Go to ICLR 2026 Conference homepageBrain-Mimetic Staged Representation Learning with Disentangled Coarse and Fine Semantic for EEG Visual Decoding
Keywords: EEG visual decoding, brain-mimetic staged representation learning, dual-level coarse-to-fine semantics, virtual EEG channels
Abstract: Decoding visual information from electroencephalography (EEG) signals remains a fundamental challenge in brain–computer interfaces and medical rehabilitation. Most existing methods focus on refining EEG encoders to obtain stronger EEG embeddings for alignment with visual features, but they largely overlook that human visual perception is inherently staged, progressing from low-level feature detection to high-level semantic abstraction and ultimately to information integration. Inspired by neuroscientific theories of staged vision, we propose a novel EEG representation learning framework that explicitly models the three stages of brain visual processing: Phase-I for low-level visual representation learning, Phase-II for high-level semantic representation learning, and Phase-III for integrative information fusion. To further enhance semantic modelling, we propose (i) a multimodal dual-level semantic learning mechanism, which disentangles coarse label-level semantics and fine image-level semantics from visual EEG channels, and (ii) a new concept of virtual EEG channels, which expand the representational capacity of EEG signals. Extensive experiments on the largest benchmark dataset demonstrate significant improvements over state-of-the-art methods under both subject-dependent and subject-independent zero-shot settings, confirming both robustness and generalisability of our method. By explicitly modelling staged brain-mimetic processing and dual-level enriched semantic representations, our work not only advances decoding performance but also provides a biologically grounded perspective for future EEG-based brain decoding research.
Primary Area: applications to neuroscience & cognitive science
Submission Number: 22210
Loading