Task-free online continual learning (TFOCL) requires models to adapt to non-stationary data streams without task boundaries, while processing each sample only once and without revisiting past data. A key challenge in TFOCL is catastrophic forgetting (CF), where distributional shifts rapidly degrade previously acquired knowledge. Recent contrastive replay-based methods improve robustness to CF by learning discriminative representations. However, the classifier is trained in a separate phase, detached from the representation learning pipeline, which prevents the model from providing accurate predictions while processing the data stream. To overcome this limitation, we propose TORCH (Task-Free Online Replay with Contrastive Learning and Dynamic Herding), which continuously updates the representation network and classifier in an online manner, enabling true anytime inference throughout the learning process. To enhance TORCH’s performance, we introduce a memory management strategy, Online Mean-Feature Sampling (OMFS), which enforces class balance and maintains representative exemplars within a constrained memory budget. By reducing redundancy and retaining representative exemplars, OMFS facilitates more effective replay. We evaluate TORCH on three benchmark datasets and demonstrate that it significantly outperforms state-of-the-art methods, achieving improvements of up to 2.8

Federated learning (FL) has emerged as a promising solution to enable distributed learning without sharing sensitive data. However, FL is vulnerable to data poisoning attacks, where malicious clients inject malicious data during training to compromise the global model. Existing FL defenses suffer from the assumptions of independent and identically distributed (IID) model updates, asymptotic optimal error rate bounds, and strong convexity in the optimization problem. Hence, we propose a novel framework called Federated Learning Optimal Transport (FLOT) that leverages the Wasserstein barycentric technique to obtain a global model from a set of locally trained non-IID models on client devices. In addition, we introduce a loss function-based rejection (LFR) mechanism to suppress malicious updates and a dynamic weighting scheme to optimize the Wasserstein barycentric aggregation function. We provide the theoretical proof of the Byzantine resilience and convergence of FLOT to highlight its efficacy. We evaluate FLOT on four benchmark datasets: GTSRB, KBTS, CIFAR10, and EMNIST. The experimental results underscore the practical significance of FLOT as an effective defense mechanism against data poisoning attacks in FL while maintaining high accuracy and scalability. Also, we observe that FLOT serves as a robust client selection technique under no attack, which demonstrates its effectiveness.

Objectives: This study focuses on the early recurrence prediction of hepatocellular carcinoma (HCC), a prevalent and life-threatening liver cancer, following surgical resection. Addressing the challenges posed by postoperative recurrence, particularly in early stages, is pivotal for improving patient outcomes. Our aim is to develop a computer-aided diagnostic tool that integrates clinical data and whole slide images (WSIs) within a multimodal learning framework to enhance prediction performance. Methods: Given the direct correlation between tissue structure in WSIs and patient prognosis, combined with the complementary nature of comprehensive patient profiling derived from clinical data, we propose an innovative multimodal fusion approach. Traditional methods, whether based solely on WSIs or multimodal fusion (e.g., CT/MRI and clinical data), have limitations. To overcome the challenges associated with WSI-based methods, we introduce tumor annotation-guided early recurrence probability heatmaps (ER-ProbMaps) to extract robust features relevant for early recurrence prediction. For multimodal fusion, our method incorporates clinical-guided cross-attention mechanisms to facilitate seamless integration of clinical data and WSI features. Findings: The results demonstrate that our proposed Improved Multiple Instance Learning (Improved MIL) model, which leverages ER-ProbMaps, outperforms all unimodal methods. Specifically, when compared to the state-of-the-art unimodal method, our model achieves a significant AUC increase of 1.6% and an ACC increase of 17.7%. Furthermore, in comparison with multimodal fusion methods, our Improved MIL model with Clinical-Guided Cross-Attention (Improved MIL-CGCA) surpasses all multimodal approaches. Notably, when benchmarked against the state-of-the-art multimodal method, our model exhibits an AUC improvement of 3% and demonstrates superior performance. These findings underscore the critical role of integrating clinical data and WSIs within a multimodal learning framework for enhanced early recurrence prediction in HCC. Novelty: This research introduces a novel multimodal learning framework that integrates clinical data and WSIs, leveraging tumor annotation-guided ER-ProbMaps and cross-attention mechanisms for improved early recurrence prediction in HCC. Our findings emphasize the importance of considering both imaging and clinical data for accurate and reliable predictions, fostering a significant advancement in the field of computer-aided HCC diagnosis and management.

The primary objective of this study is to develop a novel object detection model that excels in both inference speed and accuracy, particularly under low-light conditions. Object detection plays a crucial role in various applications, from surveillance to autonomous vehicles, but remains challenging in low-light environments due to reduced visibility and image quality, along with high computational complexity. To tackle these issues, we propose a model that integrates cutting-edge deep learning techniques for efficient performance on resource-constrained devices. The model architecture comprises three key components: a backbone with Cross-Stage Partial (CSP) connections to reduce computational demands and enhance gradient flow; a neck for effective feature integration across scales; and a head for streamlined prediction of classifications and bounding box coordinates. The model also benefits from varifocal loss and complete IoU loss functions, which improve training convergence and accuracy. Additionally, advanced techniques such as task-assigned learning and mixed precision training enhance its performance across various benchmarks, making it well-suited for accurate object detection in low-light conditions.

Point cloud-based three-dimensional (3D) object detection is a critical task in autonomous driving, robotics, and augmented reality, where accurate localization and classification of objects are essential under diverse and challenging scenarios. This work introduces a point cloud-based multi-target 3D object detection framework using Light Detection and Ranging (LiDAR) sensors. The key contribution lies in improving the encoding technique to preserve spatial information and adapt to varying point densities, enabling efficient processing of raw LiDAR data for robust 3D detection. Specifically, we design an efficient network architecture that incorporates a single unified classification head to jointly handle positive and negative samples, simplifying the design and improving training stability. Our model detects all classes (car, pedestrian, cyclist) simultaneously within a single network, enhancing computational efficiency while avoiding separate networks for each class. Furthermore, the proposed feature extraction pipeline combines a voxel-based encoder with a sparsely embedded convolutional backbone and a feature pyramid network, facilitating multi-scale feature representation and effective detection of objects at different scales. This encoding method facilitates the efficient processing of raw LiDAR data, enabling accurate object detection across diverse scenarios. As a result, our model achieves state-of-the-art performance on the KITTI dataset, surpassing baseline methods such as PointPillars and VoxelNet by delivering superior Average Precision (AP) across all difficulty levels for the car class in both bird’s-eye view (BEV) and 3D object detection tasks. Specifically, it achieves AP40 scores of 92.84

Autonomous vehicles must excel in safety-critical perception tasks, especially in adverse weather conditions. In addition, transitional weather shifts in nature, such as sunny to rainy, rainy to cloudy, etc., pose abrupt illumination changes that can distort object boundaries and degrade segmentation performance. Existing research focuses mainly on segmentation in clear and discrete weather conditions, leaving a gap in addressing the issues of transitional weather scenarios. Hence, we propose a novel method called causal road and rest segmentation (CaRS) that utilizes causal intervention to mitigate the confounding bias due to transitional weather changes. We use dual complementary attention modules, one for causal and another for confounding feature extraction. These modules complement each other and are fine-tuned via an adversarial min-max approach to reduce confounding bias and enhance segmentation performance. Also, our CaRS method concurrently performs road semantic segmentation and instance segmentation of vehicles and pedestrians. Further, we introduce a transitional weather-driving dataset for segmentation (TWDS16) using a spurious correlation generator that leverages data interpolation to produce 16 weather transitions. We evaluate the performance of CaRS on TWDS16, along with three other benchmark datasets, namely, Foggy Cityscapes, RainCityscapes, and BDD100K. The experimental results validate the efficacy of the proposed method in mitigating confounding influences, leading to improved mIoU for semantic segmentation and mAP for instance segmentation across diverse datasets.

Small Multimodal Models (SMMs) suffer under distribution shift after fine-tuning. Unsupervised Domain Adaptation (UDA) is a common remedy for this issue, but existing theory and methods are designed primarily for single-or dual-encoder architectures, overlooking the encoder–decoder structure of SMMs, whose fusion mechanism introduces additional shift. This work bridges this gap in two steps. First, we derive a dual-divergence risk bound that separates encoder divergence from fusion divergence and illustrate its tightness compared to the classical encoder-only bound with a negation-flip example. Second, motivated by this theory, we propose Dual-level Adversarial Alignment (DuAA), a two-stage alignment algorithm. DuAA inserts domain-discriminative adapters after the encoder and within the decoder to minimize both divergences. Furthermore, DuAA employs selective pseudo-labeling to refine target semantics. Our contribution targets domain shift in encoder–decoder SMMs and is agnostic to the fine-tuning mechanism, i.e., DuAA acts on internal representations, making it orthogonal to LoRA, or other fine-tuning variants. We adopt LoRA in experiments solely as a popular, parameter-efficient instantiation to keep the protocol fixed across settings. We compile twelve new cross-domain VQA tasks with distinct visual and textual shifts from existing datasets and observe that DuAA consistently outperforms standard fine-tuning across all tasks.

Accurate 3D object detection from LiDAR data is vital for enhancing road safety, enabling efficient traffic management, and supporting reliable path planning in autonomous navigation systems. However, LiDAR point clouds suffer from inherent challenges such as sparsity, occlusion, and variations in point density, which can significantly impact detection accuracy. To address these challenges, we introduce 3DA-Net, a dual-attention-based network that integrates global and local context for enhanced 3D object detection. We begin by converting raw LiDAR point clouds into structured voxel representations, which are then processed through a hybrid dual-attention encoder. In this encoder, global attention modules capture high-level semantic dependencies across the entire scene, while local attention focuses on fine-grained geometric structures within neighborhoods. This dual-attention mechanism is further strengthened with point-wise and channel-wise attention, which enhances the model’s ability to capture both spatial and contextual information, which is essential for 3D perception. Our design incorporates a custom backbone for robust feature extraction from voxel-based pseudo-image representations, coupled with a feature pyramid network for efficient multi-scale feature learning. Evaluations on the KITTI dataset show that 3DA-Net achieves AP40 scores of 95.91

The earthquake early warning (EEW) system provides advance notice of potentially damaging ground shaking. In EEW, early estimation of magnitude is crucial for timely rescue operations. A set of thirty-four features is extracted using the primary wave earthquake precursor signal and site-specific information. In Japan's earthquake magnitude dataset, there is a chance of a high imbalance concerning the earthquakes above strong impact. This imbalance causes a high prediction error while training advanced machine learning or deep learning models. In this work, Conditional Tabular Generative Adversarial Networks (CTGAN), a deep machine learning tool, is utilized to learn the characteristics of the first arrival of earthquake P-waves and generate a synthetic dataset based on this information. The result obtained using actual and mixed (synthetic and actual) datasets will be used for training the stacked ensemble magnitude prediction model, MagPred, designed specifically for this study. There are 13295, 3989, and 1710 records designated for training, testing, and validation. The mean absolute error of the test dataset for single station magnitude detection using early three, four, and five seconds of P wave are 0.41, 0.40, and 0.38 MJMA. The study demonstrates that the Generative Adversarial Networks (GANs) can provide a good result for single-station magnitude prediction. The study can be effective where less seismic data is available. The study shows that the machine learning method yields better magnitude detection results compared with the several regression models. The multi-station magnitude prediction study has been conducted on prominent Osaka, Off Fukushima, and Kumamoto earthquakes. Furthermore, to validate the performance of the model, an inter-region study has been performed on the earthquakes of the India or Nepal region. The study demonstrates that GANs can discover effective magnitude estimation compared with non-GAN-based methods. This has a high potential for wide application in earthquake early warning systems.

Navigating autonomous vehicles in adverse weather conditions such as extreme rain, fog, and clouds poses significant challenges in efficiently identifying road elements and performing path planning. These difficulties are further compounded by the frequent transitions between diverse weather conditions, such as shifting from cloudy to rainy or rainy to sunny. Existing research for autonomous vehicles has the least focus on performing object detection in transitional weather conditions. Also, autonomous driving data is vulnerable to minor perturbations or noises, leading to unpredictable outcomes, particularly in transitional weather conditions. In addition, current adversarial attacks demand significant computational resources and have limited real-world applicability due to the large number of queries and computational resources. To address these limitations, we propose a novel minimalistic method called explainable black-box adversarial detection attack in transitional weather conditions for autonomous driving (TransWardX). Our attention-guided attack is minimalist, targeting limited image regions to deceive the model effectively. We assess our attack using a continuous weather-driving dataset called AIWD6. Later, we also evaluate our attack with other datasets like BDD100K and GTSRB. Our results demonstrate the effectiveness of TransWardX, achieving a high success rate with minimal perturbations, fewer iterations, and a drastic reduction of 50

Federated Learning (FL) enables collaborative model training across decentralized devices while maintaining data privacy. However, in many real-world scenarios, accessing both source and labeled target data on these devices is not feasible due to privacy concerns, data heterogeneity, and resource constraints. This paper introduces a novel approach to address this challenge, termed Federated Source-Free Multi-Target Domain Adaptation (Fed-SMTDA). In this framework, client devices possess only unlabeled target data, while the server has a pretrained model derived from source data. Our method leverages the intrinsic structure of the target domain at each client by clustering similar features, facilitating more effective feature extraction and assignment. To improve model performance, we incorporate an entropy regularization term to minimize class confusion, ensuring cleaner decision boundaries. Additionally, we introduce a dynamic aggregation strategy called Weight Adjustment (WA), where the server adjusts the weights assigned to client models during aggregation based on the observed generalization gap across clients. This adaptive approach improves the overall robustness and generalization of the federated model, enabling it to perform effectively across diverse and unlabeled target domains. Fed-SMTDA is evaluated on the OfficeHome and PACS datasets. Its performance is compared against two centralized baselines - one that has full access to labels across domains and is termed Oracle, and another where the model is trained only with the source domain labels, which we call Source-Only. Fed-SMTDA delivers consistently better performance than the Source-Only model, and its performance is upper-bounded by the Oracle model.

The time allocation problem in multi-function cognitive radar systems focuses on the trade-off between scanning for newly emerging targets and tracking the previously detected targets. We formulate this as a multi-objective optimization problem and employ deep reinforcement learning to find Pareto-optimal solutions and compare deep deterministic policy gradient (DDPG) and soft actor-critic (SAC) algorithms. Our results demonstrate the effectiveness of both algorithms in adapting to various scenarios, with SAC showing improved stability and sample efficiency compared to DDPG. We further employ the NSGA-II algorithm to estimate an upper bound on the Pareto front of the considered problem. This work contributes to the development of more efficient and adaptive cognitive radar systems capable of balancing multiple competing objectives in dynamic environments.

Scene classification in remote sensing (RS) images is a challenging task due to the limited availability of labeled data and the high intraclass variability in complex landscapes. Semi-supervised learning (SSL) has emerged as an effective approach to leverage the limited labeled data in utilizing a large amount of unlabeled data for improved classification. Pseudo-labeling (PL), a widely used SSL technique, determines suitable labels to unlabeled data based on high-confidence model predictions. However, traditional PL methods suffer from confirmation bias, where incorrect labels reinforce errors, degrading model performance. To address this, we propose contrapositive learning-based PL (CPL-PL), a novel method designed specifically for RS scene classification. CPL-PL introduces a contrapositive loss (CPLoss) that enforces feature consistency for similar scenes while ensuring representation separation for dissimilar ones, leading to more reliable pseudo-label assignments. Our approach mitigates pseudo-label noise, enhances feature discrimination, and improves classification robustness. Experimental results on benchmark RS datasets demonstrate that CPL-PL significantly outperforms conventional PL strategies, especially in low-label regimes. The proposed method provides a promising direction for advancing semi-supervised scene classification in RS images.

In a conventional Domain Adaptation (DA) setting, we only have one source and target domain, whereas, in many real-world applications, data is often collected from several related sources in different conditions. This has led to a more practical and challenging knowledge transfer problem called Multi-source Domain Adaptation (MDA). Several methodologies, such as prototype matching, explicit distance discrepancy, adversarial learning, etc., have been considered to tackle the MDA problem in recent years. Among them, the adversarial-based learning framework is a popular methodology for transferring knowledge from multiple sources to target domains using a minmax optimization strategy. Despite the advances in adversarial-based methods, several limitations exist, such as the need for a classifier-aware discrepancy metric to align the domains and the need to consider target samples’ consistency and semantic information while aligning the domains. To mitigate these issues, in this work, we propose a novel adversarial learning MDA algorithm, MDAMA, which aligns the target domain with a mixture distribution that consists of source domains. MDAMA uses margin-based discrepancy and augmented intermediate distributions to align the domains effectively. We also propose consistency of target samples by confidence thresholding and transfer of semantic information from multiple source domains to the augmented target domain to further improve the performance of the target domain. We extensively experiment with the MDAMA algorithm on popular real-world MDA datasets such as OfficeHome, Office31, PACS, Office-Caltech, and DomainNet. We evaluate the MDAMA model on these benchmark datasets and demonstrate top performance in all of them.

Knowledge distillation (KD) is a model compression technique that transfers knowledge from a complex and well-trained teacher model to a compact student model, thereby enabling the student to mimic the performance and behavior of the teacher. However, traditional KD methods struggle with long-tailed facial expression recognition (FER), as FER datasets often exhibit severe class imbalance. For instance, certain expressions (e.g., happiness) are overrepresented, while others (e.g., fear) have significantly fewer samples. This innate class-imbalanced property of FER leads to suboptimal knowledge transfer for the underrepresented expressions (i.e. biased learning and poor generalization for underrepresented classes). To address this issue, this paper introduces the CE-KD, a Class-wise Expert-based Knowledge Distillation framework that enables the student model to effectively learn fine-grained expression details and high-level emotional concepts from specialized emotion experts. This improves the generalizability of the FER models. Extensive experiments on benchmark datasets — FERPlus and RAF-DB — demonstrate that our CE-KD framework provides a practical solution to implement efficient FER systems in real-world applications while maintaining robust performance across different emotion expressions.

This paper explores second-order optimization methods in Federated Learning (FL), addressing the critical challenges of slow convergence and the excessive communication rounds required to achieve optimal performance from the global model. While existing surveys in FL primarily focus on challenges related to statistical and device label heterogeneity, as well as privacy and security concerns in first-order FL methods, less attention has been given to the issue of slow model training. This slow training often leads to the need for excessive communication rounds or increased communication costs, particularly when data across clients are highly heterogeneous. In this paper, we examine various FL methods that leverage second-order optimization to accelerate the training process. We provide a comprehensive categorization of state-of-the-art second-order FL methods and compare their performance based on convergence speed, computational cost, memory usage, transmission overhead, and generalization of the global model. Our findings show the potential of incorporating Hessian curvature through second-order optimization into FL and highlight key challenges, such as the efficient utilization of Hessian and its inverse in FL. This work lays the groundwork for future research aimed at developing scalable and efficient federated optimization methods for improving the training of the global model in FL.