Cross-subject EEG decoding remains a fundamental challenge due to substantial inter-subject variability in brain activity, which hinders the development of subject-independent EEG models. Despite progress in extracting cross-subject invariant features, existing studies neglect the shared neural responses that arise under similar cognitive or emotional states across individuals, limiting their ability to learn generalized and consistent EEG representations. To address the challenges, we propose State Mamba, a novel spatiotemporal EEG state-space model that explicitly models and aligns neural responses and their spatiotemporal state transitions to learn consistent and generalizable representations across subjects. Innovatively, State Mamba theoretically formulates a multi-channel Mamba architecture that jointly models spatial and temporal brain state transitions, supporting principled analysis of neural responses. To enhance spatiotemporal feature coupling, we introduce the LGANN module, which adopts global-local attention to integrate long- and short-term brain activity into a compact EEG representation. Furthermore, we design two self-supervised pretext tasks to extract consistent neural patterns across subjects: (1) representation alignment to align EEG representation, and (2) pattern alignment to align their transition rules under identical conditions, jointly promoting subject-invariant EEG representations. Extensive experiments on three benchmark datasets, FACED, DEAP, and ISRUC, demonstrate the superior performance of State Mamba in cross-subject emotion and sleep recognition tasks, validating its robust generalization capability.

Unsupervised anomaly detection is to detect previously unseen rare samples without any prior knowledge about them. With the emergence of deep learning, many methods employ normal data reconstruction to train detection models, which is expected to yield relatively large errors when reconstructing anomalies. However, recent studies find that anomalies can be overgeneralized, resulting in reconstruction errors as small as normal samples. In this paper, we examine the anomaly overgeneralization problem and propose global semantic information learning. Normal and anomalous samples may share the same local feature such as textures, edges, and corners, but have separability at the global semantic level. To address this, we propose the global semantic memory with novel cascade architecture designed to capture global semantic information in the latent space and introduce a configurable sparsification and random forgetting mechanism. Our proposed method achieves state-of-the-art experimental results on different public benchmarks, without the introduction of any additional auxiliary loss.

Artificial Intelligence (AI) has significantly impacted healthcare, yet shifting applications from laboratory development to real-world deployment remains challenges: the inevitable scenario change leads to incompatible data inputs and potentially unreliable AI outputs. Human-centered monitoring and intervention during data collection should be reconsidered as a solution. By choosing Parkinson's wearable assessment as a scenario (existing application PD Assist, which failed in deployment), we conducted a user study and distilled the primary usability issues into three categories: understanding, timing, and disturbance. Accordingly, we redesigned an interactive version of PD Assist+, which contained active monitoring and intervention to ensure users' behavior meets expectations. Through evaluation involving 38 Parkinson's participants, our design increased accuracy by 16.9% and participants' use willingness by 34.8% compared to the original version. Our work derived awareness of data work in AI-based healthcare applications and proposed a pipeline to bridge the gap from development to deployment.

Cell fate decisions are orchestrated by intricate gene regulatory networks (GRNs), which govern gene expression with precise spatiotemporal control. However, accurately capturing context-specific nature of gene regulation remains challenging, particularly when integrating multi-omics data at bulk and single-cell level across diverse cellular contexts. Here, we present CellPolaris, a unified computational framework designed to decode the roles of transcription factors (TFs) in developmental processes. CellPolaris performs TF-centered GRN construction, master TF identification, and TF perturbation simulation. By leveraging transfer learning, the framework generates tissue-specific or cell-type-specific GRNs using pre-constructed high-confidence GRNs of diverse contexts and requires only transcriptomic data as input. Using these learned GRNs, CellPolaris identifies underlying master TFs critical for cell fate transitions and simulates the effects of TF perturbations on developmental processes. Benchmarking tests demonstrate the robust performance of CellPolaris in GRN construction. The efficacy of CellPolaris is supported by the significant overlap between predicted top-ranked master regulators and known TF combinations experimentally validated in cell fate conversion experiments. Furthermore, CellPolaris accurately simulates the developmental consequences of Rfx2 knockout during round spermatid differentiation. In summary, we present CellPolaris, a comprehensive framework that enables GRN construction through transfer learning, identification of key TFs driving cell fate transitions, and simulation of TF perturbations. This tool allows us to further elucidate the regulatory mechanisms underlying developmental processes and cell state transitions.

In recent years, unlearnable data have emerged as an effective way to prevent unauthorized exploitation of personal data by intentionally preventing model learning meaningful patterns during training, while ensuring that the processed data remain visually indistinguishable from the original. Previous approaches, such as error-minimizing noise methods, have yielded promising results in disrupting model learning but often lack robustness against adversarial training. More recently, convolution-based techniques have been introduced to address this limitation, enhancing resistance to adversarial training, but at the cost of sacrificing high-frequency image details. In this study, we propose a novel method called Wavelet Domain Steganography (WDS) that steganographically writes class-specific information in the wavelet transform domain. By leveraging the inherent multi-resolution and localized properties of wavelets, WDS allows for precise control over the steganographic information injected into different frequency subbands. The proposed method preserves high-frequency details and maintains visual fidelity while effectively inhibiting the learning of semantic information by deep neural networks. Experimental evaluations show that WDS exhibits robust unlearnability under common data augmentations and adversarial training scenarios, thereby achieving an improved balance between imperceptibility and unlearnability relative to prior methods.

The future of artificial intelligence (AI) systems lies in large-scale, heterogeneous agent collaboration to solve increasingly diverse and complex tasks. This collaborative paradigm promises greater flexibility and scalability but brings a fundamental challenge: how to balance collaboration efficiency with task quality. Existing Model Context Protocol (MCP) and Agent-to-Agent (A2A) architectures struggle to achieve this trade-off, facing either scalability bottlenecks or unstable coordination. To address these issues, we propose MoFedNet (Model Federalization Network), a novel semantic link-guided model collaboration system that integrates disparate models into a federated collective. MoFedNet features a dual-layer architecture with centralized monitoring and decentralized collaboration. Its core semantic link (SL) mechanism enables scalable, efficient, and dynamic collaboration among heterogeneous models while supporting intelligent task orchestration and network evolution. We enhance MoFedNet with three key modules. The Link-to-Link Semantic Protocol (L2L) organizes interactions over the hypergraph. The Contextual Memory Enhanced Retrieval Module (CoMER) introduces structured memory modules to store and retrieve representations across abstraction levels. The Evolutionary Local-global Optimization Module (EvoLOM) drives continuous improvement by accumulating memory and optimizing the hypergraph. We propose a novel quality-efficiency (QE) index to evaluate system performance. In a large-scale generic task simulation with over 15 000 nodes, MoFedNet achieves a QE score of 0.74, significantly outperforming the MCP and A2A approaches. Furthermore, when applied to a real-world medical task, MoFedNet attains a QE of 0.63, demonstrating clear superiority over existing multi-agent methods alongside stable temporal and spatial complexity.

Training deep learning models for photoplethysmography(PPG)-based cuff-less blood pressure estimation often requires a substantial amount of labeled data collected through sophisticated medical instruments, posing significant challenges in practical applications. To address this issue, we propose Physiological Knowledge-Aware Contrastive Learning (PhysCL), a novel approach designed to reduce the dependence on labeled PPG data while improving blood pressure estimation accuracy. Specifically, PhysCL tackles the semantic consistency problem in contrastive learning by introducing a knowledge-aware augmentation bank, which generates positive physiological signal pairs using knowledge-based constraints during the contrastive pair generation. Additionally, we propose a contrastive feature reconstruction method to enhance feature diversity and prevent model collapse through feature re-sampling and re-weighting. We evaluate PhysCL on data from 106 subjects across the MIMIC III, MIMIC IV, and UQVS datasets under cross-dataset validation settings, comparing it against state-of-the-art contrastive learning methods and blood pressure estimation models. PhysCL achieves an average mean absolute error of 9.5/5.9 mmHg (systolic/diastolic) across the three datasets, using only 2% labeled data combined with 98% unlabeled data for pre-training and 5 samples for personalization, which represents a 6.2%/4.3% improvement, respectively, over the current best supervised methods. The ablation study provides further convincing evidence that the unlabeled data can be utilized to improve the existing cuff-less blood pressure estimation models and shed light on unsupervised contrastive learning for physiological signals.

Text-to-image (T2I) diffusion models have achieved remarkable progress in generating realistic images from textual descriptions. However, ensuring consistent high-quality image generation with complete backgrounds, object appearance, and optimal texture rendering remains challenging. This paper presents a novel fine-grained pixel-level image editing method based on pre-trained diffusion models. The proposed dual-branch architecture, consisting of Guidance and Generation branches, employs U-Net Denoisers and Self-Attention mechanisms. An improved DDIM-like inversion method obtains the latent representation, followed by multiple denoising steps. Cross-branch interactions, such as KV Replacement, Classifier Guidance, and Feature Correspondence, enable precise control while preserving image fidelity. The iterative refinement and reconstruction process facilitates finegrained editing control, supporting attribute modification, image outpainting, style transfer, and face synthesis with Clickand-Drag style editing using masks. Experimental results demonstrate the effectiveness of the proposed approach in enhancing the quality and controllability of T2I-generated images, surpassing existing methods while maintaining attractive computational complexity for practical real-world applications.

Large Language Models (LLMs) have showcased exceptional capabilities in various domains, attracting significant interest from both academia and industry. Despite their impressive performance, the substantial size and computational demands of LLMs pose considerable challenges for practical deployment, particularly in environments with limited resources. The endeavor to compress language models while maintaining their accuracy has become a focal point of research. Among the various methods, knowledge distillation has emerged as an effective technique to enhance inference speed without greatly compromising performance. This paper presents a thorough survey from three aspects: method, evaluation, and application, exploring knowledge distillation techniques tailored specifically for LLMs. Specifically, we divide the methods into white-box KD and black-box KD to better illustrate their differences. Furthermore, we also explored the evaluation tasks and distillation effects between different distillation methods, and proposed directions for future research. Through in-depth understanding of the latest advancements and practical applications, this survey provides valuable resources for researchers, paving the way for sustained progress in this field.

Federated Hetero-Task Learning (FHTL) enables the simultaneous learning of multiple heterogeneous tasks on federated learning clients, offering enhanced flexibility. However, the inconsistency between optimization objectives and evaluation metrics for these heterogeneous tasks poses challenges in achieving performance fairness among clients. This study proposes a fairness-aware FHTL method, FairFHTL. It employs adversarial multi-task representation learning at the client level to learn the task-independent shared model. Consequently, it solves optimization objectives inspired by fair resource allocation on the server side to determine the update direction of the global shared model, ultimately achieving task-independent fair performance balance. Extensive experiments on three multi-task datasets demonstrate that FairFHTL significantly enhances performance across the majority of tasks compared to conventional federated learning and FHTL methods. Moreover, compared with other fairness-aware federated learning approaches, FairFHTL maintains a more uniform performance distribution across all tasks.

Wearable activity recognition is crucial for ubiquitous computing, enhancing human-machine interaction, medical monitoring, and personalized services. As wearable devices collect user activity data that often contain personal privacy information, federated learning (FL) is increasingly applied to protect user data privacy. However, in real-world scenarios, users' data are commonly exhibit heterogeneity, manifesting as non-independent and identically distributed (non-HD) characteristics, which presents challenges for FL methods. Traditional FL client selection approaches with heterogeneous data can cause global model drift, reducing the accuracy of activity recognition models. In this paper, we propose Data Contribution-Based Federated Learning (DCBFL) method, an adaptive FL training approach by selecting clients to counter the problem caused by heterogeneous data. Specifically, we first utilize a conditional generator on the server to construct an auxiliary dataset, which is used to train an auxiliary model as a benchmark to measure the degree of heterogeneity in each client's data. Furthermore, we reasonably differentiate the data contributions of clients based on the degree of data heterogeneity and select suitable clients for FL training, effectively utilizing heterogeneous data information, mitigating global model drift. The comprehensive experiments are conducted on five public activity recognition datasets under non-IID conditions in this work. The experimental results show that DCBFL outperforms existing baseline methods, showcasing superior performance.

BACKGROUND:The aim was to develop an attention-based model using 18F-fluorodeoxyglucose (18F-FDG) PET imaging to differentiate autoimmune encephalitis (AE) patients from controls and to discriminate among different AE subtypes. METHODS:This multi-center retrospective study enrolled 390 participants: 222 definite AE patients (comprising four subtypes: LGI1-AE, NMDAR-AE, GABAB-AE, GAD65-AE), 122 age- and sex-matched healthy controls, and 33 age- and sex-matched antibody-negative AE patients along with 13 age- and sex-matched viral encephalitis patients, both serving as disease controls. An attention-based multi-instance learning (MIL) model was trained using data from one hospital and underwent external validation with data from other institutions. Additionally, a multi-modal MIL (m-MIL) model integrating imaging features, age, and sex parameters was evaluated alongside logistic regression (LR) and random forest (RF) models for comparative analysis. RESULTS:The attention-based m-MIL model outperformed classical algorithms (LR, RF) and single-modal MIL in AE vs. all controls binary classification, achieving the highest accuracy (84.00% internal, 67.38% external) and sensitivity (90.91% internal, 71.19% external). For multiclass AE subtype classification, the MIL-based model achieved 95.05% (internal) and 77.97% (external) accuracy. Heatmap analysis revealed that NMDAR-AE involved broader brain regions, including the medial temporal lobe (MTL) and basal ganglia (BG), whereas LGI1-AE and GABAB-AE showed focal attention on the MTL and BG. In contrast, GAD65-AE demonstrated concentrated attention exclusively in the MTL. CONCLUSION:The m-MIL model effectively discriminates AE patients from controls and enables subtyping of different AE subtypes, offering a valuable diagnostic tool for the clinical assessment and classification of AE.

Large-scale neural networks have revolutionized many general knowledge areas (e.g., computer vision and language processing), but are still rarely applied in many expert knowledge areas (e.g., healthcare), due to data sparsity and high annotation expenses. Human-in-the-loop machine learning (HIL-ML) incorporates expert domain knowledge into the modeling process, effectively addressing these challenges. Recently, some researchers have started using large models to substitute for certain tasks typically performed by humans. Although large models have limitations in expert knowledge areas, after being trained on trillions of examples, they have demonstrated advanced capabilities in reasoning, semantic understanding, grounding, and planning. These capabilities can serve as proxies of human, which introduces new opportunities and challenges in HIL-ML area. Based on the above, we summarize a more comprehensive framework, Agent-in-the-Loop Machine Learning (AIL-ML), where agent represents both humans and large models. AIL-ML can efficiently collaborate human and large model to construct vertical AI models with lower costs. This paper presents the first review of recent advancements in this area. First, we provide a formal definition of AIL-ML and discuss its related fields. Then, we categorize the AIL-ML methods based on data processing and model development, providing formal definitions for each, and present representative works in detail for each category. Third, we highlight relative applications of AIL-ML. Finally, we summarize the current literature and highlight future research directions.

Sleep staging is critical for assessing sleep quality and tracking health. Polysomnography (PSG) provides comprehensive multimodal sleep-related information, but its complexity and impracticality limit its practical use in daily and ubiquitous monitoring. Conversely, unimodal devices offer more convenience but less accuracy. Existing multimodal learning paradigms typically assume that the data types remain consistent between the training and testing phases. This makes it challenging to leverage information from other modalities in ubiquitous scenarios (e.g., at home) where only one modality is available. To address this issue, we introduce a novel framework for ubiquitous Sleep staging via Supervised Multimodal Coordination, called SleepSMC. To capture category-related consistency and complementarity across modality-level instances, we propose supervised modality-level instance contrastive coordination. Specifically, modality-level instances within the same category are considered positive pairs, while those from different categories are considered negative pairs. To explore the varying reliability of auxiliary modalities, we calculate uncertainty estimates based on the variance in confidence scores for correct predictions during multiple rounds of random masks. These uncertainty estimates are employed to assign adaptive weights to multiple auxiliary modalities during contrastive learning, ensuring that the primary modality learns from high-quality, category-related features. Experimental results on four public datasets, ISRUC-S3, MASS-SS3, Sleep-EDF-78, and ISRUC-S1, show that SleepSMC achieves state-of-the-art cross-subject performance. SleepSMC significantly improves performance when only one modality is present during testing, making it suitable for ubiquitous sleep monitoring.

Current state-of-the-art paradigms predominantly treat Text-to-Motion (T2M) generation as a direct translation problem, mapping symbolic language directly to continuous poses. While effective for simple actions, this System 1 approach faces a fundamental theoretical bottleneck we identify as the Semantic-Kinematic Impedance Mismatch: the inherent difficulty of grounding semantically dense, discrete linguistic intent into kinematically dense, high-frequency motion data in a single shot. In this paper, we argue that the solution lies in an architectural shift towards Latent System 2 Reasoning. Drawing inspiration from Hierarchical Motor Control in cognitive science, we propose Latent Motion Reasoning (LMR) that reformulates generation as a two-stage Think-then-Act decision process. Central to LMR is a novel Dual-Granularity Tokenizer that disentangles motion into two distinct manifolds: a compressed, semantically rich Reasoning Latent for planning global topology, and a high-frequency Execution Latent for preserving physical fidelity. By forcing the model to autoregressively reason (plan the coarse trajectory) before it moves (instantiates the frames), we effectively bridge the ineffability gap between language and physics. We demonstrate LMR's versatility by implementing it for two representative baselines: T2M-GPT (discrete) and MotionStreamer (continuous). Extensive experiments show that LMR yields non-trivial improvements in both semantic alignment and physical plausibility, validating that the optimal substrate for motion planning is not natural language, but a learned, motion-aligned concept space. Codes and demos can be found in https://chenhaoqcdyq.github.io/LMR/https://chenhaoqcdyq.github.io/LMR/

Diabetic retinopathy (DR) is a leading cause of preventable low vision worldwide. Deep learning has exhibited promising performance in the grading of DR. Certain deep learning strategies have facilitated convenient regular eye check-ups, which are crucial for managing DR and preventing severe visual impairment. However, the generalization performance on cross-center, cross-vendor, and cross-user test datasets is compromised due to domain shift. Furthermore, the presence of small lesions and the imbalanced grade distribution, resulting from the characteristics of DR grading (e.g., the progressive nature of DR disease and the design of grading standards), complicates image-level domain adaptation for DR grading. The general predictions of the models trained on grade-skewed source domains will be significantly biased toward the majority grades, which further increases the adaptation difficulty. We formulate this problem as a grade-skewed domain adaptation challenge. Under the grade-skewed domain adaptation problem, we propose a novel method for image-level supervised DR grading via Asymmetric Bi-Classifier Discrepancy Minimization (ABiD). First, we propose optimizing the feature extractor by minimizing the discrepancy between the predictions of the asymmetric bi-classifier based on two classification criteria to encourage the exploration of crucial features in adjacent grades and stretch the distribution of adjacent grades in the latent space. Moreover, the classifier difference is maximized by using the forward and inverse distribution compensation mechanism to locate easily confused instances, which avoids pseudo-label bias on the target domain. The experimental results on two public DR datasets and one private DR dataset demonstrate that our method outperforms state-of-the-art methods significantly.

Electroencephalogram (EEG) is a non-invasive technique to record bioelectrical signals. Integrating supervised deep learning techniques with EEG signals has recently facilitated automatic analysis across diverse EEG-based tasks. However, the label issues of EEG signals have constrained the development of EEG-based deep models. Obtaining EEG annotations is difficult that requires domain experts to guide collection and labeling, and the variability of EEG signals among different subjects causes significant label shifts. To solve the above challenges, self-supervised learning (SSL) has been proposed to extract representations from unlabeled samples through well-designed pretext tasks. This paper concentrates on integrating SSL frameworks with temporal EEG signals to achieve efficient representation and proposes a systematic review of the SSL for EEG signals. In this paper, 1) we introduce the concept and theory of self-supervised learning and typical SSL frameworks. 2) We provide a comprehensive review of SSL for EEG analysis, including taxonomy, methodology, and technique details of the existing EEG-based SSL frameworks, and discuss the difference between these methods. 3) We investigate the adaptation of the SSL approach to various downstream tasks, including the task description and related benchmark datasets. 4) Finally, we discuss the potential directions for future SSL-EEG research.

AI-aided dermatological diagnosis models require large amounts of high-quality data for effective training. While mobile devices make data collection easier, the lack of professional annotations and privacy concerns hinder their use in model training Semi-supervised Federated Learning allows for collaborative training on both labeled and unlabeled data while protecting privacy. However, existing pseudo-labeling methods struggle with non-HD dermatosis data from different clients. To address this, we propose a multi-stage hierarchical semi-supervised federated learning framework (HSSFL), enabling labeled clients to assist unlabeled ones in generating pseudo-labels and maximizing data use. HSSFL consists of three stages: federated client clustering, iterative self-training, and opportunistic fine-tuning Using correlation awareness for federated client clustering, followed by an ally-prioritized approach for self-training, enhances the reliability of pseudo-labels. Opportunistic fine-tuning further minimizes label errors. Experiments on both general and real-world dermatological datasets demonstrate that HSSFL outperforms state-of-the-art semi-supervised federated learning methods.

Segmentation of ultra-high resolution (UHR) images is a critical task with numerous applications, yet it poses significant challenges due to high spatial resolution and rich fine details. Recent approaches adopt a dual-branch architecture, where a global branch learns long-range contextual information and a local branch captures fine details. However, they struggle to handle the conflict between global and local information while adding significant extra computational cost. Inspired by the human visual system's ability to rapidly orient attention to important areas with fine details and filter out irrelevant information, we propose a novel UHR segmentation method called Boundary-enhanced Patch-merging Transformer (BPT). BPT consists of two key components: (1) Patch-Merging Transformer (PMT) for dynamically allocating tokens to informative regions to acquire global and local representations, and (2) Boundary-Enhanced Module (BEM) that leverages boundary information to enrich fine details. Extensive experiments on multiple UHR image segmentation benchmarks demonstrate that our BPT outperforms previous state-of-the-art methods without introducing extra computational overhead.