EEG-JEPA

Most EEG foundation models are pretrained by reconstructing masked segments of the recorded waveform. Scalp EEG, however, has a low signal-to-noise ratio, and much of what is predictable in it is background rhythm, acquisition idiosyncrasy, and physiological artifact — all of which a reconstruction objective rewards recovering just as much as the neural structure that transfers to new tasks. EEG-JEPA, from the Tsinghua Laboratory of Brain and Intelligence at Tsinghua University with a collaborator at Peking University, reframes this as a question of objective design: not how much EEG to pretrain on, but what an EEG foundation model should be trained to predict.

The answer is a joint-embedding predictive architecture. A masked context encoder and a predictor infer the contextual latent states produced by an exponential-moving-average target encoder that sees the complete recording crop, so supervision never touches raw voltage. The framework organizes latent-target design along three axes: target content (what representation is predicted), target support (which electrode-time positions must be inferred), and target depth (at which encoder layers prediction is enforced). That places EEG-JEPA alongside CBraMod, LaBraM, CSBrain, and LUNA, but supervising in representation space rather than signal space. The work was posted as an arXiv preprint on 31 July 2026 and has not yet been peer reviewed.

#Key Features

• Contextual latent targets: An EMA-updated target encoder observing the full crop supplies stop-gradient reference states, which the context encoder and predictor infer from visible context plus a position query that reveals where a target sits but not what it contains.
• Neurotopology-aware masking (N-MET): Targets are sampled as structured regions — temporal stripes, channel stripes, local blocks, bilateral topology pairs, and whole dropped channels — rather than independent patches, so predictions cannot be solved from immediate neighbors alone.
• Hierarchical multi-depth supervision: Separate prediction heads supervise layers 3, 6, 9, and 12, contributing the single largest ablation gain (+3.66 points).
• Nested parameter-efficient encoders: Because intermediate depths are directly supervised, the L3, L6, and L9 prefixes of one pretrained model are usable on their own at 1.26M, 2.47M, and 3.68M parameters, giving an accuracy-size trade-off without retraining.
• Frozen multitask evaluation: The backbone is held fixed, normalization statistics included, while only small per-task heads train — a stricter transfer test than per-dataset fine-tuning.

#Technical Details

EEG-JEPA adopts CBraMod's fixed 19-channel tokenization and spatio-temporal encoder as its backbone, so objective comparisons hold architecture, corpus, and training budget constant. Recordings are resampled to 200 Hz, mapped to a canonical 19-channel montage, and cut into 30-second crops of non-overlapping 1-second patches. Stage 1 trains for 100 epochs on TUEG (39,758 records, 6.54 million crops); Stage 2 continues for 25 epochs on a 75/15/10 mixture of TUEG, TDBRAIN, and HBN. N-MET selects 40–55% of valid tokens per crop. Training minimizes a Smooth L1 latent prediction loss across the four supervised depths, plus context-consistency and variance-covariance regularization terms.

On the 14-task EEG-FM-Bench frozen protocol, the 4.92M-parameter encoder reaches 52.94% macro balanced accuracy, ahead of EEGPT at 52.15% (25.29M parameters), REVE at 51.50% (69.19M), CSBrain at 45.44%, LaBraM at 44.91%, and the parameter-matched CBraMod reconstruction baseline at 40.49%. Progressive ablation traces that gap: latent targets lift the baseline to 45.12%, N-MET to 46.47%, hierarchical prediction to 50.13%, context consistency to 50.42%, and multi-source continuation to 52.94%. Under protocol-matched full fine-tuning, the nine-task average balanced accuracy rises from 68.98% to 70.65%, with the largest gains on imagined speech (+4.13), motor imagery (+4.00), and mental stress (+3.50).

#Applications

The downstream suite spans the practical range of EEG analysis: clinical abnormality detection and event typing on the Temple University corpora, sleep staging on ISRUC, motor imagery on BCI-2a and PhysioMI, emotion recognition on FACED, mental-stress detection, imagined speech decoding, and depression screening. Because a single frozen encoder serves all of them through lightweight heads, groups without the data or compute to pretrain can simply attach a task head to a fixed representation. The nested prefix encoders matter most for embedded and wearable brain-computer interfaces, where a 1.26M-parameter model clearing 48% macro balanced accuracy fits a tight compute budget.

#Impact

EEG-JEPA's contribution is evidence that pretraining objective, not scale, is the binding constraint for EEG foundation models: a 4.92M-parameter encoder outperforms a 69M-parameter one on frozen transfer, and matched ablations attribute the improvement to specific design choices rather than to more data. Layerwise probing supports the mechanism, showing that cross-channel and interaction information becomes accessible in deeper layers under latent prediction but stays front-loaded under masked reconstruction. Three caveats bound the result. The paper is a preprint awaiting peer review. The public repository ships training, fine-tuning, and preprocessing scripts but no downloadable pretrained weights, so the frozen-transfer workflow cannot yet be reproduced without running pretraining. And the repository declares no license of its own — only a NOTICE crediting the upstream MIT terms on CBraMod-derived code — leaving reuse rights for the original contributions unspecified.