laya-eeg-foundation

name	laya-eeg-foundation
description	Laya: A LeJEPA approach to EEG via Latent Prediction over Reconstructed Activity. Self-supervised EEG foundation model using Joint Embedding Predictive Architecture (JEPA) with latent-space prediction for EEG representation learning. Enables transfer learning across EEG tasks without labeled data. Activation: EEG foundation model, self-supervised EEG, JEPA, EEG pretraining, EEG representation learning, brain signal embedding, 脑电基础模型, 自监督脑电
version	1.0.0
metadata	{"hermes":{"source_paper":"Laya: A LeJEPA Approach to EEG via Latent Prediction over Reconstructed Activity","arxiv_id":"2603.16281","tags":["eeg","foundation-model","self-supervised","japa","representation-learning"]}}

Laya EEG Foundation Model

Overview

Self-supervised EEG foundation model using LeJEPA (Latent JEPA) architecture. Predicts latent representations of future/reconstructed EEG activity, learning general-purpose neural embeddings transferable across BCI, clinical, and cognitive neuroscience tasks.

Architecture

EEG Input → Encoder → Latent z_t
                      ↓
              Predictor → ẑ_{t+Δ}
                      ↑
        Context Encoder → z_{t+Δ}

Key Design Principles

1. Latent-Space Prediction: Predict future latents rather than raw EEG, avoiding blurry reconstructions
2. Multi-Scale Temporal Context: Capture both short-term dynamics and long-range dependencies
3. Subject-Invariant Features: Learn representations that generalize across individuals
4. Channel-Agnostic Handling: Robust to varying EEG channel configurations

Training Pipeline

class LeJEPAEEG:
    def __init__(self, encoder_dim=256, predictor_dim=128):
        self.encoder = EEGEncoder(dim=encoder_dim)
        self.context_encoder = EEGEncoder(dim=encoder_dim)  # stop-grad
        self.predictor = Predictor(encoder_dim, predictor_dim)
    
    def forward(self, eeg_target, eeg_context):
        z_target = self.encoder(eeg_target)          # no grad through context
        z_context = self.context_encoder(eeg_context) # stop-gradient
        z_pred = self.predictor(z_context)
        loss = mse_loss(z_pred, z_target.detach())
        return loss

Applications

• Zero-shot EEG classification
• Transfer learning for BCI
• Cross-subject generalization
• Clinical EEG anomaly detection

Related Skills

• eeg-ieeg-bridge, meta-learning-in-context-brain-decoding, eeg-foundation-model-adapters