EEG Foundation Model Interpretability
An end-to-end EEG machine-learning pipeline for preprocessing, adapting, and evaluating CBraMod on the SEED-VIG dataset.
Overview
Research into how pretrained EEG transformers represent neurological states, using sparse feature analysis and activation-level interpretation.
Highlights
- Benchmarked linear probing, partial fine-tuning, LoRA, and full fine-tuning with subject-independent validation.
- Analyzed 153,600-dimensional representations across 12 transformer layers using hypothesis tests, effect sizes, and permutation testing.
- Investigated Cross-Layer Transcoders and sparse features associated with drowsiness.
Tools and methods