Decoding human vision from neural signals in brain-computer interfaces (BCIs) represents a longstanding research objective in neuroscience and machine learning. Recently, electroencephalography (EEG) has attracted significant interest due to its non-invasive nature and practical accessibility, thereby offering a viable pathway for visual decoding and reconstruction. In this study, we propose an end-to-end Fine-Grained EEG-Vision Alignment learning framework (FGEVA) for visual decoding and reconstruction. The framework consists of an EEG Wavelet Transformer Encoder (EEG-WTE) and a fine-grained EEG-vision alignment learning process. Among them, EEG-WTE leverages the advantages of wavelet transformation to activate low-frequency information for comprehensive EEG features. During alignment learning, the input EEG signal is processed by our hypergraph-based EEG Brain-region Feature Decomposer (EBFD) to obtain global and local brain-region EEG features, while the corresponding image is processed by our graph-based Visual Brain-region Feature Decomposer (VBFD) to obtain global and local brain-region visual features. Similarly, the corresponding input image is also processed by our proposed graph-based Visual Brain-region Feature Decomposer (VBFD) for the global and local brain-region visual features. Finally, the derived features are aligned through regional one-to-one mapping. This design enables both global and local fine-grained alignment between EEG signals and image object representations within the shared mapping space. Extensive experiments show that our FGEVA achieves better results in both visual decoding and reconstruction. The code is available at https://github.com/sol924/FGEVA.
This paper presents a Reinforcement Learning–Guided Hyperheuristic (RL–HH) for dynamic task scheduling in heterogeneous edge–cloud systems. A Deep Q-Network (DQN) policy observes real-time system feedback (latency, energy, and load balance) and selects among low-level heuristics, Min-Min, Max-Min, Shortest Job First, Round-Robin, and a Resource-Aware Greedy rule, to construct schedules that adapt to workload and network variability. Evaluated on synthetic workloads and Google Cluster traces, RL–HH reduces average task response time by up to 37% versus Round-Robin and 21.6% versus a Genetic Algorithm, lowers energy consumption by 28% and 15%, respectively, and improves load balance by up to 65%, while incurring only 5.1 ms heuristic-switching overhead. Ablation studies confirm that (i) the RL controller, (ii) the diversity of the heuristic portfolio, and (iii) explicit reward shaping for load balance are all critical to the observed gains. These results demonstrate that learning-based selection over a portfolio of simple heuristics provides robust, energy-aware, and scalable performance for next-generation distributed edge–cloud infrastructures.
Solving the feature selection problem in incomplete multi-label environments with missing features is a significant challenge in data mining. In response to this, we propose a new method for incomplete multi-label feature selection with relational embedding and tolerance compensation, called RETC. To fully exploit the various relationships in the data and effectively guide feature selection, the RETC model adopts a dual quantification strategy to quantify and embed these relationships into the model. First, the model explores the relationships between features using the relative feature interaction coefficient. This coefficient not only accounts for feature redundancy but also uniquely reveals the positive interactions between features. Next, the tolerance implication granularity mutual information is proposed to estimate the strength of correlation between features and labels. Additionally, a two-layer mechanism is designed to handle missing features. First, the tolerance relation is introduced to process incomplete data. Then, a tolerance compensation strategy is designed to mitigate the effects caused by the looseness of the tolerance relation. Specifically, a non-negative sparse tolerance compensation matrix is applied to further refine the computed feature–label correlation, ensuring that the feature selection process is more reliable and stable. Finally, the effectiveness and superiority of the RETC model are verified through comparative experimental analysis on multiple datasets.
Graph Neural Networks (GNNs) often experience performance degradation when training and test data follow different distributions, which limits their reliability in real-world applications. To address this problem, we propose Neighbor Environment Risk Extrapolation (NERE), an invariant learning framework that generates virtual environments through topology-aware uncertainty modeling. Unlike existing methods that rely on predefined environment partitions or fixed graph structures, NERE first applies global structure perturbations to simulate coarse distributional variations, and then models uncertainty within local neighborhoods to generate fine-grained environment views. By constraining the generated features using neighborhood statistics, NERE reduces the risk of producing unrealistic extrapolations. It further penalizes empirical risk variance across the generated environments, encouraging the GNN to learn representations that are less sensitive to environment-specific correlations. Experiments on multiple graph OOD benchmarks show that NERE improves OOD generalization under both covariate and concept shifts, while introducing limited additional computational overhead for scalable node classification.
Deep Reinforcement Learning (DRL) excels in complex control tasks, but its “black-box” nature hinders deployment in safety-critical domains. Current Explainable AI relies on post-hoc methods (e.g., saliency maps) that function as learned approximations rather than exact representations, often fostering trust that is persuasive without being faithful. A critical gap remains: the need for inherently interpretable, directly auditable models. Trustworthy deployment requires that these models be formally verifiable without sacrificing operational competence. This work addresses this by introducing a depth-constrained, value-aware policy distillation framework, motivated by rate-distortion theory, that treats tree depth as a hard capacity constraint on the decision logic extracted from a deep neural network into an auditable decision tree. Experimental validation on LunarLander-v3 demonstrates that distilled trees achieve functional competence (reward = 103.9) despite moderate behavioral fidelity (44%) by discarding implementation-specific policy variance. Comparative evaluation against Behavioral Cloning and VIPER reveals that value-function fidelity (Fs = 0.97) correlates more strongly with competence than action agreement, validating value-based distillation over pure imitation. These findings offer a step toward more auditable AI systems in high-stakes environments, though further validation would be needed before supporting formal regulatory certification. To facilitate reproducibility, our code is available publicly.
