Subject-aware contrastive learning for EEG foundation models
File(s)
Accepted version
Author(s)
Type
Conference Paper
Abstract
Foundation models are beginning to reshape EEG representation learning, but existing approaches remain dominated by self-supervised reconstruction objectives. In this work, we introduce the first subject-aware contrastive EEG foundation model, leveraging subject identity as a natural supervisory signal. Building on a patch-based architecture inspired by recent Large Brainwave Foundation Models (LBMs), we pretrain a lightweight transformer encoder using contrastive learning, where positive pairs are drawn from different segments and sessions of the same subject. Unlike contrastive foundation models in other domains, which depend on augmentations to construct positive samples, our method relies on naturally occurring intra-subject variability across EEG sessions. We evaluate the model through both representation metrics (alignment, uniformity and smooth effective rank) and downstream tasks (under linear probing and full fine-tuning). Results show that our model produces well-structured representation spaces, achieving strong representation quality and competitive performance compared to other LBMs.
Date Issued
2025-09-23
Date Acceptance
2025-10-15
Citation
2025
Publisher
OpenReview
Source
NeurIPS 2025 Workshop on Learning from Time Series for Health
Publication Status
Published online
Start Date
2025-12-07
Coverage Spatial
San Diego, CA, USA