FUSED: Foundation Model Guided EEG Decoding Framework

Researchers have introduced FUSED, a framework for source-free EEG decoding that is guided by a foundation model. This innovative approach combines a large-scale foundation model with a specialized compact model through dual-branch co-adaptation. FUSED aims to overcome the shortcomings of current source-free domain adaptation (SFDA) techniques, which depend on source-pretrained models and often exhibit inadequate cross-domain generalization and unreliable pseudo-labels. By utilizing EEG foundation models trained on extensive datasets, FUSED implements a co-adaptation strategy featuring linear and prototype views for generating cross-branch pseudo-labels, along with a consensus filtering mechanism to harness the foundation model's insights. This research is available on arXiv with the ID 2605.00857.