Search

• A linearly combined form of the objective is reformulated to control the trade-off between the return and the divergence more effectively. • An improved proximal policy optimization method (i.e., alphaPPO) is proposed, with a more elaborative alpha divergence for two adjacent policies. • The effectiveness of our alphaPPO is validated using detailed experimental comparison and analysis for six benchmark environments. Proximal policy optimization (PPO) is a recent advancement in reinforcement learning, which is formulated as an unconstrained optimization problem including two terms: accumulative discount return and Kullback–Leibler (KL) divergence. Currently, there are three PPO versions: primary, adaptive, and clipping. The most widely used PPO algorithm is the clipping version, in which the KL divergence is replaced by a clipping function to measure the difference between two policies indirectly. In this paper, we revisit this primary PPO and improve it in two aspects. One is to reformulate it as a linearly combined form to control the trade-off between two terms. The other is to substitute a parametric alpha divergence for KL divergence to measure the difference of two policies more effectively. This novel PPO variant is referred to as alphaPPO in this paper. Experiments on six benchmark environments verify the effectiveness of our alphaPPO, compared with clipping and combined PPOs. [ABSTRACT FROM AUTHOR]

Stochastic optimization acceleration is extremely significant and challenging for deep neural networks (DNNs). In recent years, several novel proportional-integral–differential-based (PID-based) optimizers have been proposed to speed up the optimization by alleviating the oscillation behavior of stochastic gradient descent with momentum (SGD-M), yet lacked theoretical analysis. Along this line of research, this paper adopts dynamical system theory to design a new hybrid optimizer and present theoretical analysis. Firstly, it is found that DNN optimization is equivalent to a discrete time dynamical system. Building upon the equivalence, high order augmented dynamical system viewpoint is utilized to design a PI-like optimizer for ensuring high accuracy, which is more stable than SGD-M. Then, hybrid dynamical system viewpoint is employed to improve the PI-like optimizer as a new hybrid form for suppressing oscillation and accelerating optimization. Lyapunov method, Taylor series, matrix theory and equilibrium are combined to theoretically investigate the convergence and the oscillation of loss function, showing that the proposed hybrid optimizer can alleviate oscillation, boost optimization speed, and maintain high accuracy. In theoretical analyses, explicit conditions of hyper-parameters that guarantee training stability are calculated and presented, practically guiding the adjustment of hyper-parameters and promoting the application of hybrid optimizer. Experiments are presented on three commonly used benchmark datasets, i.e., MNIST, CIFAR10 and CIFAR100, demonstrating that the hybrid optimizer obtains up to 42% acceleration with competitive accuracy relative to state-of-the-art optimizers. In short, this paper not only presents a new hybrid optimizer for accelerating optimization, but also provides a novel, theoretical and systematic perspective to find and analyze new optimizer for DNNs. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNNs) have been shown to be vulnerable to backdoor attacks during training. Most of the existing backdoor defense methods are designed for specific types of backdoor attacks, and the work of detecting backdoors and mitigating backdoors is mostly separate. Currently, few general and complete defense frameworks have been developed. In this paper, we propose a lightweight, general, and complete defense framework against three main types of backdoor attacks. It can efficiently detect poisoned images and remove trigger patterns on poisoned images without costly retraining of the backdoor model. First, we use the feature difference between clean samples and poisoned samples in the middle layer of the model to distinguish them. Then, we remove the backdoor using image inpainting algorithm to remove the backdoor triggering pattern on the poisoned samples. We deploy three of the most popular backdoor attacks on three datasets to test the effectiveness of our defenses. Extensive experimental results show that our method can effectively defend against various backdoor attacks with a relatively small cost. In particular, we reduce the attack success rate of the more stealthy clean-label poisoning attack from 94.9% to 0.02% with little impact on the classification accuracy of the inpainted images. [ABSTRACT FROM AUTHOR]

[Display omitted] With the widespread applications of Deep Neural Networks (DNNs), the safety of DNNs has become a significant issue. The vulnerability of the neural networks against adversarial examples deepens concerns about the safety of DNNs applications. This paper proposed a novel defence method to improve the adversarial robustness of DNN classifiers without using adversarial training. This method introduces two new loss functions. First, a zero-cross-entropy loss is used to punish overconfidence and find the appropriate confidence for different instances. Second, a logit balancing loss is proposed to protect DNNs from non-targeted attacks by regularising incorrect classes' logits distribution. This method achieved competitive adversarial robustness compared to advanced adversarial training methods. Meanwhile, a novel robustness diagram is proposed to analyse, interpret and visualise the robustness of DNN classifiers against adversarial attacks. Furthermore, a Log-Softmax-pattern-based adversarial attack detection method is proposed. This detection method can distinguish clean inputs and multiple adversarial attacks via one multi-classification MLP. In particular, it is state-of-the-art in identifying white-box gradient-based attacks; it achieved at least 95.5% accuracy for classifying four white-box gradient-based attacks with maximum 0.1% false positive ratio. [ABSTRACT FROM AUTHOR]

• New subspace-based method for image set recognition. • Theoretically-grounded integration of subspace representation into DNNs while keeping end-to-end trainability. • Our new method generalizes the classic learning subspace method. This paper addresses the problem of object recognition given a set of images as input (e.g., multiple camera sources and video frames). Convolutional neural network (CNN)-based frameworks do not exploit these sets effectively, processing a pattern as observed, not capturing the underlying feature distribution as it does not consider the variance of images in the set. To address this issue, we propose the Grassmannian learning mutual subspace method (G-LMSM), a NN layer embedded on top of CNNs that can process image sets more effectively and can be trained in an end-to-end manner. The image set is first represented by a low-dimensional input subspace and then this input subspace is matched with dictionary subspaces by a similarity of their canonical angles, an interpretable and easy to compute metric. The key idea of G-LMSM is that the dictionary subspaces are learned as points on the Grassmann manifold, optimized with Riemannian stochastic gradient descent. This learning is stable, efficient and theoretically well-grounded. We demonstrate the effectiveness of our proposed method on hand shape recognition, face identification, and facial emotion recognition. [ABSTRACT FROM AUTHOR]

A substantial amount of research has been carried out in Explainable Artificial Intelligence (XAI) models, especially in those which explain the deep architectures of neural networks. A number of XAI approaches have been proposed to achieve trust in Artificial Intelligence (AI) models as well as provide explainability of specific decisions made within these models. Among these approaches, global interpretation methods have emerged as the prominent methods of explainability because they have the strength to explain every feature and the structure of the model. This survey attempts to provide a comprehensive review of global interpretation methods that completely explain the behaviour of the AI models. We present a taxonomy of the available global interpretations models and systematically highlight the critical features and algorithms that differentiate them from local as well as hybrid models of explainability. Through examples and case studies from the literature, we evaluate the strengths and weaknesses of the global interpretation models and assess challenges when these methods are put into practice. We conclude the paper by providing the future directions of research in how the existing challenges in global interpretation methods could be addressed and what values and opportunities could be realized by the resolution of these challenges. [ABSTRACT FROM AUTHOR]

Deep neural networks provide end-to-end tools to learn effective representations from data directly. The deep structure makes it possible to model a complicated pattern, even if it has a variety of changes. This leads to a problem that noises and outliers are usually treated as a specific pattern, which is also learned in the network. It is one reason for overfitting incapable of being addressed sufficiently in deep networks. This paper proposes a new method called subdomain contraction (SDC) to tackle the problem. The idea is that our approach inclines to learn more about the shared features between the subsets of the samples but less about the specific features found in only one or two subsets. To this end, the SDC loss penalizes the distribution distance between sub-domains in the feature space to constrain the sharing level of features. By applying the SDC loss term, the data drive the learning process to an optimal tradeoff between modeling noises and the varieties of the pattern. In this manner, the SDC models the pattern as much as possible and ignores most noises, thus improving the generalization ability. The SDC loss can be efficiently computed in minibatches and can also work collaboratively with other regularization methods such as dropout to further improve the performance. Extensive experiments demonstrate that SDC can improve the effectiveness and robustness of representation learning in deep networks against noises, and the superiority is most remarkable with noisy data. [ABSTRACT FROM AUTHOR]

Neural network pruning is a critical field aimed at reducing the infrastructure costs of neural networks by removing parameters. Traditional methods follow a fixed paradigm including pretraining, pruning, and fine-tuning. Despite the close relationship among these three stages, most pruning methods treat them as independent processes. In this paper, we propose a novel two-stage pruning method, which includes pretraining a network that is instructive for subsequent pruning, and a unified optimization model that integrates pruning and fine-tuning. Specifically, in the first stage, we design a group sparse regularized model for pretraining. This model not only safeguards the network from irreversible damage but also offers valuable insights for the pruning process. In the second stage, we introduce an element-wise sparse regularization into pruning model. This model enables us to pinpoint sparse weights more precisely than pretrained network. It automatically derives effective pruning criteria, and omits the step of fine-tuning. To implement the two-stage process in practice, we utilize stochastic gradient algorithm for the pretraining and design a threshold algorithm for pruning stage. Extensive experiments confirm the competitive performance of our proposed method in terms of both accuracy and memory cost when compared to various benchmarks. Furthermore, ablation experiments validate the effectiveness of the proposed pretraining model's guidance for the pruning process. • We propose SOGP to enhance the connection between pretraining and pruning. • We propose a novel non-convex group sparse regularization GT l 1 . • We construct a unified optimization model to integrate pruning and fine-tuning. • Extensive experiments validate the good pruning and accuracy performance of SOGP. [ABSTRACT FROM AUTHOR]

Model pruning is a commonly used technique for compressing DNNs and reducing computation requirements to accelerate inference. However, the required granularity of pruning varies across different application scenarios, making it difficult and cumbersome to customize different pruning methods for each hardware or computing platform. Therefore, a unified framework is necessary to accommodate various levels of granularity of pruning. Furthermore, some available methods require additional fine-tuning or model retraining to restore accuracy, which can result in significant time costs. With this motivation, this paper proposes a Multi-Granularity Pruning Framework, namely MGPF, to obtain sparse models of different granularity without fine-tuning the remaining connections. Specifically, a deformable soft mask is introduced as the pruning initiator to achieve different levels of pruning granularity, such as weight pruning, channel pruning, and filter pruning, etc. The model parameters and soft masks are jointly trained, and we just apply L 1 regularization on soft masks for sparsity to ensure that the model can be repaired during training without fine-tuning or retraining. After pruning, the soft masks are absorbed into the model parameters in the form of element product without changing the model structures. Experimental results on three image classification benchmarks CIFAR-10/100 and ImageNet-1K demonstrate the effectiveness of our method for various CNN architectures, datasets, and pruning rates. Particularly, for ResNet-50 on ImageNet-1K, we achieve a higher accuracy under the pruning rate of 98% for unstructured pruning which leads the advanced method by 12%. [ABSTRACT FROM AUTHOR]

• It is demonstrated that deep networks can generate multiple local optima during training process which can be combined to form a stronger model. • An efficient and easy-to-implement ensemble learning approach designed for deep networks is introduced. • The approach improves the performance of neural networks without additional training cost. • The approach can be directly applied to all kinds of neural networks without tuning the network architecture. • Comparison experiments with a range of ensemble strategies have shown that the simplest ensemble strategy performs best. The present work is motivated by the desire to find an efficient approach that can improve the performance of deep neural networks in a general sense. To this end, an easy-to-implement ensemble approach is proposed in this paper leveraging the 'local sub-optima' of deep networks, which is referred as to history-state ensemble (HSE) method. We demonstrated that neural networks can naturally generate multiple 'local sub-optima' with diversity during training process, and their combination can effectively improve the accuracy and stability of the single network. The merits of HSE are twofold: (1) It does not require additional training cost in order to acquire multiple base models, which is one of the main drawbacks limiting the generalization of ensemble techniques in deep learning. (2) It can be easily applied to any types of deep networks without tuning of network architectures. We proposed the simplest way to perform HSE and investigated more than 20 ensemble strategies for HSE as comparison. Experiments are conducted on six datasets and eight popular network architectures for the case of rotating machinery fault diagnosis. It is demonstrated that the stability and accuracy of neural networks can be generally improved through the simplest ensemble strategy proposed in this paper. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNNs) are currently being deployed as machine learning technology in a wide range of important real-world applications. DNNs consist of a huge number of parameters that require millions of floating-point operations (FLOPs) to be executed both in learning and prediction modes. A more effective method is to implement DNNs in a cloud computing system equipped with centralized servers and data storage sub-systems with high-speed and high-performance computing capabilities. This paper presents an up-to-date survey on current state-of-the-art deployed DNNs for cloud computing. Various DNN complexities associated with different architectures are presented and discussed alongside the necessities of using cloud computing. We also present an extensive overview of different cloud computing platforms for the deployment of DNNs and discuss them in detail. Moreover, DNN applications already deployed in cloud computing systems are reviewed to demonstrate the advantages of using cloud computing for DNNs. The paper emphasizes the challenges of deploying DNNs in cloud computing systems and provides guidance on enhancing current and new deployments. [ABSTRACT FROM AUTHOR]

Inference centers need more data to have a more comprehensive and beneficial learning model, and for this purpose, they need to collect data from data providers. On the other hand, data providers are cautious about delivering their datasets to inference centers in terms of privacy considerations. In this paper, by modifying the structure of the autoencoder, we present a method that manages the utility-privacy trade-off well. To be more precise, the data is first compressed using the encoder, then confidential and non-confidential features are separated and uncorrelated using the classifier. The confidential feature is appropriately combined with noise, and the non-confidential feature is enhanced, and at the end, data with the original data format is produced by the decoder. The suggested architecture additionally enables data providers to modify the degree of privacy needed for private features and the level of utility for non-private features. The proposed method has been examined for both image and categorical databases, and the results show a significant performance improvement compared to previous methods. [ABSTRACT FROM AUTHOR]

With the increasing practicality of deep learning applications, practitioners are inevitably faced with datasets corrupted by noise from various sources such as measurement errors, mislabeling, and estimated surrogate inputs/outputs that can adversely impact the optimization results. It is a common practice to improve the optimization algorithm's robustness to noise, since this algorithm is ultimately in charge of updating the network parameters. Previous studies revealed that the first-order moment used in Adam-like stochastic gradient descent optimizers can be modified based on the Student's t-distribution. While this modification led to noise-resistant updates, the other associated statistics remained unchanged, resulting in inconsistencies in the assumed models. In this paper, we propose AdaTerm, a novel approach that incorporates the Student's t-distribution to derive not only the first-order moment but also all the associated statistics. This provides a unified treatment of the optimization process, offering a comprehensive framework under the statistical model of the t-distribution for the first time. The proposed approach offers several advantages over previously proposed approaches, including reduced hyperparameters and improved robustness and adaptability. AdaTerm achieves this by considering the interdependence of gradient dimensions. In particular, upon detection, AdaTerm excludes aberrant gradients from the update process and enhances its robustness for subsequent updates. Conversely, it performs normal parameter updates when the gradients are statistically valid, allowing for flexibility in adapting its robustness. This noise-adaptive behavior contributes to AdaTerm's exceptional learning performance, as demonstrated through various optimization problems with different and/or unknown noise ratios. Furthermore, we introduce a new technique for deriving a theoretical regret bound without relying on AMSGrad, providing a valuable contribution to the field. [ABSTRACT FROM AUTHOR]

Recent developments in multimodal training methodologies, including CLIP and ALIGN, obviate the necessity for individual data labeling. These approaches utilize pairs of data and corresponding textual information found online as a form of weak supervision signal. However, models employing this kind of weak supervision are not as competitive as their supervised and semi-supervised counterparts when sufficient labeled data is accessible. This performance gap constrains the applicability of weekly supervised models. In this paper, we narrow the gap by proposing a combined scaling method, named BASIC, that achieves 85.7% top-1 accuracy on the ImageNet ILSVRC-2012 validation set without learning from any labeled ImageNet example. This accuracy surpasses best-published similar models, CLIP and ALIGN, by 9.3%. Our BASIC model also shows significant improvements in robustness benchmarks. For instance, on 5 test sets with natural distribution shifts such as ImageNet-{A,R,V2,Sketch} and ObjectNet, our model achieves 84.3% top-1 average accuracy, only a small drop from its original ImageNet accuracy. To achieve these results, we first develop a theoretical framework which shows that larger contrastive batch sizes lead to smaller generalization gaps for image-text models such as CLIP and ALIGN. Based on this theoretical result, we scale up the contrastive learning framework of CLIP and ALIGN in three dimensions (data size, model size, and batch size) by proposing a new method using gradient checkpointing and model parallelism. As a result, our dataset has 6.6B noisy image-text pairs, which is 4x larger than ALIGN, and 16x larger than CLIP. Our largest model has 3B weights, which is 3.75x larger in parameters and 8x larger in FLOPs than ALIGN and CLIP. Finally, our batch size is 65536 which is 2x more than CLIP and 4x more than ALIGN. [ABSTRACT FROM AUTHOR]

With the development of artificial intelligence, the most representative deep learning has been applied to various fields, which is greatly influencing human society. However, deep neural networks (DNNs) are still a black-box model, and the process how they make decisions internally is still difficult to understand and control. At the same time, DNNs take up more hardware resources, resulting in high energy consumption. Therefore, it is significant to study the characteristics of deep AI models and deeply understand the interactions between parameters within AI models so as to improve the interpretability of DNNs, optimize their structure and increase their computational efficiency. In this paper, we propose a graph-based interpretability method for deep neural networks (GIMDNN). The running parameters of DNNs are modeled as a graph by using a kernel function or the Graph Transformer Networks (GTN), where the nodes of the graph are obtained by dimensional mapping of the parameters of the DNNs, and the edges are calculated by the Gaussian kernel function. The generated graphs are classified by a graph convolutional network (GCN). The association relationship between the adjacent layers and the running mechanism of DNNs are analyzed, and the importance of the parameters of each layer in the DNNs for the final classification result can be obtained. Convolutional neural networks (CNNs) are one of the most representative models in DNNs. The proposed method is experimentally evaluated on the CNNs. The experimental results show that the proposed method can interpret the associations among the weight parameters as well as the correlation between two adjacent layers. Therefore, the DNNs for special tasks, such as portable applications, edge computing, and so on, can be customized, the number of parameters can be reduced. It is valuable to interpret the operation and principle of CNNs. [ABSTRACT FROM AUTHOR]

Video temporal grounding is a challenging task in computer vision that involves localizing a video segment semantically related to a given query from a set of videos and queries. In this paper, we propose a novel weakly-supervised model called the Multi-level Attentional Reconstruction Networks (MARN), which is trained on video-sentence pairs. During the training phase, we leverage the idea of attentional reconstruction to train an attention map that can reconstruct the given query. At inference time, proposals are ranked based on attention scores to localize the most suitable segment. In contrast to previous methods, MARN effectively aligns video-level supervision and proposal scoring, thereby reducing the training-inference discrepancy. In addition, we incorporate a multi-level framework that encompasses both proposal-level and clip-level processes. The proposal-level process generates and scores variable-length time sequences, while the clip-level process generates and scores fix-length time sequences to refine the predicted scores of the proposal in both training and testing. To improve the feature representation of the video, we propose a novel representation mechanism that utilizes intra-proposal information and adopts 2D convolution to extract inter-proposal clues for learning reliable attention maps. By accurately representing these proposals, we can better align them with the textual modalities, and thus facilitate the learning of the model. Our proposed MARN is evaluated on two benchmark datasets, and extensive experiments demonstrate its superiority over existing methods. [Display omitted] • We introduce a novel framework for weakly-supervised video temporal grounding. • We introduce a multi-level mechanism for constraining and fine-tuning predictions. • We propose an attention module to reduce training-inference discrepancy. • The conducted extensive experiments show MARN's superiority. [ABSTRACT FROM AUTHOR]

• We are the first to formulate the interpretation of deep models as a bandit problem. • Our statistical perturbations retain regional interaction without any prior involved. • The Upper Confidence Bound guarantees a fair selection of critical image features. • Our method provides more precise explanations with a smaller area. Interpretability of black-box deep models is yet challenging because existing model-agnostic methods mainly locally explain the behavior of the classifier by learning a linear proxy around the instance being predicted. The explanation can be faithful locally, but may not be accurate globally. In this paper, we for the first time formulate the interpretation of classifiers as a bandit problem and introduce a Bandit Interpretation method via Confidence Selection (BICS). We statistically impose disturbances on different arms (image regions) and examine non-linear changes of the model's output to fairly select important regions via Upper Confidence Bounds (UCB). Unlike previous model-agnostic methods that directly occlude super-pixels, our method softly applies perturbations at a pixel level and thus can fully explore more regions with multiple granularities, leading to a more precise and robust interpretation. Quantitative and qualitative experimental results demonstrate that our approach provides reasonable and precise explanations for various image recognition tasks on different models. [ABSTRACT FROM AUTHOR]

• The layer sustainability analysis (LSA) framework is introduced to evaluate the behavior of layer-level representations of DNNs in dealing with network input perturbations using Lipschitz theoretical concepts. • A layer-wise regularized adversarial training (AT-LR) approach significantly improves the generalization and robustness of different deep neural network architectures for significant perturbations while reducing layer-level vulnerabilities. • AT-LR loss landscapes for each LSA MVL proposal can interpret layer importance for different layers, which is an intriguing aspect. Deep neural network models are used today in various applications of artificial intelligence, the strengthening of which, in the face of adversarial attacks is of particular importance. An appropriate solution to adversarial attacks is adversarial training, which reaches a trade-off between robustness and generalization. This paper introduces a novel framework (Layer Sustainability Analysis (LSA)) for the analysis of layer vulnerability in an arbitrary neural network in the scenario of adversarial attacks. LSA can be a helpful toolkit to assess deep neural networks and to extend the adversarial training approaches towards improving the sustainability of model layers via layer monitoring and analysis. The LSA framework identifies a list of Most Vulnerable Layers (MVL list) of the given network. The relative error, as a comparison measure, is used to evaluate representation sustainability of each layer against adversarial inputs. The proposed approach for obtaining robust neural networks to fend off adversarial attacks is based on a layer-wise regularization (LR) over LSA proposal(s) for adversarial training (AT). This means that the AT-LR procedure could be used with any benchmark adversarial attack to reduce the vulnerability of network layers and to improve conventional adversarial training approaches. The proposed idea performs well theoretically and experimentally for state-of-the-art multilayer perceptron and convolutional neural network architectures. Additionally, a measure named robustness and generalization score or R&G score is defined to better evaluate each adversarially trained model over a variety of significant perturbations. Compared with the AT-LR and its corresponding base adversarial training, the R&G score on Moon, MNIST, and CIFAR-10 benchmark datasets was increased by 56.52%, 75.82%, and 6.54%, respectively for more significant perturbations. The LSA framework is available and published at https://github.com/khalooei/LSA. [ABSTRACT FROM AUTHOR]

The applications of traditional statistical feature selection methods to high-dimension, low-sample-size data often struggle and encounter challenging problems, such as overfitting, curse of dimensionality, computational infeasibility, and strong model assumptions. In this paper, we propose a novel two-step nonparametric approach called Deep Feature Screening (DeepFS) that can overcome these problems and identify significant features with high precision for ultra high-dimensional, low-sample-size data. This approach first extracts a low-dimensional representation of input data and then applies feature screening on the original input feature space based on multivariate rank distance correlation recently developed by Deb and Sen (2021). This approach combines the strengths of both deep neural networks and feature screening, thereby having the following appealing features in addition to its ability of handling ultra high-dimensional data with small number of samples: (1) it is model free and distribution free; (2) it can be used for both supervised and unsupervised feature selection; and (3) it is capable of recovering the original input data. The superiority of DeepFS is demonstrated via extensive simulation studies and real data analyses. [ABSTRACT FROM AUTHOR]