EEG foundation models have become a popular way to learn general representations from large-scale multi-subject data, which can then be adapted to specific downstream tasks. However, according to a new preprint, the embeddings these models produce after pretraining are not clean: they still retain substantial unwanted information that can interfere with fine-tuning.
The paper, posted on arXiv, proposes a method called DEFT, or Disentanglement-Enhanced Fine-Tuning, to tackle this problem. Rather than modifying the pretraining stage, DEFT focuses on the fine-tuning stage, aiming to separate the useful signal from the residual noise in the embeddings.
The abstract is truncated in the source, so details on how DEFT works and its experimental results are not available from the provided text. The paper's contribution, as stated, is the introduction of this fine-tuning strategy for EEG foundation models. No comparisons to other methods or quantitative outcomes are reported in the available excerpt.