Brain Signal Rendering: Unifying EEG Video Representations for Subject-level Few-shot Learning

ICLR 2026 Conference Submission15953 Authors

Keywords: EEG, Brain-Computer Interface

TL;DR: Rendering brain signals into videos and large model enpowered EEG few-shot learning

Abstract: EEG modeling faces two core challenges: nonlinear, non-stationary dynamics and severe channel mismatch across datasets. We introduce Brain Signal Rendering (BSR), a new paradigm that reframes EEG representation learning as a rendering problem. BSR transforms EEG spectrograms into spatialized dynamic 'EEG videos', making representations invariant to electrode layouts and sampling protocols while preserving neural topology. Building on this, we propose EEG Consolidation — a unified multi-task training paradigm that integrates heterogeneous EEG-video data to adapt models to EEG-specific dynamics, improve data efficiency, reduce overfitting, and boost cross-task generalization. Crucially, BSR with EEG Consolidation enables subject-level few-shot learning, where each subject is treated as a distinct task requiring adaptation from minimal data. We validate this setting as a realistic benchmark and demonstrate substantial performance gains, establishing a scalable and interpretable framework toward foundation models for brain signals.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 15953