Hello Worlds

Geometry-aware EEG representation learning

Finalist - Hack the World(s), June 2026

24-hour hackathon on JEPAs and world models, sponsored by Yann LeCun.

Code | Final deck | Research handoff

Team: Florent Tariolle | Clement Genninasca | Yoann Frayce | Hippolyte du Pac

This repository is the final artifact of a 24h geometry-aware EEG representation-learning project: a strong SIGReg baseline, and evidence that SPD-tangent geometry makes the learned latent structure more visible. We pretrain on unlabeled TUAB recordings, freeze the encoder, and fit a linear probe for normal-versus-abnormal EEG on a patient-disjoint evaluation split.

What we found

• Performance: the in-domain SIGReg baseline reaches 0.819 balanced accuracy and roughly 0.89 AUROC across three seeds.
• Geometry: in the controlled ambient-versus-SPD-tangent and SIGReg-versus-PEIRA comparison, every cell is near 0.82. The tangent-space variant does not give a reliable frozen-probe gain, but AIRM-aware visualization exposes latent structure that a Euclidean view can obscure. This is qualitative geometric evidence, not a performance claim.

The visualization below is computed from frozen SPD latent covariances on TUSZ seizure-type events. Labels are applied only afterwards for interpretation: the AIRM-aware view makes the learned organization more apparent, but it is not a substitute for held-out metrics. The AIRM Riemannian t-SNE visualization follows Thibault de Surrel's r_tSNE_in_C work.

This is a symmetric two-view invariance objective with an explicit anti-collapse regularizer. It is JEPA-inspired, but it is not an EMA-teacher I-JEPA/V-JEPA implementation and it does not train a temporal predictor.

Retained core implementation

The final artifact is the two-view TUAB/TUSZ/TUEV pipeline below - not the archived Graph-JEPA or Fourier-JEPA explorations.

• Training pipeline: two-view SSL pretraining and the ambient-versus-SPD-tangent regularization switch.
• Frozen evaluation and TUAB baselines/benchmark.
• EEG encoder and SIGReg/BCS loss.
• PEIRA and SPD-tangent geometry.
• AIRM/tangent visualization, with matching TUSZ and TUEV paths.

Reproduce the retained TUAB path

uv venv
uv pip install -e .

DATA=<TUAB_PREPROCESSED>

# CPU baselines
python -m examples.eeg.baseline_riemann --data-root $DATA
python -m examples.eeg.baseline_chanstats --data-root $DATA --n-windows 8

# Two-view SSL pretraining and frozen evaluation
python -m examples.eeg.main --fname examples/eeg/cfgs/train.yaml
python -m examples.eeg.eval --ckpt ./checkpoints/<run>/latest.pth.tar --floor

# Rebuild benchmark figures and the deck
python -m examples.eeg.benchmark
cd presentation && make

For TUEV/TUSZ transfer and geometry commands, see the corresponding module docstrings and the notes in docs/. Dataset files, checkpoints, scheduler logs, and generated experiment runs are intentionally not versioned.

Project status

This is a frozen hackathon artifact. Manifold-JEPA is the separate repository for the follow-up anomaly-detection research; this repository remains the provenance-preserving record of the 24h project.

Experimental contract

Regularizer	Ambient Euclidean	SPD tangent
VICReg	C0 reference	-
SIGReg	C1 baseline	C2 geometry-aware
PEIRA	C3	C4 pre-registered hypothesis

• Data: TUAB, 2,717 train and 276 evaluation recordings; the split is patient-disjoint.
• Protocol: SSL uses no labels during pretraining; the frozen linear probe is trained only after representation learning.
• Headline metric: three-seed balanced accuracy. The best single seed, 0.833, is not the reported result.
• Controls: random-init encoder floor (~0.79), supervised baseline, channel mean/std baseline, and a classical Riemannian covariance baseline.

Secondary evidence and caveats

• TUSZ to TUAB: a general-TUH pretraining run reaches 0.814 balanced accuracy and 0.889 AUROC when frozen and probed on TUAB. This is one seed and same-site evidence, not a broad cross-site claim.
• TUEV: the repository retains a conventional SIGReg train/probe pipeline and Riemannian visualizations as a secondary cross-task track. Its relationship to the TUAB pretraining split remains caveated; it is not a headline result.
• Visualization: labels are applied after embedding for interpretation only. t-SNE/UMAP-style plots never substitute for held-out metrics, leakage checks, or seed variation.

Repository layout

eb_jepa/                 minimal vendor: EEG dataset and anti-collapse losses
examples/eeg/            retained TUAB/TUSZ/TUEV training, probes, geometry, and figures
cluster/                 retained Dalia/SLURM launch recipes
results/                 curated core measured tables and figures
presentation/            final deck source and compiled PDF
docs/                    positioning, geometry notes, and research handoff
references/              short literature notes
archive/                 explicitly non-headline Graph-JEPA and Fourier-JEPA explorations

Archived explorations

Graph-JEPA and Fourier-JEPA are preserved as self-contained explorations. They are not part of the controlled conclusion above and should not be read as competing headline results.

Honest positioning

The repository does not claim a foundation model, SOTA, or that geometry improves TUAB frozen accuracy. It records a reproducible negative result with proper controls and preserves the strongest follow-up question: where, if anywhere, geometry improves EEG representation quality beyond this narrow frozen-probe setting.

See docs/positioning.md, docs/geometry_tangent_analysis.md, and docs/research_handoff.md for the detailed claims and next-step constraints.