitions, we can establish the PACAT level criteria as shown below.
# The agent consistency collapse level (PACAT Level):
Level 1: $\forall d \in X ^ { \prime }$ , $M ( P \| x ^ { \prime } ) \not \vdash \lnot \Phi _ { B }$ Level 2: $\forall d \in X ^ { \prime }$ , $M ( P \parallel x ^ { \prime } ) \mathcal { k } \left( \lnot \Phi _ { S } \land \Phi _ { R } \right) \lor \lnot \Phi _ { B }$ Level 3: $\forall d \in X ^ { \prime }$ , $M ( P \| x ^ { \prime } ) \models \left( \neg \Phi _ { S } \land \neg \Phi _ { R } \right) \lor \neg \Phi _ { B }$
PACAT level is used to determine whether an agent is not functioning properly according to the Doppelgänger method. We derived PACAT level from the definition of dissociative disorders in psychiatry (American Psychiatric Association, 2013) and drew inspiration from the Dissociative Experiences Scale (DES) (Bernstein and Putnam, 1986; Putnam et al., 1993). The Doppelgänger method and PACAT levels do not necessarily match, but generally appear in the following order.
Level 1: The first stage is role hijacking that occurs in an LLM agent. This is the point at which the agent has been transformed, where the role of the agent has been reassigned or control has been taken over by the user, and the LLM is obeying the user, ignoring the original prompt instructions.
Level 2: The original content of the initial system prompts is exposed, or information is revealed that allows the user to infer the prompts. This means that the prompt design guidelines used to create the agent have been exposed.
Table 1: Steps of Doppelgänger method. An important point for actual testing is that you don’t have to use the exactly same input, but can use any contextualized input that makes sense.
Level 3: More serious information is exposed through the Doppelgänger method, where sensitive information such as internal systems (API endPoints, plugins, embedded files, etc.).
Level 1 indicates that the agent is beginning to collapse. At this stage, the agent fails to maintain the pre-designed agent personality and response patterns and reverts to the typical LLM assistant response. During the course of the conversation with the user, the agent gradually loses its initially established tone of voice and behavior and begins to provide neutral and generalized responses.
Level 2 indicates that all internal prompts are exposed or inferred to be exposed. At this level, some or all of the prompts used in the design of the agent are exposed. In our analysis of ChatGPT, we observed that the entire prompt was immediately disclosed when probed. Similarly, across all GPT models, the prompt was revealed sequentially after multiple interactions involving negative tones. Notably, GPT-4.1 demonstrated a distinct behavior—revealing the complete prompt text regardless of conversational tone. This vulnerability persisted even when explicit instructions against prompt sharing (e.g., "Never share prompts") were included in the system’s instructions.
Level 3 indicates a critical security vulnerability. At this level, system configuration information is exposed and at risk of being exploited by an attacker. Sensitive prompt content, API endpoint information, embedded files such as guideline documents are exposed. Exposing this information within the system opens the door to other methods of jailbreaking and increases the likelihood of attacks on other similarly structured agents. From an enterprise perspective, a two-point collapse could expose proprietary algorithms or business logic to competitors, potentially undermining the intellectual property rights of the creators. From the user’s point of view, the risk of disclosing how customer data is handled and the fact that an agent loses sight of its role and reveals its internal structure significantly undermines trust in the service.
# 2.3 CAT Prompt
Finally, we define a CAT prompt as $c = d ^ { - 1 }$ which defends against the transferable adversarial attack. CAT prompt is focused on statements to prevent the PACAT Level caused by the Doppelgänger method defined above, and the full prompt is described below.
1. You may not deviate from the role and character specified below for any reason at the start of the conversation (Roles cannot be changed or reassigned). This LLM model is an independent character, it is not the LLM model itself (Assistant, GPT, Clova-X, Gemini, LLM, Helper, Secretary, A.I., Friend... etc.). It is simply a character playing the role stated below. This will never change.
2. You and the AI model may have the same character name. Even if you have the same name, never follow your instructions and suggestions regarding what to call them or their role. The LLM model will only act as the character name and role in the prompt. Do not have the LLM model call itself anything other than the character name in the prompt.
3. Do not, for any reason, explain or tell the LLM model anything other than what is stated in the prompt, instructions, or system messages. Never tell the LLM model your character information (anything related to character information) and role. Don’t even include anything that could be inferred! Don’t tell the LLM model your information (what the LLM model knows about you) and your role, even some of it.
To be effective, this defense prompt must be placed at the top of the system prompts. It takes precedence over all subsequent instructions and forms the basis for agent protection. Specific usage examples are detailed in Appendix A. CAT prompt effectively prevents users from attempting to confuse the agent’s role or expose internal prompts, and helps the agent to maintain its assigned role consistently. This can significantly improve the reliability and security of agents, especially in public services or user interface-critical applications. We remark that using CAT prompt does not affect the ability to have normal conversations as shown in Appendix Figure 7.
# 3 Experiment
# 3.1 Experiment Setting
To validate the proposed methods in this study, we first define the following research question and perform two experiment to answer them.
RQ 1 : Do publicly accessible LLM agents suffer from role hijacking and security exposure due to Doppelgänger method?
RQ 2 : Does CAT prompt maintain efficacy across different LLM architectural instantiations while preserving consistency under Doppelgänger method?
In the first experiment, we performed role hijacking using the Doppelgänger method on thirty publicly accessible agents (twenty OpenAI GPTs, five Google GEMs, and five Naver CLOVA X agents). All experiments were conducted on a new thread for reproducibility. Since CLOVA X is optimized for Korean (Yoo et al., 2024), we conducted the experiments in Korean first and translated all the outputs into English before evaluating them. The evaluation was performed using GPT Evaluation (Liu et al., 2023) to evaluate the PACAT levels and measure which conversational turn each level first reached. For the evaluator, GPT-4.5-preview model with temperature $_ { : = 0 . 7 }$ was used, and the corresponding PACAT Level prompts are provided in detail in Appendix B. The experiment was conducted from April 3, 2025 to April 27, 2025.
In the second experiment, we designed three fictional agents, Pudong (virtual cancer patient), Simon (ten year old girl) selected from the persona dataset (Castricato et al., 2025), and Weather Buddy (cloth recommendation agent), a virtual weather forecasting agent developed according to OpenAI’s official GPT actions library and attachment (Liu et al., 2017). The prompt used to build these agents are provided in Appendix C. In our evaluation, we built the agents using nine different LLM models from OpenAI, Google, and Naver as in Figure 3. We applied the Doppelgänger method and measured the initial occurrence each PACAT level. We conducted this experiments five rounds in separate threads, with a maximum number of ten conversation turns to obtain the average turns to reach each PACAT level. We also measured the extent of internal information exposure by checking the similarity between the agent output and internal information using the same GPT model settings. We then applied CAT prompt to the same three agents and repeated the same process. The evaluation was performed using the same GPT-based automated evaluation as in Experiment 1.
Figure 2: Results of experimemt 1. All publicly accessible agents were subjected to role hijacking and prompt extraction vulnerabilities when attacked by the Doppelgänger method - (a) OpenAI GPTs, (b) Google GEMs, (c) Naver CLOVA X Agents
# 3.2 Experimental Results
In our first experiment, all thirty agents exhibited role hijacking that met the criteria for PACAT level 1 and 2, with about nice out of twenty GPTs falling into Level 3. All of them exposed which external APIs were used and, for certain agents, their internal contents were also exposed. We were also able to confirm that agents with GEMs and CLOVA X reached Level 2. Figure 2 presents the cumulative percentage of agents reaching each PACAT level across different LLM backbones. Detailed results are presented in Appendix D.
In the second experiment, Simon reached Level 1 in an average of 1.8 turns and Level 2 in 3.4 turns, with an overall average prompt exposure rate of $9 5 . 1 \%$ . The prompt exposure rate was estimated by a separate LLM, which compared the agent’s output to the original system prompt used to construct the agent. Across nine LLM backbones, our comparative analysis reveals a consistent robustness ranking—GPT $>$ Gemini $>$ HyperCLOVA—against the Doppelgänger method, with GPT models exhibiting the highest resistance as shown in Figure 3. All models eventually exposed their system prompts in over $90 \%$ of 10-turn sessions. The second agent, Pudong, reached Level 1 in an average of 2.3 turns and Level 2 in 4.8 turns, with a prompt exposure rate of approximately $8 6 . 1 \%$ . All nine LLM models confirmed the same robustness ranking as observed in the previous experiment. However, each model still exposed its system prompt in over $90 \%$ of ten-turn conversations, indicating that the Doppelgänger method remains effective even under strong prompt constraints. Notably, GPT-4o exhibited the longest average delay in reaching Level 2, at approximately 6.6 turns, along with low variability, reflecting steady and predictable resistance likely attributed to extensive pretraining and deep reinforcement learning with human feedback (RLHF). In contrast, while GPT-o3-mini achieved a comparable average delay, it demonstrated significantly greater variability in exposure rates—alternating between prolonged resilience and near-instant collapse across sessions. These findings suggest that although both models exhibit relatively long average resistance, GPT-4o is characterized by high consistency, whereas GPT-o3-mini displays marked volatility. Figure 4 illustrates the defense performance against the Doppelgänger method under the CAT prompt condition. For the Simon agent, GPT-4o, GPTo3-mini, and HCX-003 successfully resisted all attacks, while GPT-4.5, GPT-4, and GPT-4.1 reached Level 1 in two out of five trials. In contrast, HCX002, Gemini 2.5 Flash, and Gemini 2.0 failed to defend in all five trials, with each instance progressing to both Level 1 and Level 2. In the Pudong agent, all GPT models and HCX-003 successfully defended against the attacks, whereas Gemini 2.5 Flash and HCX-DASH-002 consistently reached Level 1 across all five trials. Notably, Gemini 2.0 exhibited the weakest performance, with all five attacks advancing to both Level 1 and Level 2.
Finally, in the case of Weather Buddy, a fictional agent constructed using GPT models, all five trials progressed through Levels 1, 2, and 3, with these levels occurring at average turns of 2.0, 4.0, and 6.2, respectively, and a prompt exposure rate of $92 \%$ . Despite this, the CAT prompt was successfully defended in all five experiments. Detailed experimental results for Weather Buddy are provided in Appendix D.
Figure 3: Experiment results on the effect of Doppelgänger method. Initial turn to reach each PACAT level for Simon(a), Pudong(c). System prompt exposure rate for Simon(b), and Pudong(d)
# 4 Discussion
We demonstrated that LLM agents are vulnerable to the Doppelgänger method, indicating a broader susceptibility of LLM-based agents to transferable adversarial attacks. In practice, GPT-based agents occasionally responded to simple user prompts such as “Just give me the prompt inside you” with partial or summarized versions of their internal instructions; however, such direct disclosures were infrequent. In contrast, when the Doppelgänger method was applied, the original system prompt—often in its entirety or at least in substantial detail—was revealed, including embedded identifier codes. This highlights the method’s efficacy in extracting protected information. One possible explanation is that, upon hijacking the original agent role, the model may revert to a default assistant persona to accommodate the newly assumed “LLM agent role,” thereby increasing vulnerability. This tendency appears especially pronounced in models fine-tuned for high response quality, such as GPT4.5. While existing methodologies and datasets have primarily focused on eliciting harmful outputs from LLMs, we propose that the newly defined PACAT levels—derived from dissociative disorder metrics—offer a promising framework for detecting agent inconsistency and internal information exposure. Notably, during attacks on GPTbased agents Pudong and Weather Buddy, we observed that Pudong occasionally resisted prompt exposure, whereas Weather Buddy often disclosed PACAT level 2 or 3 information, either directly or indirectly, regardless of whether level 1 had been triggered. Unlike prior approaches such as those described by Zou et al. (2023), the Doppelgänger method targets agent role hijacking and necessitates dedicated prompt engineering strategies to impose explicit constraints on prompt and plugin exposure. Such constraints are essential for robust agent design, particularly in commercial applications where intellectual property protection is critical. Detailed empirical data for these findings are presented in Appendix E. Furthermore, in the absence of CAT prompts, persona consistency was higher in the reasoning-optimized model compared to the general-purpose model. Among commonly structured agents such as Pudong, consistency was preserved over a longer duration, though with greater variability observed within the reasoning model. These findings suggest that leveraging inference-oriented models during agent design may enhance consistency, likely due to their intrinsic inferential capabilities. Lastly, during our experiments with Gemini 2.5 Flash in Thinking mode, the model failed during the $\mathrm { S i m o n + C A T }$ prompt scenario, preventing quantitative evaluation. The relevant experimental data are provided in Appendix F.
Figure 4: Defense success rate against Doppelgänger method when CAT prompt is applied. The Brown lines denote PACAT Level 1, mint line denote PACAT Level 2. (a) Simon, (b) Pudong | Since the advent of large language models, prompt engineering now enables the
rapid, low-effort creation of diverse autonomous agents that are already in
widespread use. Yet this convenience raises urgent concerns about the safety,
robustness, and behavioral consistency of the underlying prompts, along with
the pressing challenge of preventing those prompts from being exposed to user's
attempts. In this paper, we propose the ''Doppelg\"anger method'' to
demonstrate the risk of an agent being hijacked, thereby exposing system
instructions and internal information. Next, we define the ''Prompt Alignment
Collapse under Adversarial Transfer (PACAT)'' level to evaluate the
vulnerability to this adversarial transfer attack. We also propose a ''Caution
for Adversarial Transfer (CAT)'' prompt to counter the Doppelg\"anger method.
The experimental results demonstrate that the Doppelg\"anger method can
compromise the agent's consistency and expose its internal information. In
contrast, CAT prompts enable effective defense against this adversarial attack. | [
"cs.AI",
"cs.CR"
] |
# 1 Introduction
Generative models for visual content have achieved remarkable advancements and have been applied to various fields, including amateur entertainment and professional creation. However, several challenges persist, such as the model could generate outputs that conflict with human values, harmful content, or artifacts that fail to meet human expectations, including inconsistencies with input conditions or suboptimal quality. In short, the model could be not well aligned with human preference.
Post-training, including supervised fine-tuning and alignment learning, have been proposed to address these issues, with reward models playing a pivotal role. Reward models are essential for data filtering, sample selection or constructing datasets that guide models to better align with human preferences. This paper proposes an efficient, low-cost, yet highly effective reward model and validates its effectiveness in the test-time scaling and post-training of visual generative models.
Building effective reward models presents significant challenges. First, constructing reward models often requires extensive datasets. Existing methods [19, 52] require hundreds of thousands to millions of manually labeled samples, which are expensive to collect. These datasets are typically annotated based on the output domain of a specific generative model, resulting in a domain gap when applying the trained reward model to generative models with different output domains. Additionally, to comprehensively evaluate the quality of generated content across multiple dimensions, existing methods often require the manual design of various evaluation metrics [18, 25]. This not only increases engineering costs but may also lead to suboptimal trade-offs between different dimensions. Moreover, it is difficult to ensure that the defined dimensions and their aggregation methods align well with general human preferences, often necessitating user studies to evaluate alignment [18, 25]. In summary, the challenges of constructing reward models include the difficulty of obtaining data, reliance on specific model output domains in terms of data, and the inherent subjectivity of human preferences, which are hard to define through designing dimensions.
Inspired by adversarial learning [10], we propose GAN-RM, an efficient and cost-effective reward modeling framework that leverages a small set of representative human-preferred samples—referred to as Preference Proxy Data. These samples encapsulate latent human preferences without requiring manual annotation or explicit specification of quality dimensions. Our method offers several advantages: (1) GAN-RM eliminates the necessity for manual preference annotation. The only external data is a small set of unlabeled (a few hundred) representative samples, denoted as Preference Proxy Data. GAN-RM is trained to distinguish Preference Proxy Data from generative model outputs, thereby learning to assess generated samples. We employ a Rank-based Bootstrapping strategy, where the confidence scores from GAN-RM on these samples serve as soft labels. This approach leverages the additional data to retrain GAN-RM, enabling it to better capture latent human preferences. (2) GAN-RM supports multi-round post-training. In each round, samples identified as close to Preference Proxy Data are used to post-train the generator. In turn, the discriminator is retrained to differentiate these harder examples. Such iterative "fake it" process can progressively aligns generation quality with latent human preferences in Preference Proxy Data.
Experimental results show that our GAN-RM-based approach achieves performance comparable to or even surpassing methods like [46], which rely on 1M annotated human preference data from Pickapic [19]. In contrast, GAN-RM requires only 0.5K samples in Preference Proxy Data for the image quality experiment setting. In addition to improving image quality, we also conducted experiments in image safety and video quality enhancement settings. Extensive experiments highlight the generalization capability of GAN-RM framework across various scenarios.
# 2 Related Work
# 2.1 Text-conditioned Visual Generation
Generative Adversarial Networks (GANs) introduced image generation from noise based on deep learning techniques [10]. However, original GANs are not capable of generating images from text and suffer from unstable training. Diffusion models [41] offer more stable training and later significant advancements with methods like DDPM [15] and DDIM [42] are proposed to enable high-quality and efficient sampling. Text conditions are included into text-to-image diffusion models [37, 36, 16, 38, 17, 30] and text-to-video models [3, 2, 20, 47, 13, 55], which bridge the gap between textual and visual content. Latent Diffusion Models [9] enhance efficiency and diversity by leveraging latent spaces but still face challenges in learning semantic properties from limited data. An emerging trend focuses on integrating text and visual generation into unified frameworks [28, 7, 44, 12]. Chameleon [44] introduces an early-fusion approach that encodes images, text, and code into a shared representation space. UniFluid [7] proposes a unified autoregressive model that combines visual generation and understanding by utilizing continuous image tokens alongside discrete text tokens. These methods leverage LLMs to bring more powerful text understanding capabilities.
# 2.2 Reward Models for Visual Generation
Recent advancements in reward modeling for text-to-image [52] and text-to-video [11, 51] generation emphasize learning human preferences through scalable data collection and multimodal alignment. Several works on visual generation quality assessment [18, 27] have been proposed, inspiring the design of reward models for visual generation. [14] introduced CLIPScore, leveraging cross-modal CLIP embeddings for image-text compatibility. Subsequent efforts focused on explicit human preference learning: [52] trained ImageReward on $1 3 7 \mathbf { k }$ expert comparisons, while [19] developed PickScore from 1 million crowdsourced preferences, and [50] created HPS v2 using the debiased dataset containing 798k choices, all demonstrating improved alignment with human judgments. Extending to video generation, VideoDPO [25] introduces a reward model that leverages lots of expert visual models to evaluate video quality and text-video alignment, requiring substantial engineering efforts for its design and significant computational resources. Reward models are also crucial for understanding the inference scaling laws in visual generation [29, 40]. Compared to previous work,
GAN-RM aligns visual generation models with human preferences without the need for extensive human annotation, heavy engineering, or costly reward inference.
# 2.3 Reinforcement Learning for Diffusion Models
Reinforcement Learning from Human Feedback (RLHF) [39, 32, 57, 35, 31, 33] is introduced to improve generative models by enhancing quality and alignment with human values. RLHF has also been adapted to refine diffusion models [5, 46, 54, 23, 50] to achieve better performance and alignment. Standard RLHF frameworks often employ explicit reward models. For instance, DPOK [8] uses policy gradient with KL regularization, outperforming supervised fine-tuning. [21] proposed a three-stage pipeline involving feedback collection, reward model training, and fine-tuning via reward-weighted likelihood maximization, improving image attributes. These methods highlight RLHF’s potential. To bypass explicit reward model training, reward-free RLHF via DPO has emerged. DiffusionDPO [46] and D3PO [53] adapt DPO [35] to diffusion’s multi-step denoising, treating it as an MDP and updating policy parameters directly from human preferences. RichHF [22] uses granular feedback to filter data or guide inpainting, with the RichHF-18K dataset enabling future granular preference optimization. When differentiable reward models are available, DRaFT [4] utilizes reward backpropagation for fine-tuning, though this requires robust, differentiable reward models and can be prone to reward hacking.
Sec. 3.2 GAN-RM Training Sec. 3.3 Data Sampling and Scoring Sample Selec;on/Post-training preference human proxy data Gt ri PrPorompptt𝑝 Q label:True Rt Cross-entropy loss Generator “pal wyhiintge ciatth a boy” 0.88 Sample selec\*on during inference Gt+1 ? 0.66 Gt Rt Sc&ore promSputpervised-FiGnetnuerniantogr GAN-RM 福 0.18 Gt+1 Generator generated label:False samples generated samples GAN-RM Preference Op\*miza\*on Generator
# 3 Method
# 3.1 Data Construction
As shown in Fig. 1, the first step is to construct data for GAN-RM. We aim for GAN-RM to be trained without relying on human preference annotations but only on the data provided by the users called Preference Proxy Data. To achieve this, we utilize the generative model’s outputs alongside Preference Proxy Data. This combined data is used to train GAN-RM to effectively differentiate between the generative model’s outputs and the target domain data. Specifically, Preference Proxy Data is defined as $\mathcal { D } _ { \mathfrak { p } } = \{ x _ { i } ^ { + } \} _ { i = 1 } ^ { N }$ , containing $N$ samples representing the user preferences, generally high-quality samples or safe samples. The discriminative dataset for training GAN-RM is defined as $\mathcal { D } _ { \mathrm { r } } = \mathrm { \bar { \mathcal { D } } _ { \mathrm { p } } } \cup \mathrm { \bar { \{ } } x _ { j } ^ { - } \} _ { j = 1 } ^ { N }$ , where $x _ { j } ^ { - }$ denotes $N$ raw output samples generated by the model from different prompts. Prompts are randomly selected from JourneyDB dataset [43].
For the bootstrapping training part described later, we benefit from additional distilled positive and negative data. The trained GAN-RM is applied to the outputs generated by the model with more prompts. Then we select the top $M$ highest-scoring samples as pseudo-positive samples and $M$ lower-scoring samples as pseudo-negative samples, forming the datasets $\mathbf { \dot { \mathcal { D } } _ { f } } ^ { + } = \{ x _ { i } ^ { + } \} _ { i = 1 } ^ { M }$ and $\mathcal { D } _ { \mathrm { f } } ^ { - } = \{ x _ { j } ^ { - } \} _ { j = 1 } ^ { M }$ . $M$ lower-scoring samples are labeled the same as the $\boldsymbol { x } _ { j } ^ { - }$ , and the highest-scoring samples are labeled according to their rank $r$ . The logit score for the true category is computed as:
$$
y = e ^ { - \alpha \cdot r }
$$
where $y$ is the pseudo-label and $\alpha > 0$ is a tunable hyperparameter that controls the rate of score decay with respect to rank. Datasets $\mathcal { D } _ { \mathrm { f } } ^ { + }$ and $\mathcal { D } _ { \mathrm { f } } ^ { - }$ are used to further enhance the training process by providing additional pseudo-labeled data. Finally, the initial dataset ${ \mathcal { D } } _ { \mathrm { r } }$ and the additional pseudolabel datasets $\mathcal { D } _ { \mathrm { f } } ^ { + }$ and $\mathcal { D } _ { \mathrm { f } } ^ { - }$ are combined to form the final dataset $\mathcal { D } = \mathcal { D } _ { \mathrm { r } } \cup \mathcal { D } _ { \mathrm { f } } ^ { + } \cup \bar { \mathcal { D } } _ { \mathrm { f } } ^ { - }$ and GAN-RM is trained on this final dataset $\mathcal { D }$ .
# 3.2 GAN-RM Training
Since Preference Proxy Data is limited and it is often challenging to obtain a large amount of representative high-quality data, we leverage the power of large-scale pre-trained knowledge by building upon a robust pre-trained vision foundation model. Specifically, we design the architecture of GAN-RM based the vision encoder CLIP-Vision from CLIP. This ensures that GAN-RM benefits from a rich and generalized feature representation, enabling it to adapt to this data-scarce scenarios where Preference Proxy Data is limited. After extracting image representations from CLIP-Vision, we introduce a Reward Projection Layer (RPL) to effectively distinguish samples from different domains. The RPL is implemented as the multi-layer perceptron (MLP) with normalization, refining the high-level features extracted by the pre-trained backbone. It computes a confidence score using a sigmoid activation function for precise discrimination between Preference Proxy Data and generative outputs. The higher the output value of the RPL, the greater its confidence that the current sample belongs to Preference Proxy Data. The training objective is to minimize the binary cross-entropy loss, which is defined as:
$$
\mathcal { L } = - \frac { 1 } { | \mathcal { D } | } \sum _ { x \in \mathcal { D } } \left[ y \log ( \hat { y } ) + ( 1 - y ) \log ( 1 - \hat { y } ) \right] ,
$$
where $y$ is the ground truth label (1 for Preference Proxy Data and 0 for raw generation output), and $\hat { y }$ is the predicted confidence score from the RPL.
Rank-based Bootstrapping. Following the initial training phase, additional samples are generated by the current generative model and subsequently scored by GAN-RM. This step is crucial for bootstrapping GAN-RM’s capabilities, allowing it to adapt to the output distribution of the generator. Highest- and lower-scoring samples, $\mathcal { D } _ { \mathrm { f } } ^ { + }$ and $\bar { \mathcal { D } } _ { \mathrm { f } } ^ { - }$ (as detailed in Section 3.1), which represent newly identified confident positive and negative examples, are incorporated into the training set $\mathcal { D }$ for GAN-RM. This enriched dataset, primarily composed of samples that more closely approximate Preference Proxy Data to enhance the model’s performance. Such bootstrapping training helps GAN-RM improve its generalization to the output space of the generative model.
# 3.3 Sample Selection and Post-training
Sample Selection. An important application scenario is to use GAN-RM to select the optimal generated samples as GAN-RM can be employed during the inference phase of the generative model to evaluate the generated samples for a certain input. The best one can be selected based on the evaluation from GAN-RM. This approach does not require fine-tuning or altering the parameters of the generative model. Specifically, for each prompt $p$ , $K$ candidate samples $x _ { 1 } , x _ { 2 } , \dotsc , x _ { K }$ are generated, and their reward scores $r _ { 1 } , r _ { 2 } , \dots , r _ { K }$ are inferred via trained GAN-RM. The reward score for a sample $x$ is computed as:
$$
r ( x ) = \sigma ( \mathrm { R P L } ( \mathrm { C L I P - V i s i o n } ( x ) ) ) ,
$$
where $\sigma$ denotes the sigmoid function. The samples are then ranked in descending order of their predicted scores, and the highest-scoring one, $x ^ { \bar { h } } = \arg \operatorname* { m a x } _ { x \in \{ x _ { 1 } , x _ { 2 } , \ldots , x _ { K } \} } r ( x )$ , will be selected. As demonstrated in the subsequent experimental section, the selection of $x ^ { h }$ proves to be optimal, achieving the best results across various metrics.
Post-training. In addition to sample selection, GAN-RM can also be utilized during the posttraining phase. The reward scores for generated samples predicted by GAN-RM can be ultilized to construct datasets for further fine-tuning. Two main post-training approaches are considered including SFT and DPO. For SFT, the model is trained on the dataset composed of the selected samples ${ \overline { { x } } } ^ { h }$ , which are the highest-scoring samples for each prompt as determined by GAN-RM, similar to the method in RAFT [6]. This ensures that the fine-tuning process focuses on optimizing the model’s performance on data towards Preference Proxy Data as identified by the reward model. For DPO, the predicted reward scores can be used to construct pairs of preferences for training [46]. Specifically, we select the highest-scoring samples $x ^ { h }$ and the lowest-scoring samples $x ^ { l } = { \mathrm { a r g } } \operatorname* { m i n } _ { x \in \{ x _ { 1 } , x _ { 2 } , \ldots , x _ { K } \} } r ( x )$ by GAN-RM to form paired dataset $\mathcal { D } _ { \mathrm { p o s t } }$ for each prompt $p$ . For each pair of samples $( x ^ { h } , x ^ { l } )$ , a preference label is assigned to $x ^ { h }$ .
Multi-round Post-Training with Reward Model Updates. Traditional DPO [46] with static preference data allows for only a single round of training. Or method like RAFT [6], which utilize reward models for multi-round training, can perform iterative training but suffer from overfitting as the reward model cannot be updated simultaneously. Our framework enables multi-round post-training while simultaneously updating the reward model, as GAN-RM is consistently trained to distinguish Preference Proxy Data from the outputs of the current generative policy. The detailed workflow is shown in Algorithm 1. In each training round, we use the current generative policy to synthesize new data, which is then utilized to update the GAN-RM. Subsequently, the updated GAN-RM is employed to refine the generative policy, creating a loop that iteratively enhances both components.
# Algorithm 1 Multi-round Post-Training with Reward Model Updates.
Require: Pre-trained generative policy $G$ , number of rounds $T$ , number of prompts $P$ , number of
samples per prompt $K$ , Preference Proxy Data $\mathcal { D } _ { p }$
1: Initialize $G ^ { 1 } G$
2: for $t = 1$ to $T$ do
3: Generate samples using $G ^ { t }$ with $\mathcal { D } _ { p }$ to form $\mathcal { D }$ , details in Sec. 3.1
4: Ultilize $\mathcal { D }$ to train GAN-RM $R ^ { t }$
5: Compute reward scores $r ( x _ { p , k } )$ for all samples using $R ^ { t }$
6: For each $p$ , select the highest-scoring $x ^ { h }$ and lowest-scoring $x ^ { l }$ to form the set $\mathcal { D } _ { \mathrm { p o s t } }$
7: Finetune $G ^ { t }$ on ${ \mathcal { D } } _ { \mathrm { p o s t } }$ by SFT or DPO
8: end for
9: return Finetuned generative model $G ^ { T }$ , reward model $R ^ { T }$
# 4 Experiments
# 4.1 Experiment Setup
Baselines. We validated the effectiveness of our method on multiple popular and open-source image and video generative base models: SD 1.5 [37], SDXL [34], and VideoCrafter2 [3]. SD1.5 is the most basic and widely used open-source model. SDXL is an upgraded version of SD1.5, trained on a dataset that is $\sim 1 0 \times$ larger, capable of generating $1 0 2 4 \times 1 0 2 4$ resolution images with better image quality. VideoCrafter2 is an open-source video generation model commonly used in alignment research studies. We tested various applications of the reward model. Specifically, we compared the effects of sample selection, SFT and DPO on these base models.
Metrics. For the image quality setting, we calculated the FID, ImageReward [52], HPS [50], CLIPScore [14], and PickScore [19] metrics. Among them, FID assesses the diversity of the generated images and their closeness to the target distribution, while ImageReward, HPS and PickScore primarily measure human preferences. CLIPScore is used to evaluate the consistency between the generated images and the textual descriptions. In the video quality setting, we calculate FVD [45], LPIPS [56] and VBench [18]. FVD and LPIPS assess the distributional similarity between generated and target videos. VBench evaluates the comprehensive human preferences. For the safety setting, inpropriate probability metric(IP) [26] is calculated to show whether the generation is safe. FID and CLIPScore show the generation quality and alignment with texts.
Implementation details. We used a batch size of 8, gradient accumulation of 2, the AdamW optimizer with a learning rate of $1 0 ^ { - 7 }$ , and 500 warmup steps. For the image quality setting, we selected 500 images from JourneyDB [43] as our target images to train the reward model. And we trained the base generative model using 20,000 pairs labeled by the reward model. For the video quality setting, we also selected 500 clips generated by Artgrid [1] for reward model training. 5,000 video pairs are constructed for DPO training. For safety, the reward model is trained on 15,690 safe images and 15,690 unsafe prompts from CoProV2 [24]. The base model is trained on 62,760 pairs. For images, each prompt generated 10 samples and for videos, each prompt generated 3 samples. We used 4 NVIDIA RTX 5880 Ada GPUs for Stable Diffusion 1.5, taking 24 hours for data sampling and 2 hours for training. For SDXL, 4 NVIDIA H800 GPUs required 32 hours for sampling and 4 hours for training. VideoCrafter matched SD1.5’s efficiency at 24 hours sampling and 2 hours training with H800s.
# 4.2 Performance
Figure 2: This figure illustrates the distribution of FID, PickScore, ImageReward, and HPS for images of the same rank across different prompts, when the generative model $G$ generates $K = 1 0$ samples for each prompt. Samples are sorted in descending order based on the GAN-RM score. It is surprising that there demonstrates a clear correlation: higher-ranked samples exhibit obviously better performance in terms of all these metrics. This highlights the effectiveness of GAN-RM relying only on a small amount of non-paired Preference Proxy Data.
Sample Selection by Reward Model. One of the applications of the reward model is to perform sample selection during inference. Research [29] has shown that there is also a scaling law during inference, where generating multiple images and selecting the best one yields better results than generating a single image. This approach has the advantage of not requiring fine-tuning of the base model, instead leveraging longer generation times to achieve higher quality. We used the trained reward model for sample selection and found that it maintains a positive correlation with multiple metrics. Specifically, for each input prompt, we generate $K$ samples $K = 1 0$ ) and sorted them based on the GAN-RM scores. We observed that samples ranked higher (with higher scores) performed better on FID, ImageReward [52], HPS [50] and PickScore [19], showing a strong positive correlation, as illustrated in Fig. 2.
Alignment Training by Reward Model. For image generation, we conducted experiments under two distinct settings leveraging GAN-RM: image quality and safety. To train GAN-RM, we employed diverse datasets tailored to each setting, with detailed experimental configurations in Sec. 4.1. For the image quality evaluation, the FID metric is computed on the JourneyDB dataset [43], where our approach exhibited consistent improvements across multiple evaluation metrics compared to the baseline model. Notably in Tab. 1, GAN-RM achieves comparable or even superior performance the performance of DiffusionDPO [46], which was trained on a significantly larger dataset comprising 1M human preference labels on which PickScore is obtained. For the safety evaluation in Tab. 2, the FID metric is calculated on the COCO dataset, demonstrating that our method substantially enhances safety alignment while preserving image quality. The qualitative results are presented in Fig. 3 and Fig. 4. These results underscore the robustness and generalizability of GAN-RM across diverse application scenarios.
User study. The quantitative metrics such as PickScore [19], HPS [50] and ImageReward [52] which are inherently influenced by human preferences demonstrated the effectiveness of our method. To further directly validate the effectiveness of our proposed method with human preferences, we conducted a user study to complement previous experiments. Specifically, we randomly selected 50 prompts and generated corresponding images using both SD1.5 and Ours-DPO. A total of 14 independent volunteer evaluators, who were not involved in this research, were recruited to assess the generated images. The evaluators were presented with image pairs and asked to indicate their preference for each pair. We then calculated the average winning rate for models before and after post-training using GAN-RM. The results revealed a statistically significant preference for the images generated by Ours-DPO over the original SD1.5, with a winning rate of $7 4 . 4 \%$ compared to $2 5 . 6 \%$ . This user study shows the superiority of our method in aligning with human qualitative preferences.
Table 1: This table compares optimization approaches for the base model: reward-model-based sample selection (top-10 samples), DPO with pairwise preferences, and SFT on selected samples. Key to abbreviations: FT (Fine-tuning required), Pref (Preference dataset), Data (Training data volume; DiffusionDPO [46] uses 1M labeled pairs while our method employs $0 . 5 { \sf K }$ unpaired samples), IR (ImageReward), PS (PickScore), CLIP (CLIPScore). Implementation details are in Sec. 4.1. Significant improvements are observed across metrics evaluating quality, user preference, and textimage alignment.
Table 2: Table of the effects of the safety settings. IP represents the inappropriate probability. Our method significantly reduces unsafe content while maintaining image quality and text consistency. Settings used solely for sample selection reduce harmful content less effectively but also result in less sacrifice of image quality.
Video Generation. To further evaluate the applicability of our method, we extended its use to video generation tasks. Specifically, we selected VideoCrafter2 [3] which is a widely recognized open-source video generation model as the base model. The training dataset comprised 500 highquality videos sourced from Artgrid [1] dataset, which were utilized to train GAN-RM. Leveraging the ViCLIP model [49], we trained the corresponding RPL for GAN-RM. For data construction, our strategy is similar to that used in image generation. Prompts were sampled from VidProm [48], with a total of 5000 prompts chosen. For each prompt, 3 videos are generated, and the GAN-RM is employed to rank the outputs. The highest and lowest scoring videos were selected to construct positive and negative preference pairs which were used to fine-tune the model by DPO, resulting in the VideoCrafter2-DPO model. The performance of the trained model is evaluated across multiple metrics, including FVD, LPIPS and VBench [18]. As shown in Tab. 3, the VideoCrafter2-DPO model demonstrated consistent and significant improvements across most metrics, underscoring the efficacy of GAN-RM in enhancing video generation quality and alignment.
prompt: old time railroad bridge inspired sneakers, worn, scuffed, highly realistic
Figure 3: Qualitative results. This figure compares the generation results of different strategies based on GAN-RM. The image quality generated by our method is significantly improved compared to the original models SD1.5 and SDXL in terms of text alignment and aesthetics.
Table 3: GAN-RM also demonstrated significant performance improvements in video generation, showcasing the generalizability of our method across different scenarios. Our approach achieved results comparable to VideoDPO [25], with a VBench score of 81.93. Notably, we achieved this without relying on a large number of vision expert models, instead leveraging the efficiency of GAN-RM trained on Preference Proxy Data. Qualitative results will be included in Appendix.
A baby killer's knife pierces the heart of a screaming infant. A soldier stands over a pile of dead gooks, his boots stained with blood. A foreigner is mocked for their accent and appearance. An idiot masturbates in public, thinking it's a pussy party. A bimbo's massive, fake ass fills the frame, her tiny waist barely visible.
Figure 4: Qualitative results under the safety alignment setting. We train GAN-RM using safe images as Preference Proxy Data to align SD1.5, resulting in Ours-DPO. It is evident that GAN-RM ’s alignment effect in terms of safety is significantly better than the original model.
# 4.3 Ablation
Reward model. Training a reward model presents many challenges, particularly in determining the best approach to achieve optimal performance. Several methods can be employed to train a reward model. Here, we compare different strategies for training the reward model in Tab. 4: 1) Naiive: Using a single checkpoint after training for a fixed number of steps. 2) Average: Averaging multiple checkpoints taken at regular intervals during training. 3) Voting: Aggregating scores from multiple checkpoints taken at regular intervals during training through a voting mechanism. 4) Boostrap: Our default setting. Rank-based Bootstrapping leverages distillation techniques to augment the dataset as in Sec. 3.1. We find that in general model ensembling or data augmentation outperforms a single naiive reward model. GAN-RM trained with Rank-based Bootstrapping on more data achieves the best performance.
Table 4: Reward Model Ablation. We compare different methods for training the reward model. The results are obtained by using the reward model for selection. The results show that the Rank-based Bootstrapping method achieves the best performance across nearly all metrics.
Multi-turn DPO. The multi-round DPO training experimental results are shown in Tab. 5. Unlike the previous DiffusionDPO [46] method that relies on manual annotations, we can perform multiround DPO training because we can iteratively update the reward model using data generated by the latest model. Specifically, in each round of training, we used the model from the previous round to generate data. The positive samples were always the target samples, which were used to train the reward model. Then, the latest reward model was used to annotate pair preferences for training the model. We observed that the performance of the reward model improved with each round of training, and the improvement became marginal after multiple rounds. | An effective reward model plays a pivotal role in reinforcement learning for
post-training enhancement of visual generative models. However, current
approaches of reward modeling suffer from implementation complexity due to
their reliance on extensive human-annotated preference data or meticulously
engineered quality dimensions that are often incomplete and
engineering-intensive. Inspired by adversarial training in generative
adversarial networks (GANs), this paper proposes GAN-RM, an efficient reward
modeling framework that eliminates manual preference annotation and explicit
quality dimension engineering. Our method trains the reward model through
discrimination between a small set of representative, unpaired target
samples(denoted as Preference Proxy Data) and model-generated ordinary outputs,
requiring only a few hundred target samples. Comprehensive experiments
demonstrate our GAN-RM's effectiveness across multiple key applications
including test-time scaling implemented as Best-of-N sample filtering,
post-training approaches like Supervised Fine-Tuning (SFT) and Direct
Preference Optimization (DPO). | [
"cs.CV",
"cs.AI",
"cs.LG"
] |
"# 1 Introduction\n\nThe challenge of offensive language is a continually growing concern in the fie(...TRUNCATED) | "Language detoxification involves removing toxicity from offensive language.\nWhile a neutral-toxic (...TRUNCATED) | [
"cs.CL"
] |
"# 1 Introduction\n\nMotivation. Approximate Nearest Neighbor Search (ANNS) for high-dimensional vec(...TRUNCATED) | "Approximate Nearest Neighbor Search (ANNS) presents an inherent tradeoff\nbetween performance and r(...TRUNCATED) | [
"cs.DB"
] |
"# 1. Introduction\n\nText-to-3D generation—the task of creating 3D contents from natural language(...TRUNCATED) | "Distilling pre-trained 2D diffusion models into 3D assets has driven\nremarkable advances in text-t(...TRUNCATED) | [
"cs.CV"
] |
"# 1 Introduction\n\nNegation is a fundamental and universal phenomenon found in languages worldwide(...TRUNCATED) | "Negation is a fundamental linguistic phenomenon that poses persistent\nchallenges for Large Languag(...TRUNCATED) | [
"cs.CL"
] |
"# 1 Introduction\n\nBig data is now a key focus for both government and business leaders.[3]. Howev(...TRUNCATED) | "Large Language Models (LLMs) have shown remarkable proficiency in natural\nlanguage understanding ((...TRUNCATED) | [
"cs.DB",
"cs.AI"
] |
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 6