Continuous learning constitutes a fundamental capability of artificial intelligence systems, enabling them to incrementally assimilate novel information without succumbing to catastrophic forgetting. Recent research has leveraged Pre-Trained Models (PTMs) to enhance continual learning efficacy. Nevertheless, prevailing methodologies typically depend on a singular pre-trained backbone and freeze all pre-trained parameters to mitigate network forgetting, thereby constraining adaptability to emerging tasks. In this study, we introduce an innovative PTM-based framework featuring a Dual-Representation Backbone Architecture (DRBA), which integrates both invariant and evolved representation networks to concurrently capture static and dynamic features. Building upon DRBA, we propose an Adaptive and Expandable Mixture Model (AEMM) that incrementally incorporates new expert modules with minimal parameter overhead to accommodate the learning of each novel task. To further augment adaptability, we develop a Dynamic Adaptive Representation Fusion Mechanism (DARFM) that processes outputs from both representation networks and autonomously generates data-driven adaptive weights, optimizing the contribution of each representation. This mechanism yields an adaptive, semantically enriched composite representation, thereby maximizing positive knowledge transfer. Additionally, we propose a Dynamic Knowledge Calibration Mechanism (DKCM), comprising prediction and representation calibration processes, to ensure consistency in both predictions and feature representations. This approach achieves a balance between stability and plasticity, even when learning complex datasets. Empirical evaluations substantiate that the proposed approach attains state-of-the-art performance.

Visual object tracking stands out as a crucial foundational task in computer vision that enjoys a long-standing reputation and widespread application. However, in recent years, adversarial attacks on visual trackers have posed significant threats to their robustness. Regrettably, existing defense methods encounter challenges in transferring from siamese trackers to transformer-based trackers, together with the severe performance decline on clean samples. These shortcomings impede the scalability of existing defense methods on heterogeneous trackers, thereby posing challenges to real-world applications. To address these issues, we present AADN++, a more transferable adversarial defense network. Specifically, based on the observation that attacked latent features tend to deviate from original features at different convolutional scales, we introduces the Latent Feature Loss to improve defense performance and transferability. The LFL is comprised of multi-scale feature loss and classification-regression loss. Furthermore, we enhance the adversarial training process by incorporating an extra forward pass to boost tracking accuracy on clean samples. Experimental evaluations on the OTB100, VOT2018, and LaSOT benchmarks demonstrate that AADN++ exhibits superior defense transferability on heterogeneous trackers and exhibits outstanding robustness against generative and iterative attacks.

Plasticity and stability denote the ability to assimilate new tasks while preserving previously acquired knowledge, representing two important concepts in continual learning. Recent research addresses stability by leveraging pre-trained models to provide informative representations, yet the efficacy of these methods is highly reliant on the choice of the pre-trained backbone, which may not yield optimal plasticity. This paper addresses this limitation by introducing a streamlined and potent framework that orchestrates multiple different pre-trained backbones to derive semantically rich multi-source representations. We propose an innovative Multi-Scale Interaction and Dynamic Fusion (MSIDF) technique to process and selectively capture the most relevant parts of multi-source features through a series of learnable attention modules, thereby helping to learn better decision boundaries to boost performance. Furthermore, we introduce a novel Multi-Level Representation Optimization (MLRO) strategy to adaptively refine the representation networks, offering adaptive representations that enhance plasticity. To mitigate over-regularization issues, we propose a novel Adaptive Regularization Optimization (ARO) method to manage and optimize a switch vector that selectively governs the updating process of each representation layer, which promotes the new task learning. The proposed MLRO and ARO approaches are collectively optimized within a unified optimization framework to achieve an optimal trade-off between plasticity and stability. Our extensive experimental evaluations reveal that the proposed framework attains state-of-the-art performance. The source code of our algorithm is available at https://github.com/CL-Coder236/LMSRR.

Continual learning requires the model to continually capture novel information without forgetting prior knowledge. Nonetheless, existing studies predominantly address the catastrophic forgetting, often neglecting enhancements in model robustness. Consequently, these methodologies fall short in real-time applications, such as autonomous driving, where data samples frequently exhibit noise due to environmental and lighting variations, thereby impairing model efficacy and causing safety issues. In this paper, we address robustness in continual learning systems by introducing an innovative approach, the Dynamic Siamese Expansion Framework (DSEF) that employs a Siamese backbone architecture, comprising static and dynamic components, to facilitate the learning of both global and local representations over time. Specifically, the proposed framework dynamically generates a lightweight expert for each novel task, leveraging the Siamese backbone to enable rapid adaptation. A novel Robust Dynamic Representation Optimization (RDRO) approach is proposed to incrementally update the dynamic backbone by maintaining all previously acquired representations and prediction patterns of historical experts, thereby fostering new task learning without inducing detrimental knowledge transfer. Additionally, we propose a novel Robust Feature Fusion (RFF) approach to incrementally amalgamate robust representations from all historical experts into the expert construction process. A novel mutual information-based technique is employed to derive adaptive weights for feature fusion by assessing the knowledge relevance between historical experts and the new task, thus maximizing positive knowledge transfer effects. A comprehensive experimental evaluation, benchmarking our approach against established baselines, demonstrates that our method achieves state-of-the-art performance even under adversarial attacks.

The advancement of artificial intelligence security research has led to the emergence of adversarial attack technology as a critical approach for identifying potential security vulnerabilities in artificial intelligence models . When targeting natural language processing models, conducting adversarial attacks in the hard-label setting presents a more practical and challenging black-box scenario due to the difficulty in computing gradients directly from discrete word sequences. Current textual adversarial attack methods are inefficient due to the lack of consideration for the limited number of queries available during the adversarial text generation process, creating a disparity between these approaches and real-world adversarial attack scenarios. To this end, this work proposes a dual-granularity optimization strategy that consists of a single-word semantic optimization and a multi-word joint semantic optimization procedure , and presents a query-efficient hard-label attack method called DualAttack by incorporating the proposed dual-granularity optimization strategy into the mutation and crossover process of the Genetic Algorithm framework. Extensive experimental results demonstrate that DualAttack can effectively produce high-quality adversarial texts with superior semantic similarity and minimal perturbation rates within fewer queries compared to existing methods in the hard-label setting.

Evaluating the potential risks posed by adversarial examples is crucial for securely deploying deep neural network (DNN)-based Internet of Things (IoT) devices. Although adversarial patches are considered a primary physical attack strategy, recent studies suggest their real-world impact may be less significant. Research has explored using optical tools like lasers or projectors to create perturbations, but these methods are uncommon in natural settings. Given the visual challenges inherent in natural environments, the prevalent occurrence of lens flare is noteworthy. This phenomenon can obstruct human vision and may also be exploited maliciously. In this article, we emphasize for the first time that lens flares produced by light sources exhibit strong adversarial characteristics. In the physical world, lens flare that occurs around and directly impacts the target object can easily deceive advanced models in various real-world scenarios. We show that merely using a regular flashlight to generate adversarial flare is sufficient to easily deceive well-trained models during both day and night in typical traffic situations, with an average attack success rate (ASR) of over 90%. Furthermore, by utilizing real lens flares as perturbations, we introduce a novel digital-domain closed-box adversarial attack method, primarily designed for extensive experimental validation of adversarial lens flare attacks in the physical world. Experimental results demonstrate that adversarial lens flare attacks effectively deceive state-of-the-art models, including YoloV8, DinoV2, and other three baseline models, achieving average ASRs as high as 95.4% on GTSRB and 82.8% on MTSD. We also discuss the limitations and defense mechanisms against this attack.

Background: Integral equations play a crucial role in modeling complex systems across various scientific disciplines. However, traditional numerical methods and existing physics-informed neural networks (PINNs) face substantial challenges, including the curse of dimensionality, uncontrolled error propagation, and limited generalization capabilities. Objectives: This paper aims to overcome these limitations by developing a robust and scalable solver for high-dimensional and nonlinear integral equations. The primary goal is to achieve higher accuracy and efficiency compared to traditional methods and existing deep learning approaches. Methods: We present the enhanced physics-informed neural network (EPINN), a novel framework that incorporates three key innovations: 1) a variable-order operator decomposition theory that transforms integral equations into well-posed differential systems, thereby mitigating error accumulation, 2) a differentiable primal function projection layer that ensures physical consistency within the Sobolev spaces, and 3) a boundary-aware multi-objective training paradigm that improves generalization. Results: Experimental validation across five benchmark cases spanning two to four dimensions, including linear/nonlinear Volterra/Fredholm and hybrid Volterra-Fredholm integral equations, demonstrates the superior performance of EPINN. Compared with traditional methods, EPINN reduces relative errors by 1 to 2 orders of magnitude, while achieving over 92% accuracy with limited training data. When compared with existing deep learning solvers, EPINN provides significant improvements in computational efficiency (with a speedup factor of 3 to 6 times) and accuracy (error reduction of 23% to 85%). Conclusions: These advancements establish EPINN as a robust and scalable solver for high-dimensional and nonlinear integral equations, with wide-ranging applications in computational physics and engineering. The success of EPINN suggests that integrating physical principles with neural networks can lead to substantial improvements in solving complex mathematical problems.

The need to capture intra-word information in natural language processing (NLP) tasks has inspired research in learning various word representations at word, character, or morpheme levels, but little attention has been given to syllables from a syllabic alphabet. Motivated by the success of compositional models in morphological languages, we present a Convolutional-long short term memory (Conv-LSTM) model for constructing Swahili word representation vectors from syllables. The unified architecture addresses the word agglutination and polysemous nature of Swahili by extracting high-level syllable features using a convolutional neural network (CNN) and then composes quality word embeddings with a long short term memory (LSTM). The word embeddings are then validated using a syllable-aware language model ( 31.267 ) and a part-of-speech (POS) tagging task ( 98.78 ), both yielding very competitive results to the state-of-art models in their respective domains. We further validate the language model using Xhosa and Shona, which are syllabic-based languages. The novelty of the study is in its capability to construct quality word embeddings from syllables using a hybrid model that does not use max-over-pool common in CNN and then the exploitation of these embeddings in POS tagging. Therefore, the study plays a crucial role in the processing of agglutinative and syllabic-based languages by contributing quality word embeddings from syllable embeddings, a robust Conv–LSTM model that learns syllables for not only language modeling and POS tagging, but also for other downstream NLP tasks.

Machine olfaction is an intelligent system that combines a cross-sensitivity chemical sensor array and an effective pattern recognition algorithm for the detection, identification, or quantification of various odors. Data collected by the sensor array are the multivariate time series signals with a complex structure, and these signals become more difficult to analyze due to sensor drift. In this work, we focus on improving the classification performance under sensor drift by using the deep learning method, which is popular nowadays. Compared with other methods, our method can effectively tackle sensor drift by automatically extracting features, thus not only removing the complexity of designing the hand-made features but also making it pervasive for a variety of application in machine olfaction. Our experimental results show that the deep learning method can learn the features that are more robust to drift than the original input and achieves high classification accuracy.

Motion detection is a basis step for video processing. Previous works of motion detection based on deep learning need clean foreground or background images which always do not exist in practice. To address this challenge, a novel and practical method is proposed based on auto-encoder neural networks. First, the approximate background images are obtained via an auto-encoder network (called Reconstruction Network) from video frames. Then, a background model is learned based on these images by using another auto-encoder network (called Background Network). To be more resilient, our background model can be updated on-line to absorb more training samples. Our main contributions are 1) the architecture of the couple of auto-encoder networks which can model the background very efficiently; 2) the online learning algorithm in which a method of searching the minimizing effect parameters is adopted to accelerate the training of the Reconstruction Network. Our approach improves the motion detection performance on three data sets.

Background learning is a pre-processing of motion detection which is a basis step of video analysis. For the static background, many previous works have already achieved good performance. However, the results on learning dynamic background are still much to be improved. To address this challenge, in this paper, a novel and practical method is proposed based on deep auto-encoder networks. Firstly, dynamic background images are extracted through a deep auto-encoder network (called Background Extraction Network) from video frames containing motion objects. Then, a dynamic background model is learned by another deep auto-encoder network (called Background Learning Network) using the extracted background images as the input. To be more flexible, our background model can be updated on-line to absorb more training samples. Our main contributions are 1) a cascade of two deep auto-encoder networks which can deal with the separation of dynamic background and foregrounds very efficiently; 2) a method of online learning is adopted to accelerate the training of Background Extraction Network. Compared with previous algorithms, our approach obtains the best performance over six benchmark data sets. Especially, the experiments show that our algorithm can handle large variation background very well.

Recently, a few adaptive algorithms for generalized eigen-decomposition have been proposed, which are very useful in many applications such as digital mobile communications, Blind signal separation, etc. These algorithms are all focusing on extracting principal generalized eigenvectors. However, in many practical applications such as dimension reduction and signal processing, extracting the minor generalized eigenvectors adaptively are needed. Because of little literatures in the community, we discuss several approaches that lead to a few novel algorithms for extracting minor generalized eigenvectors. First, we derive an adaptive algorithms by using a single-layer linear forward neural network from the viewpoint of linear discriminant analysis(LDA). And the algorithm to extract multiple minor generalized eigenvectors are also proposed by using orthogonality property. Second, by using gradient ascent approach of some objective functions, we can derive more algorithms and explain the first algorithm. Then, we extend these algorithms to minor generalized eigenvector problem. Theoretical analysis shows that these algorithms are stable and convergent to the minor generalized eigenvectors. Simulations have been conducted for illustration of the efficiency and effectiveness of our algorithms.