EEG Foundation-Model probing

Goal: reuse a frozen pretrained EEG backbone and fit only a light linear/MLP probe on labelled MI data — i.e. no full retraining. This is how we'll test the 32-ch cap data later, and it's how we get an apples-to-apples comparison against the pyRiemann baselines.

Where things run

M5 Mac (MPS): fine for inference / linear-probing of the small models (EEGPT ≈ 10M params, CBraMod). Use device="mps".
GPU server: needed only for fine-tuning the backbone or pretraining.

Setup

bash src/foundation/setup_foundation.sh   # clone repos + fetch weights

Repos land in third_party/, weights in checkpoints/.

Candidates (all channel-flexible, weights public)

Model	Repo	Input it expects	License
EEGPT	`BINE022/EEGPT`	58ch/256Hz/4s, patch 64	see repo
CBraMod	`wjq-learning/CBraMod`	(ch, seg, 200 pts/patch)=200Hz	MIT
LaBraM	`935963004/LaBraM`	200Hz, patch, channel names	see repo

Probe recipe (identical across models)

Pull MI epochs from MOABB (same LeftRightImagery paradigm as the baseline) so the comparison is fair.
Resample to the model's native rate (EEGPT 256 Hz, CBraMod/LaBraM 200 Hz); map our channel names to the model's channel embedding.
Forward pass through the frozen backbone → per-trial embedding.
Fit LogisticRegression (or a 1-layer MLP) on embeddings; evaluate cross-subject exactly like the baseline.

probe_moabb.py implements this loop; encoders live in encoders.py.

Status (cross-subject 2a, ROC-AUC — see ../../results/RESULTS.md)

encoder	wired?	2a AUC	notes
`raw`	✅	0.521	floor
`cbramod`	✅ (MPS)	0.560	weights in `checkpoints/CBraMod/`
`labram`	✅ (MPS)	0.595	weights ship in `third_party/LaBraM/checkpoints/`
`eegpt`	⬜	—	needs the figshare checkpoint (below)

Run: python src/foundation/probe_moabb.py --dataset 2a --subjects 9 --encoder labram --fmin 0.5 --fmax 45 (set PYTORCH_ENABLE_MPS_FALLBACK=1 for the transformer ops MPS doesn't cover).

All frozen probes underperform the classic Riemannian baseline (~0.78) — expected: MI is a subtle spatial ERD contrast that a frozen generic backbone doesn't expose linearly. FMs need fine-tuning (not frozen probing) to compete.

Gotchas already handled (in encoders.py)

Pool over time only, keep channels — averaging over channels drives MI to chance.
Scale = µV/100 (scale=0.01); MOABB returns µV-scale here. Feed broadband 0.5–45 Hz.
LaBraM + timm 1.x: _timm1x_shim() aliases the old timm.models.layers/.registry.
PyTorch 2.6 torch.load: LaBraM ckpt needs weights_only=False (trusted repo file).

EEGPT (not yet wired — needs manual download)

Checkpoint is on figshare (blocked from automated fetch): https://figshare.com/s/e37df4f8a907a866df4b → eegpt_mcae_58chs_4s_large4E.ckpt (58 ch, 256 Hz, 4 s, patch 64). Download it into checkpoints/, then mirror third_party/EEGPT/downstream/linear_probe_EEGPT_BCIC2A.py (EEGPTClassifier in downstream/Modules/models/EEGPT_mcae_finetune.py) as an EEGPTEncoder. Note its native rate is 256 Hz — run the probe with --sfreq 256.