MB2L: A Biomimetic Framework for EEG-Based Visual Decoding

Researchers propose MB2L, a Multi-Level Bidirectional Biomimetic Learning framework, to improve EEG-based visual neural decoding. The framework addresses the mismatch between digital images and biological visual perception by incorporating physiological inductive biases. Key components include Adaptive Blur with Visual Priors, which reweights visual inputs based on retinotopic priors, and Biomimetic Visual Feature Extraction, which learns multi-level visual representations consistent with hierarchical cortical processing. These modules are jointly optimized via multi-level bidirectional learning to enhance subject-invariant encoding and cross-modal alignment. The approach aims to overcome limited paired data and neuroanatomical variability in tasks such as image retrieval.