Aligning EEG with text for decoding imagined speech
A new framework maps EEG trials to language representations to decode 39 intent categories.
Researchers developed a contrastive learning framework to decode imagined speech. This method aligns EEG trials with language representations in the embedding space of a frozen pretrained text encoder.
The approach uses an EEG encoder trained with a composite objective. This objective combines semantically reweighted EEG-sentence alignment and text-prototype-based class discrimination.
The framework was evaluated on the Chisco imagined speech dataset across five subjects. It measured Top-k accuracy, macro-F1, and balanced pairwise matching accuracy (2v2).
Maximum EEG-text cosine similarity, or MaxCosine, was used as an indicator of decoding reliability. This metric identified subsets of held-out trials with higher Top-1 and Top-5 accuracy.
The findings suggest that EEG-language alignment can retain fixed-intent decoding. It also supports similarity-based retrieval and the evaluation of broader semantic structure.
Key facts
- The framework aligns EEG trials with language representations.
- The method was tested on the Chisco imagined speech dataset using five subjects.
- The approach preserves separability among 39 intent categories.
- MaxCosine can identify trials with higher Top-1 and Top-5 accuracy.