In Internet of Things (IoT), artificial intelligence (AI) agents have become a core paradigm for edge intelligence, supporting on-device sensor data processing and privacy-preserving analytics. In practice, however, the sensor data streams evolve continuously and often exhibit nonindependent and identically distributed (non-IID) characteristics. Federated class-incremental learning (FCIL) has thus emerged as a promising approach to endow edge AI agents with long-tail learning capabilities, allowing models to adapt autonomously to persistent changes in the data. Despite recent progress, most existing FCIL methods assume balanced data distributions, which rarely hold in real-world deployments. Under realistic long-tailed distributions, the local optimization objectives of individual AI agents diverge from the desired global convergence, leading to head-class bias and rapid forgetting of tail classes. To overcome these limitations, we propose gradient-aligned and imbalance-aware FCIL model (FedGRAIL), a unified FCIL framework that resolves the conflict between local optimization bias and global consistency in dynamic IoT environments. FedGRAIL integrates local training and global aggregation within a structurally aligned learning process, rather than treating them as separate stages. It further introduces a gradient-correction mechanism that coordinates dispersed local optima and stabilizes global optimization via diversity-driven generative replay. This design supports effective retention of minority concepts while respecting the zero-storage constraints of edge agents. Extensive experiments on long-tailed benchmarks show that FedGRAIL substantially improves tail-class accuracy under both independent and identically distributed (IID) and non-IID conditions - achieving a 59% average gain over the strongest baseline, while preserving the competitive accuracy on head classes. These results demonstrate that FedGRAIL provides a storage-efficient, privacy-preserving, and robust solution for continual learning in IoT edge intelligence.

Fatigued driving is a major societal risk factor in the road safety system, affecting the efficiency of the entire transportation network and public safety. Fatigue detection methods based on multimodal physiological signals can objectively reflect a driver’s neurocognitive state, offering higher monitoring reliability. However, existing methods face two main challenges: 1) the high cost of annotating high-quality physiological signals leads to challenges in learning with limited data; and 2) the signal-to-noise ratio of multimodal physiological signals is variable, leading to imbalances in the confidence in the decision between modalities. To address these issues, this article proposes the meta-evidential dual-stream network (MEDS-Net), which integrates a meta-learning framework with evidential deep learning to establish a solution for driver fatigue detection for the first time. Specifically, MEDS-Net employs a dual-stream evidence generation network to extract heterogeneous features from multimodal physiological signals and uses a Dirichlet distribution to quantify cognitive uncertainty for each modality. Additionally, MEDS-Net introduces a meta-learning framework to enhance the model’s ability to handle data scarcity through the design of task simulation mechanisms and a two-layer optimization strategy. Extensive experiments on the SEED-VIG dataset demonstrate that MEDS-Net outperforms current state-of-the-art methods in terms of effectiveness and generalizability. Furthermore, MEDS-Net can precisely quantify the importance of global features in explaining modal contributions and average activation patterns.

The emergence of large language models (LLMs) has been a key enabler of technological innovation in healthcare. People can conveniently obtain a more accurate medical consultation service by utilizing LLMs' powerful knowledge inference capability. However, existing LLMs require users to upload explicit requests during remote healthcare consultations, which involves the risk of exposing personal privacy. Furthermore, the reliability of the response content generated by LLMs is not guaranteed. To tackle the above challenges, this paper proposes a novel privacy-preserving LLM for user-activated health, called Adaptive Compressed-based Privacy-preserving LLM (ACP2LLM). Specifically, an adaptive token compression method based on information entropy is carefully designed to ensure that ACP2LLM can preserve user-sensitive information when invoking the medical consultation of LLMs deployed on the cloud platform. Moreover, a multi-doctor one-chief physician mechanism is proposed to rationally split and collaboratively infer the patients' requests to achieve the privacy-utility trade-off. Notably, the proposed ACP2LLM also provides highly competitive performance in various token compression rates. Extensive experiments on multiple Medical Question and Answers datasets demonstrate that the proposed ACP2LLM has strong privacy protection capabilities and high answer precision, outperforming current state-of-the-art LLM methods.

In multimodal sentiment analysis (MSA), recent works introduced uncertainty quantification to evaluate model prediction reliability, aiming to achieve well-performing and trustworthy models. However, a trustworthy model must not only quantify predictive uncertainty but also ensure that the quantified uncertainty is accurate, which recent studies have neglected. Therefore, we propose an epistemic uncertainty (EU) verification task, which aims to estimate and compare the accuracy of the uncertainty quantification of different models by constructing a test dataset with samples of hierarchical prediction difficulty. Inspired by this task, we propose a Reliable Uncertainty Trustworthy Model (RUTM), which is validated against the EU verification task to assess its accuracy in predicting uncertainty. RUTM improves generalizable sentiment modeling through Information Entropy Disentanglement method and achieves more accurate uncertainty predictions via the EU Modeling Module. Extensive experiments on three public MSA datasets confirm that our model achieves state-of-the-art results, surpassing all baselines in uncertainty prediction accuracy.

In Multimodal Sentiment Analysis (MSA), data noise arising from various sources can lead to uncertainty in Aleatoric Uncertainty (AU), significantly impacting model performance. Current efforts to address AU have insufficiently explored its sources. They primarily focus on modeling noise rather than implementing targeted modeling based on its origin. Consequently, these approaches struggle to effectively mitigate the influence of AU, resulting in sustained limitations in model performance. Our research identifies that the AU primarily stems from two problems: subjective bias in the annotation process and the complex set relationships of sentiment features. To specifically address them, we propose DFMU, a Distribution-based Framework for Modeling Aleatoric Uncertainty, which incorporates an uncertainty modeling block capable of encoding uncertainty distributions and adaptively adjusting optimization objectives. Furthermore, we introduce distribution-based contrastive learning with sentiment words replacement to better capture the complex relationships among features. Extensive experiments on three public MSA datasets, i.e., MOSI, MOSEI, and SIMS, demonstrate that the proposed model maintains robust performance even under high noise conditions and achieves state-of-the-art results on these popular datasets.

Physiological multimodal emotion recognition (PMER) has become a key research direction for advancing human-computer interaction and affective computing. However, current PMER methods are affected by significant individual differences and the limited number of samples, making it challenging to capture complex emotional states comprehensively. To address aforementioned issues, this letter proposes a novel multimodal Few-Shot emotion recognition model, called Hierarchical Interactive Alignment Network (HIA-Net). Specifically, the Hierarchical Adaptive Interactive Attention (HAIA) module of HIA-Net is proposed to capture multidimensional emotional features and aggregate the cross-modal information effectively. Additionally, a cross-domain optimization strategy based on the maximum mean discrepancy is proposed to enhance the HIA-Net's adaptability across varying data distributions. Experimental results show that HIA-Net achieves state-of-the-art performance under Few-Shot experimental paradigms on the SEED and SEED-FRA datasets. Our codes are available athttps://github.com/As1yk/HIA-Net.

Although multimodal physiological data from the central and peripheral nervous systems can objectively respond to human emotional states, the individual differences caused by non-stationary and low signal-to-noise properties bring several challenges to cross-subject emotion recognition tasks. Many previous studies usually focused on learning high correlation information between different modalities, which easily leads to incomplete descriptions of different physiological signals and difficulties in aligning critical emotional information. To tackle these challenges, this paper proposes a novel multimodal emotion recognition model for improving the generalization performance to unseen target domain subjects, termed Completeness-induced Adaptative Broad Learning (CiABL). The proposed CiABL can gradely explore the completeness modality representation that encompasses both modality-relevant and modality-independent information, avoiding the loss of performance due to spurious correlations from different modalities. Subsequently, a well-designed weighted representation distribution alignment mechanism of CiABL can appropriately align the marginal and conditional distributions to reduce the influences of individual differences greatly. Extensive experiments on the SEED and SEED-FRA datasets demonstrate the effectiveness and generalization of the proposed CiABL, which outperforms current state-of-the-art methods. In addition, CiABL can precisely quantify the importance of global features to properly explain the modality contribution and averaged activation patterns of the brain under cross-subject emotion recognition tasks.

With increasing social globalization, interaction between people from different cultures has become more frequent. However, there are significant differences in the expression and comprehension of emotions across cultures. Therefore, developing computational models that can accurately identify emotions among different cultures has become a significant research problem. This study aims to investigate the similarities and differences in emotion cognition processes in different cultural groups by employing a fusion of electroencephalography (EEG) and eye movement (EM) signals. Specifically, an effective adaptive region selection method is proposed to investigate the most emotion-related activated brain regions in different groups. By selecting these commonly activated regions, we can eliminate redundant features and facilitate the development of portable acquisition devices. Subsequently, the multiple stacked broad learning system (MSBLS) is designed to explore the complementary information of EEG and EM features and the effective emotional information still contained in the residual value. The intracultural subject-dependent (ICSD), intracultural subject-independent, and cross-cultural subject-independent (CCSI) experiments have been conducted on the SEED-CHN, SEED-GER, and SEED-FRA datasets. Extensive experiments manifest that MSBLS achieves superior performance compared with current state-of-the-art methods. Moreover, we discover that some brain regions (the anterior frontal, temporal, and middle parieto-occipital lobes) and Gamma frequency bands show greater activation during emotion cognition in diverse cultural groups.

Recognizing novel sub-categories with scarce samples is an essential and challenging research topic in computer vision. Existing literature addresses this challenge by employing local-based representation approaches, which may not sufficiently facilitate meaningful object-specific semantic understanding, leading to a reliance on apparent background correlations. Moreover, they primarily rely on high-dimensional local descriptors to construct complex embedding space, potentially limiting the generalization. To address the above challenges, this article proposes a novel model, Robust Saliency-aware Distillation (RSaD), for few-shot fine-grained visual recognition. RSaD introduces additional saliency-aware supervision via saliency detection to guide the model toward focusing on the intrinsic discriminative regions. Specifically, RSaD utilizes the saliency detection model to emphasize the critical regions of each sub-category, providing additional object-specific information for fine-grained prediction. RSaD transfers such information with two symmetric branches in a mutual learning paradigm. Furthermore, RSaD exploits inter-regional relationships to enhance the informativeness of the representation and subsequently summarize the highlighted details into contextual embeddings to facilitate the effective transfer, enabling quick generalization to novel sub-categories. The proposed approach is empirically evaluated on three widely used benchmarks, demonstrating its superior performance.

Sentiment analysis uses a series of automated cognitive methods to determine the author's or speaker's attitudes toward an expressed object or text's overall emotional tendencies. In recent years, the growing scale of opinionated text from social networks has brought significant challenges to humans' sentimental tendency mining. The pretrained language model designed to learn contextual representation achieves better performance than traditional learning word vectors. However, the existing two basic approaches for applying pretrained language models to downstream tasks, feature-based and fine-tuning methods, are usually considered separately. What is more, different sentiment analysis tasks cannot be handled by the single task-specific contextual representation. In light of these pros and cons, we strive to propose a broad multitask transformer network (BMT-Net) to address these problems. BMT-Net takes advantage of both feature-based and fine-tuning methods. It was designed to explore the high-level information of robust and contextual representation. Primarily, our proposed structure can make the learned representations universal across tasks via multitask transformers. In addition, BMT-Net can roundly learn the robust contextual representation utilized by the broad learning system due to its powerful capacity to search for suitable features in deep and broad ways. The experiments were conducted on two popular datasets of binary Stanford Sentiment Treebank (SST-2) and SemEval Sentiment Analysis in Twitter (Twitter). Compared with other state-of-the-art methods, the improved representation with both deep and broad ways is shown to achieve a better F1 -score of 0.778 in Twitter and accuracy of 94.0% in the SST-2 dataset, respectively. These experimental results demonstrate the abilities of recognition in sentiment analysis and highlight the significance of previously overlooked design decisions about searching contextual features in deep and broad spaces.

In recent years, the appearance of the broad learning system (BLS) is poised to revolutionize conventional artificial intelligence methods. It represents a step toward building more efficient and effective machine-learning methods that can be extended to a broader range of necessary research fields. In this survey, we provide a comprehensive overview of the BLS in data mining and neural networks for the first time, focusing on summarizing various BLS methods from the aspects of its algorithms, theories, applications, and future open research questions. First, we introduce the basic pattern of BLS manifestation, the universal approximation capability, and essence from the theoretical perspective. Furthermore, we focus on BLS's various improvements based on the current state of the theoretical research, which further improves its flexibility, stability, and accuracy under general or specific conditions, including classification, regression, semisupervised, and unsupervised tasks. Due to its remarkable efficiency, impressive generalization performance, and easy extendibility, BLS has been applied in different domains. Next, we illustrate BLS's practical advances, such as computer vision, biomedical engineering, control, and natural language processing. Finally, the future open research problems and promising directions for BLSs are pointed out.

In the era of big data, mining the emotional tendency of opinions through natural language processing technology is meaningful for the timely understanding of human behavior data. Nowadays, Generalized Autoregressive Pre-training Language Modeling (XLNet) can not only capture bidirectional contextual knowledge but also learn the word dependency. However, its sentence-level representation didn't take broad features into account. In this paper, we design a novel architecture, called Broad Autoregressive Language Model (BroXLNet), to automatically process sentiment analysis task. BroXLNet integrates the advantage of generalized autoregressive language modeling and broad learning system, which has the ability of extracting deep contextual features and randomly searching high-level contextual representation in broad spaces. We evaluated our algorithm on binary Stanford Sentiment Treebank dataset. Compared with the state-of-the-art methods, e.g., BiLSTM, ELMo, OpenAI GPT, BERT and XLNet, BroXLNet achieved the best result of 94.0% in sentiment analysis task of binary Stanford Sentiment Treebank. The result demonstrated the excellent classifying ability of BroXLNet in sentiment analysis.