EEG-JEPA: Structured Latent Prediction for EEG Foundation Models
A novel framework for EEG foundation models, named EEG-JEPA, has been presented in a paper on arXiv (2608.00114). This framework tackles a key issue in EEG modeling: determining what a foundation model should predict to develop transferable representations. In contrast to typical pretraining methods that aim to reconstruct masked waveforms, EEG-JEPA employs a masked context encoder and predictor to deduce contextual latent states generated by an exponential-moving-average target encoder that analyzes the entire input. The design of the target is structured across three complementary aspects: target content indicates the predicted representation. This announcement is cross-type and can be accessed at https://arxiv.org/abs/2608.00114.
Key facts
- EEG-JEPA is a structured latent-prediction framework for EEG foundation modeling.
- It is introduced in a paper on arXiv with ID 2608.00114.
- The framework avoids masked waveform reconstruction, instead predicting latent states.
- It uses a masked context encoder and predictor.
- A target encoder with exponential moving average observes the complete input.
- Target design includes three dimensions: target content, and others.
- The paper is a cross-type announcement.
- The source URL is https://arxiv.org/abs/2608.00114.
Entities
Institutions
- arXiv