Search

Large deep learning models are impressive, but they struggle when real-time data is not available. Few-shot class-incremental learning (FSCIL) poses a significant challenge for deep neural networks to learn new tasks from just a few labeled samples without forgetting the previously learned ones. This setup can easily leads to catastrophic forgetting and overfitting problems, severely affecting model performance. Studying FSCIL helps overcome deep learning model limitations on data volume and acquisition time, while improving practicality and adaptability of machine learning models. This paper provides a comprehensive survey on FSCIL. Unlike previous surveys, we aim to synthesize few-shot learning and incremental learning, focusing on introducing FSCIL from two perspectives, while reviewing over 30 theoretical research studies and more than 20 applied research studies. From the theoretical perspective, we provide a novel categorization approach that divides the field into five subcategories, including traditional machine learning methods, meta learning-based methods, feature and feature space-based methods, replay-based methods, and dynamic network structure-based methods. We also evaluate the performance of recent theoretical research on benchmark datasets of FSCIL. From the application perspective, FSCIL has achieved impressive achievements in various fields of computer vision such as image classification, object detection, and image segmentation, as well as in natural language processing and graph. We summarize the important applications. Finally, we point out potential future research directions, including applications, problem setups, and theory development. Overall, this paper offers a comprehensive analysis of the latest advances in FSCIL from a methodological, performance, and application perspective. • In-depth survey on FSCIL methods, applications, performance. • Categorizing FSCIL into five subcategories for clear analysis. • Evaluating FSCIL research on benchmarks for strengths, weaknesses. • FSCIL applications in computer vision, NLP, and graph. [ABSTRACT FROM AUTHOR]

Deep neural networks have achieved outstanding performance in computer vision tasks. Convolutional Neural Networks (CNNs) typically operate in the spatial domain with raw images, but in practice, images are usually stored and transmitted in their compressed representation where JPEG is one of the most widely used encoder. Also, these networks are computationally intensive and slow. This paper proposes performing the learning and inference processes in the compressed domain in order to reduce the computational complexity and improve the speed of popular CNNs. For this purpose, a novel graph-based frequency channel selection method is proposed to identify and select the most important frequency channels. The computational complexity is reduced by retaining the important frequency components and discarding the insignificant ones as well as eliminating the unnecessary layers of the network. Experimental results show that the modified ResNet-50 operating in the compressed domain is up to 70% faster than the spatial-based traditional ResNet-50 while resulting in similar classification accuracy. Moreover, this paper proposes a preprocessing step with partial encoding to improve the resilience to distortions caused by low-quality encoded images. Finally, we show that training a network with highly compressed data can achieve a good classification accuracy with up to 93% reduction in the storage requirements of the training data. [ABSTRACT FROM AUTHOR]

Due to the ubiquity of graph-structured data, Graph Neural Network (GNN) have been widely used in different tasks and domains and good results have been achieved in tasks such as node classification and link prediction. However, there are still many challenges in representation learning of heterogeneous networks. Existing graph neural network models are partly based on homogeneous graphs, which do not take into account the rich semantic information of nodes and edges due to their different types; And partly based on heterogeneous graphs, which require predefined meta-structures (include meta-paths and meta-graphs) and do not take into account the different effects of different meta-structures on node representation. In this paper, we propose the MS-GAN model, which consists of four parts: graph structure learner, graph structure expander, graph structure filter and graph structure parser. The graph structure learner automatically generates a graph structure consisting of useful meta-paths by selecting and combining the sub-adjacent matrices in the original graph using a 1 × 1 convolution. The graph structure expander further generates a graph structure containing meta-graphs by Hadamard product based on the previous step. The graph structure filterer filters out graph structures that are more effective for downstream classification tasks based on diversity. The graph structure parser assigns different weights to graph structures consisting of different meta-structures by a semantic hierarchical attention. Finally, through experiments on four datasets and meta-structure visualization analysis, it is shown that MS-GAN can automatically generate useful meta-structures and assign different weights to different meta-structures. • Semi-supervised learning based graph neural network. • Automatically generate meta-structures. • Assign weights to meta-structures. [ABSTRACT FROM AUTHOR]

Recent years have witnessed increasing interest in adversarial attacks on images, while adversarial video attacks have seldom been explored. In this paper, we propose a sparse adversarial attack strategy on videos (DeepSAVA). Our model aims to add a small human-imperceptible perturbation to the key frame of the input video to fool the classifiers. To carry out an effective attack that mirrors real-world scenarios, our algorithm integrates spatial transformation perturbations into the frame. Instead of using the l p norm to gauge the disparity between the perturbed frame and the original frame, we employ the structural similarity index (SSIM), which has been established as a more suitable metric for quantifying image alterations resulting from spatial perturbations. We employ a unified optimisation framework to combine spatial transformation with additive perturbation, thereby attaining a more potent attack. We design an effective and novel optimisation scheme that alternatively utilises Bayesian Optimisation (BO) to identify the most critical frame in a video and stochastic gradient descent (SGD) based optimisation to produce both additive and spatial-transformed perturbations. Doing so enables DeepSAVA to perform a very sparse attack on videos for maintaining human imperceptibility while still achieving state-of-the-art performance in terms of both attack success rate and adversarial transferability. Furthermore, built upon the strong perturbations produced by DeepSAVA, we design a novel adversarial training framework to improve the robustness of video classification models. Our intensive experiments on various types of deep neural networks and video datasets confirm the superiority of DeepSAVA in terms of attacking performance and efficiency. When compared to the baseline techniques, DeepSAVA exhibits the highest level of performance in generating adversarial videos for three distinct video classifiers. Remarkably, it achieves an impressive fooling rate ranging from 99.5% to 100% for the I3D model, with the perturbation of just a single frame. Additionally, DeepSAVA demonstrates favourable transferability across various time series models. The proposed adversarial training strategy is also empirically demonstrated with better performance on training robust video classifiers compared with the state-of-the-art adversarial training with projected gradient descent (PGD) adversary. • Sparse attacks on video models: perturb fewer frames to gain high fooling rate. • Combining additive and spatial perturbations to enhance attacking performance. • Using SSIM instead of l p -norm to maintain the human perception. • Applying Bayesian Optimisation to identify the most critical frame to perturb. • A new adversarial training method based on combination of diverse perturbations. [ABSTRACT FROM AUTHOR]

Recently, leveraging deep neural networks for automated colorectal polyp segmentation has emerged as a hot topic due to the favored advantages in evading the limitations of visual inspection, e.g., overwork and subjectivity. However, most existing methods do not pay enough attention to the uncertain areas of colonoscopy images and often provide unsatisfactory segmentation performance. In this paper, we propose a novel boundary uncertainty aware network (BUNet) for precise and robust colorectal polyp segmentation. Specifically, considering that polyps vary greatly in size and shape, we first adopt a pyramid vision transformer encoder to learn multi-scale feature representations. Then, a simple yet effective boundary exploration module (BEM) is proposed to explore boundary cues from the low-level features. To make the network focus on the ambiguous area where the prediction score is biased to neither the foreground nor the background, we further introduce a boundary uncertainty aware module (BUM) that explores error-prone regions from the high-level features with the assistance of boundary cues provided by the BEM. Through the top-down hybrid deep supervision, our BUNet implements coarse-to-fine polyp segmentation and finally localizes polyp regions precisely. Extensive experiments on five public datasets show that BUNet is superior to thirteen competing methods in terms of both effectiveness and generalization ability. • A boundary uncertainty aware network is proposed for accurate polyp segmentation. • A boundary exploration module is proposed to explore boundary cues of polyps. • A boundary uncertainty aware module is proposed to seek error-prone regions. • The proposed network achieves high performance on six public datasets. [ABSTRACT FROM AUTHOR]

Neural network ensembling is a common and robust way to increase model efficiency. In this paper, we propose a new neural network ensemble algorithm based on Audibert's empirical star algorithm. We provide optimal theoretical minimax bound on the excess squared risk. Additionally, we empirically study this algorithm on regression and classification tasks and compare it to most popular ensembling methods. [ABSTRACT FROM AUTHOR]

Supervised learning in deep neural networks is commonly performed using error backpropagation. However, the sequential propagation of errors during the backward pass limits its scalability and applicability to low-powered neuromorphic hardware. Therefore, there is growing interest in finding local alternatives to backpropagation. Recently proposed methods based on forward-mode automatic differentiation suffer from high variance in large deep neural networks, which affects convergence. In this paper, we propose the Forward Direct Feedback Alignment algorithm that combines Activity-Perturbed Forward Gradients with Direct Feedback Alignment and momentum. We provide both theoretical proofs and empirical evidence that our proposed method achieves lower variance than forward gradient techniques. In this way, our approach enables faster convergence and better performance when compared to other local alternatives to backpropagation and opens a new perspective for the development of online learning algorithms compatible with neuromorphic systems. • Forward Gradient methods suffer from high variance that hinders convergence. • Activity-Perturbed Forward Gradients can be used to learn derivatives as direct feedback connections. • Feedback learning acts as a momentum that reduces the gradient variance closer to backpropagation. • Feedback learning reduces the biasness of Direct Feedback Alignment. • Feedback learning enables learning in convolutional layers. [ABSTRACT FROM AUTHOR]

Sharpness aware minimization (SAM) optimizer has been extensively explored as it can generalize better for training deep neural networks via introducing extra perturbation steps to flatten the landscape of deep learning models. Integrating SAM with adaptive learning rate and momentum acceleration, dubbed AdaSAM, has already been explored empirically to train large-scale deep neural networks without theoretical guarantee due to the triple difficulties in analyzing the coupled perturbation step, adaptive learning rate and momentum step. In this paper, we try to analyze the convergence rate of AdaSAM in the stochastic non-convex setting. We theoretically show that AdaSAM admits a O (1 / b T) convergence rate, which achieves linear speedup property with respect to mini-batch size b. Specifically, to decouple the stochastic gradient steps with the adaptive learning rate and perturbed gradient, we introduce the delayed second-order momentum term to decompose them to make them independent while taking an expectation during the analysis. Then we bound them by showing the adaptive learning rate has a limited range, which makes our analysis feasible. To the best of our knowledge, we are the first to provide the non-trivial convergence rate of SAM with an adaptive learning rate and momentum acceleration. At last, we conduct several experiments on several NLP tasks and the synthetic task, which show that AdaSAM could achieve superior performance compared with SGD, AMSGrad, and SAM optimizers. • Adaptive SAM with momentum acceleration ensures automatic learning rate adjustment. • Adaptive SAM shows linear speedup, converging at O (1 / b T). • Delaying 2nd-order momentum separates grad from adaptive rate & perturbed gradients. [ABSTRACT FROM AUTHOR]

This paper examines the issue of almost periodic quasi-projective synchronization of delayed fractional-order quaternion-valued neural networks. First, using a direct method rather than decomposing the fractional quaternion-valued system into four equivalent fractional real-valued systems, using Banach's fixed point theorem, according to the basic properties of fractional calculus and some inequality methods, we obtain that there is a unique almost periodic solution for this class of neural network with some sufficient conditions. Next, by constructing a suitable Lyapunov functional, using the characteristic of the Mittag-Leffler function and the scaling idea of the inequality, the adequate conditions for the quasi-projective synchronization of the established model are derived, and the upper bound of the systematic error is estimated. Finally, further use Matlab is used to carry out two numerical simulations to prove the results of theoretical analysis. [ABSTRACT FROM AUTHOR]

Currently, through proposing discontinuous control strategies with the signum function and discussing separately short-term memory (STM) and long-term memory (LTM) of competitive artificial neural networks (ANNs), the fixed-time (FXT) synchronization of competitive ANNs has been explored. Note that the method of separate analysis usually leads to complicated theoretical derivation and synchronization conditions, and the signum function inevitably causes the chattering to reduce the performance of the control schemes. To try to solve these challenging problems, the FXT synchronization issue is concerned in this paper for competitive ANNs by establishing a theorem of FXT stability with switching type and developing continuous control schemes based on a kind of saturation functions. Firstly, different from the traditional method of studying separately STM and LTM of competitive ANNs, the models of STM and LTM are compressed into a high-dimensional system so as to reduce the complexity of theoretical analysis. Additionally, as an important theoretical preliminary, a FXT stability theorem with switching differential conditions is established and some high-precision estimates for the convergence time are explicitly presented by means of several special functions. To achieve FXT synchronization of the addressed competitive ANNs, a type of continuous pure power-law control scheme is developed via introducing the saturation function instead of the signum function, and some synchronization criteria are further derived by the established FXT stability theorem. These theoretical results are further illustrated lastly via a numerical example and are applied to image encryption. [ABSTRACT FROM AUTHOR]

Spiking Neural Networks (SNNs) have become one of the most prominent next-generation computational models owing to their biological plausibility, low power consumption, and the potential for neuromorphic hardware implementation. Among the various methods for obtaining available SNNs, converting Artificial Neural Networks (ANNs) into SNNs is the most cost-effective approach. The early challenges in ANN-to-SNN conversion work revolved around the susceptibility of converted SNNs to conversion errors. Some recent endeavors have attempted to mitigate these conversion errors by altering the original ANNs. Despite their ability to enhance the accuracy of SNNs, these methods lack generality and cannot be directly applied to convert the majority of existing ANNs. In this paper, we present a framework named DNISNM for converting ANN to SNN, with the aim of addressing conversion errors arising from differences in the discreteness and asynchrony of network transmission between ANN and SNN. The DNISNM consists of two mechanisms, Data-based Neuronal Initialization (DNI) and Signed Neuron with Memory (SNM), designed to respectively address errors stemming from discreteness and asynchrony disparities. This framework requires no additional modifications to the original ANN and can result in SNNs with improved accuracy performance, simultaneously ensuring universality, high precision, and low inference latency. We verify it experimentally on challenging object recognition datasets, including CIFAR10, CIFAR100, and ImageNet-1k. Experimental results show that the SNN converted by our framework has very high accuracy even at extremely low latency. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) have gained immense popularity in recent years, finding their utility in diverse fields such as image recognition, natural language processing, and bio-informatics. Despite the remarkable progress made in deep learning theory, most studies on CNNs, especially in regression tasks, tend to heavily rely on the least squares loss function. However, there are situations where such learning algorithms may not suffice, particularly in the presence of heavy-tailed noises or outliers. This predicament emphasizes the necessity of exploring alternative loss functions that can handle such scenarios more effectively, thereby unleashing the true potential of CNNs. In this paper, we investigate the generalization error of deep CNNs with the rectified linear unit (ReLU) activation function for robust regression problems within an information-theoretic learning framework. Our study demonstrates that when the regression function exhibits an additive ridge structure and the noise possesses a finite p th moment, the empirical risk minimization scheme, generated by the maximum correntropy criterion and deep CNNs, achieves fast convergence rates. Notably, these rates align with the mini-max optimal convergence rates attained by fully connected neural network model with the Huber loss function up to a logarithmic factor. Additionally, we further establish the convergence rates of deep CNNs under the maximum correntropy criterion when the regression function resides in a Sobolev space on the sphere. [ABSTRACT FROM AUTHOR]

Adversarial training has become the mainstream method to boost adversarial robustness of deep models. However, it often suffers from the trade-off dilemma, where the use of adversarial examples hurts the standard generalization of models on natural data. To study this phenomenon, we investigate it from the perspective of spatial attention. In brief, standard training typically encourages a model to conduct a comprehensive check to input space. But adversarial training often causes a model to overly concentrate on sparse spatial regions. This reduced tendency is beneficial to avoid adversarial accumulation but easily makes the model ignore abundant discriminative information, thereby resulting in weak generalization. To address this issue, this paper introduces an Attention-Enhanced Learning Framework (AELF) for robustness training. The main idea is to enable the model to inherit the attention pattern of standard pre-trained model through an embedding-level regularization. To be specific, given a teacher model built on natural examples, the embedding distribution of teacher model is used as a static constraint to regulate the embedding outputs of the objective model. This design is mainly supported with that the embedding feature of standard model is usually recognized as a rich semantic integration of input. For implementation, we present a simplified AELFs that can achieve the regularization with single cross entropy loss via the parameter initialization and parameter update strategy. This avoids the extra consistency comparison operation between embedding vectors. Experimental observations verify the rationality of our argument, and experimental results demonstrate that it can achieve remarkable improvements in generalization under the high-level robustness. [ABSTRACT FROM AUTHOR]

Deep neural networks tend to suffer from the overfitting issue when the training data are not enough. In this paper, we introduce two metrics from the intra-class distribution of correct-predicted and incorrect-predicted samples to provide a new perspective on the overfitting issue. Based on it, we propose a knowledge distillation approach without pretraining a teacher model in advance named Tolerant Self-Distillation (TSD) for alleviating the overfitting issue. It introduces an online updating memory and selectively stores the class predictions of the samples from the past iterations, making it possible to distill knowledge across the iterations. Specifically, the class predictions stored in the memory bank serve as the soft labels for supervising the samples from the same class for the current iteration in a reverse way, i.e. the correct-predicted samples are supervised with the incorrect predictions while the incorrect-predicted samples are supervised with the correct predictions. Consequently, the premature convergence issue caused by the over-confident samples would be mitigated, which helps the model to converge to a better local optimum. Extensive experimental results on several image classification benchmarks, including small-scale, large-scale, and fine-grained datasets, demonstrate the superiority of the proposed TSD. [ABSTRACT FROM AUTHOR]

The dreaming Hopfield model constitutes a generalization of the Hebbian paradigm for neural networks, that is able to perform on-line learning when "awake" and also to account for off-line "sleeping" mechanisms. The latter have been shown to enhance storing in such a way that, in the long sleep-time limit, this model can reach the maximal storage capacity achievable by networks equipped with symmetric pairwise interactions. In this paper, we inspect the minimal amount of information that must be supplied to such a network to guarantee a successful generalization, and we test it both on random synthetic and on standard structured datasets (i. e. , MNIST, Fashion-MNIST and Olivetti). By comparing these minimal thresholds of information with those required by the standard (i. e. , always "awake") Hopfield model, we prove that the present network can save up to ∼ 90 % of the dataset size, yet preserving the same performance of the standard counterpart. This suggests that sleep may play a pivotal role in explaining the gap between the large volumes of data required to train artificial neural networks and the relatively small volumes needed by their biological counterparts. Further, we prove that the model Cost function (typically used in statistical mechanics) admits a representation in terms of a standard Loss function (typically used in machine learning) and this allows us to analyze its emergent computational skills both theoretically and computationally: a quantitative picture of its capabilities as a function of its control parameters is achieved and consistency between the two approaches is highlighted. The resulting network is an associative memory for pattern recognition tasks that learns from examples on-line, generalizes correctly (in suitable regions of its control parameters) and optimizes its storage capacity by off-line sleeping: such a reduction of the training cost can be inspiring toward sustainable AI and in situations where data are relatively sparse. [ABSTRACT FROM AUTHOR]

Spiking Neural Networks (SNN) are recognised as well-suited for processing spatiotemporal information with ultra-low energy consumption. However, proposals based on SNN for classification tasks are more common than for forecasting problems. In this sense, this paper presents a new general training methodology for univariate time-series forecasting based on SNN. The methodology is focused on one-step ahead forecasting problems and combines a PulseWidth Modulation based encoding–decoding algorithm with a Surrogate Gradient method as supervised training algorithm. In order to validate the generality of the presented methodology sine-wave, 3 UCI and 1 available real-world datasets are used. The results show very satisfactory forecasting results (M A E ∈ [ 0. 0094 , 0. 2891 ]) regardless of the characteristics of the dataset or the application field. In addition, weights can be initialised just once to achieve robust results, boosting the advantages of computational and energy cost of SNN. • New supervised training methodology for univariate time-series forecasting with SNN. • A PWM based encoding–decoding algorithm and a Surrogate Gradient method are combined. • The methodology is characterised by ultra-low latency and high robustness. • 3 UCI datasets and air pollution data for Greater London Area are used for validation. • Satisfactory forecasting results regardless of the characteristics of the dataset. [ABSTRACT FROM AUTHOR]

Spiking neural networks (SNNs) are brain-inspired models that utilize discrete and sparse spikes to transmit information, thus having the property of energy efficiency. Recent advances in learning algorithms have greatly improved SNN performance due to the automation of feature engineering. While the choice of neural architecture plays a significant role in deep learning, the current SNN architectures are mainly designed manually, which is a time-consuming and error-prone process. In this paper, we propose a spiking neural architecture search (NAS) method that can automatically find efficient SNNs. To tackle the challenge of long search time faced by SNNs when utilizing NAS, the proposed NAS encodes candidate architectures in a branchless spiking supernet which significantly reduces the computation requirements in the search process. Considering that real-world tasks prefer efficient networks with optimal accuracy under a limited computational budget, we propose a Synaptic Operation (SynOps)-aware optimization to automatically find the computationally efficient subspace of the supernet. Experimental results show that, in less search time, our proposed NAS can find SNNs with higher accuracy and lower computational cost than state-of-the-art SNNs. We also conduct experiments to validate the search process and the trade-off between accuracy and computational cost. [ABSTRACT FROM AUTHOR]

Multi-view deep neural networks have shown excellent performance on 3D shape classification tasks. However, global features aggregated from multiple views data often lack content information and spatial relationship, which leads to difficult identification the small variance among subcategories in the same category. To solve this problem, in this paper, a novel multiscale dilated convolution neural network termed as MSDCNN is proposed for multi-view fine-grained 3D shape classification. Firstly, a sequence of views are rendered from 12-viewpoints around the input 3D shape by the sequential view capturing module. Then, the first 22 convolution layers of ResNeXt50 is employed to extract the semantic features of each view, and a global mixed feature map is obtained through the element-wise maximum operation of the 12 output feature maps. Furthermore, attention dilated module (ADM), which combines four concatenated attention dilated block (ADB), is designed to extract larger receptive field features from global mixed feature map to enhance context information among the views. Specifically, each ADB is consisted by an attention mechanism module and a dilated convolution with different dilation rates. In addition, prediction module with label smoothing is proposed to classify features, which contains 3 × 3 convolution and adaptive average pooling. The performance of our method is validated experimentally on the ModelNet10, ModelNet40 and FG3D datasets. Experimental results demonstrate the effectiveness and superiority of the proposed MSDCNN framework for 3D shape fine-grained classification. [ABSTRACT FROM AUTHOR]

An increasing need of running Convolutional Neural Network (CNN) models on mobile devices encourages the studies on efficient and lightweight neural network model. In this paper, an Inverse Residual Multi-Branch Network named IremulbNet is proposed to solve the problem of insufficient classification accuracy in existing lightweight network models. The core module of this model is to reconstruct an inverse residual structure, in which a special feature fusion method, multi-branch feature extraction, and depthwise separable convolution techniques are used to improve the classification accuracy. The performance of model is tested using image databases. Experimental results show that for the fine-grained image dataset Imagenet-woof, IremulbNet achieved 10.9%, 12.2%, and 15.3% higher accuracy than that of MobileNet V3, ShuffleNet V2, and PeleeNet, respectively. Moreover, it can reduce inference time (GPU) about 42.09% and 75.56% compared to classic ResNet50 and DenseNet121. • A lightweight IremulbNet is proposed to achieve high accuracy image recognition with a small number of parameters and operations. • A special feature fusion method, multi-branch feature extraction, and depthwise separable convolution are used to improve the performance of the network. • A significant improvement in recognition accuracy can be obtained in image classification tasks with low computation load. [ABSTRACT FROM AUTHOR]

Since the physical meaning of the fields of the dataset is unknown, we have to use the feature interaction method to select the correlated features and exclude uncorrelated features. The current state-of-the-art methods employ various methods based on feature interaction to predict advertisement Click-Through Rate (CTR); however, the feature interaction based on potential new feature mining is rarely considered, which can provide effective assistance for feature interaction. This motivates us to investigate methods that combine potential new features and feature interactions. Thus, we propose a potential feature excitation learning network ( PeNet ), which is a neural network model based on feature combination and feature interaction. In PeNet , we treat the row compression and column compression of the original feature matrix as potential new features, and proposed the excitation learning mechanism that is a weighted mechanism based on residual principle. Through this excitation learning mechanism, the original embedded features and potential new features are subjected to weighted interaction based on the residual principle. Moreover, a deep neural network is exploited to iteratively learn and iteratively combine features. The excitation learning structure of PeNet neural network is well demonstrated in this paper, that is, the control flow of embedding, compression, excitation and output, which further strengthens the correlated features and weakens the uncorrelated features by compressing and expanding the features. Experimental results on multiple benchmark datasets indicate the PeNet as a general-purpose plug-in has more superior performance and better efficiency than previous state-of-the-art methods. • Combine the row compression and column compression and treat as new features. • Propose a feature excitation learning mechanism to guide the new features. • Propose [ABSTRACT FROM AUTHOR]

Spiking Neural Networks (SNNs) have been considered a potential competitor to Artificial Neural Networks (ANNs) due to their high biological plausibility and energy efficiency. However, the architecture design of SNN has not been well studied. Previous studies either use ANN architectures or directly search for SNN architectures under a highly constrained search space. In this paper, we aim to introduce much more complex connection topologies to SNNs to further exploit the potential of SNN architectures. To this end, we propose the topology-aware search space, which is the first search space that enables a more diverse and flexible design for both the spatial and temporal topology of the SNN architecture. Then, to efficiently obtain architecture from our search space, we propose the spatio-temporal topology sampling (STTS) algorithm. By leveraging the benefits of random sampling, STTS can yield powerful architecture without the need for an exhaustive search process, making it significantly more efficient than alternative search strategies. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate the effectiveness of our method. Notably, we obtain 70.79% top-1 accuracy on ImageNet with only 4 time steps, 1.79% higher than the second best model. Our code is available under https://github.com/stiger1000/Random-Sampling-SNN. • The SNN architectures significantly benefit from much more complex connection topologies. • Incorporating synaptic delay provides a novel perspective to the design of SNN architectures. • The novel sampling method greatly accelerates the process of obtaining SNN architectures. [ABSTRACT FROM AUTHOR]

Although great success has been achieved in various computer vision tasks, deep neural networks (DNNs) suffer dramatic performance degradation when evaluated on out-of-distribution data. Domain generalization (DG) is proposed to handle this problem by learning domain-agnostic information from multiple source domains to generalize well on unseen target domains. Several methods resort to Fourier transform due to its simplicity and efficiency. They argue that amplitude spectra imply domain-specific information, which should be suppressed, while phase counterparts imply domain-agnostic information, which should be preserved. However, these methods only suppress the domain-specific information in source domains and neglect the relationship with target domains, leading to the persistence of the domain gap. Besides, these methods preserve domain-agnostic information by keeping phase components unchanged, causing them to be underutilized. In this paper, we propose Dual Branch Augmentation Module (DBAM) by leveraging Fourier transform and taking advantage of both amplitude and phase spectra. For the amplitude branch, we propose Inner-domain Amplitude Distribution Rectification (IADR) and Cross-domain Amplitude Dirichlet Mixup (CADM) to stabilize the training process and explore more feature space. In addition, we propose Test-time Amplitude Prototype Calibration (TAPC) to construct the connection between source and target domains during evaluation to further mitigate the domain gap. For the phase branch, we propose Random Symmetric Phase Perturbation (RSPP) to enhance the robustness for recognizing domain-agnostic information. With the combined contributions of the two branches, DBAM significantly surpasses other state-of-the-art (SOTA) methods. Extensive experiments on four benchmarks and further analysis demonstrate the effectiveness of DBAM. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNNs) are vulnerable to the attacks of adversarial examples, which bring serious security risks to the learning systems. In this paper, we propose a new defense method to improve the adversarial robustness of DNNs based on stochastic neural networks (SNNs), termed as Margin-SNN. The proposed Margin-SNN mainly includes two modules, i.e., feature uncertainty learning module and label embedding module. The first module introduces uncertainty to the latent feature space by giving each sample a distributional representation rather than a fixed point representation, and leverages the advantages of variational information bottleneck method in achieving good intra-class compactness in latent space. The second module develops a label embedding mechanism to take advantage of the semantic information underlying the labels, which maps the labels into the same latent space with the features, in order to capture the similarity between sample and its class centroid, where a penalty term is equipped to elegantly enlarge the margin between different classes for better inter-class separability. Since no adversarial information is introduced, the proposed model can be learned in standard training to improve adversarial robustness, which is much more efficient than adversarial training. Extensive experiments on data sets MNIST, FASHION MNIST, CIFAR10, CIFAR100 and SVHN demonstrate superior defensive ability of the proposed method. Our code is available at https://github.com/humeng24/Margin-SNN. [ABSTRACT FROM AUTHOR]

Many large and complex deep neural networks have been shown to provide higher performance on various computer vision tasks. However, very little is known about the relationship between the complexity of the input data along with the type of noise and the depth needed for correct classification. Existing studies do not address the issue of common corruptions adequately, especially in understanding what impact these corruptions leave on the individual part of a deep neural network. Therefore, we can safely assume that the classification (or misclassification) might be happening at a particular layer(s) of a network that accumulates to draw a final correct or incorrect prediction. In this paper, we introduce a novel concept of corruption depth , which identifies the location of the network layer/depth until the misclassification persists. We assert that the identification of such layers will help in better designing the network by pruning certain layers in comparison to the purification of the entire network which is computationally heavy. Through our extensive experiments, we present a coherent study to understand the processing of examples through the network. Our approach also illustrates different philosophies of example memorization and a one-dimensional view of sample or query difficulty. We believe that the understanding of the corruption depth can open a new dimension of model explainability and model compression , where in place of just visualizing the attention map, the classification progress can be seen throughout the network. • A novel metric corruption depth is proposed to explain functioning of neural network. • The proposed research helps in deciphering the flow of classification decisions. • The proposed research overcomes the limitations of understanding example difficulty. • Corruption depth notation helps in pruning the networks while improving the accuracy. [ABSTRACT FROM AUTHOR]