SUPA-EEG: Scale-Unified Parieto-occipital Architecture for EEG Data

Lydia Liya He, Huy Hai Nguyen, Finn Beckmann, Natthanan Bhukan

Keywords: Electroencephalography, Encoding, Zero-Shot Visual Stimuli Decoding, Object Classifcation

TL;DR: Zero-shot EEG-based visual decoding via multi-scale representation alignment, reflecting human neural recognition steps.

Abstract: Electroencephalography (EEG) is a non-invasive spatio-temporal recording of cerebral cortical activity with the ability to measure brain waves as patients view various image categories. Decoding visual perception by concurrent EEG-probes offers a path towards applications including brain-computer interfaces, restorative prosthetics and cognitive-load monitoring. Several limitations persist across existing methods, including aligning EEG representations only to the final CLIP layer, discarding low- and mid-level visual structure preserved in neural signals, and treating all EEG channels equally, ignoring both spatial relevance and multi-scale structure. We propose Scale-Unified Parieto-occipital Architecture for EEG (SUPA-EEG) for zero-shot visual retrieval on THINGS-EEG, inspired by human neural hierarchical visual processing to address existing limitations. SUPA-EEG unifies scale-aware channel weighting, multi-scale spatiotemporal representation learning, and intermediate feature alignment within a single framework, producing stable and interpretable EEG embeddings for zero-shot visual decoding.

Submission Number: 6