EEG-Based Visual Decoding

EEG-based visual decoding asks what visual content a person has seen or imagined, ranging from choosing an image from a candidate set to reconstructing 2D or 3D scenes. Common approaches encode EEG’s temporal, spectral, and spatial patterns, then align them with pretrained vision or vision-language representations using contrastive learning; diffusion and other generative models can turn aligned signals into images. Evaluation often uses THINGS-EEG and THINGS-EEG2 for within-subject, cross-subject, and zero-shot retrieval, with newer tests probing fine-grained edits and dynamic perception. A central challenge is that noisy, person-specific signals can support semantic retrieval more reliably than faithful detail, so cross-subject generalization and evaluation of perceptual versus semantic fidelity remain open issues.