Reproduce “EEG-JEPA: Structured Latent Prediction for EEG Foundation Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinhao Li, Zhiyuan Ma, Xueqiao Han, Zhongye Xia, Xinche Zhang, Shanghong Xie, Yixuan Liu, Yongjian Li and 3 more
EEG-JEPA learns transferable EEG features by predicting a slowly updated teacher’s representations of hidden electrode-and-time regions from visible signals, rather than reconstructing missing voltages. It structures targets across masking patterns and encoder depths. On EEG-FM-Bench’s 14-task frozen-transfer evaluation, Stage-1 reaches 50.42% macro balanced accuracy versus 40.49% for a matched waveform-reconstruction control; the encoder stays fixed, but task-specific heads are trained.
Key insights
- Predicting structured latent states, particularly those with neurotopology-aware support and hierarchical supervision, effectively overcomes limitations of direct waveform reconstruction, yielding more transferable representations.
- Hierarchical supervision encourages a clear organization of learned features, with deeper layers specializing in complex within-signal interactions and cross-channel relations, crucial for diverse EEG tasks.
- The framework's design allows for parameter-efficient transfer, enabling the use of smaller, intermediate encoder prefixes as pre-trained models, providing flexible accuracy-size trade-offs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.