A compact pipeline for motor-imagery EEG: align each recording's covariance, run it through a small EEGNet-style network, then adapt the model to the target subject at test time. It runs on two BCI Competition IV datasets — IV-2a (4-class, 22 channels) and IV-2b (2-class, 3 channels), 9 subjects each — using the usual session-holdout and leave-one-subject-out (LOSO) protocols.

Within-subject uses the official session split (IV-2a: train session 1, test session 2; IV-2b: train sessions 1–3, test sessions 4–5). LOSO trains on eight subjects and tests on the held-out ninth. A validation split carved from the training data picks the checkpoint; the test set is scored once at the end.

Alignment helps in every setting. The test-time adaptation step is where the cross-subject gains come from — roughly +3 points on both datasets' LOSO. Within-subject it helps IV-2a and comes out about even with plain alignment on IV-2b. For context, recent within-subject IV-2a results: CTNet 82.5, MSCARNet 82.7, EEGEncoder 86.5. Full tables and ablations are in RESULTS.md.

1. Alignment — whiten each recording by its mean spatial covariance, either Euclidean (ea) or Riemannian/centroid (ra). It's computed from the signals alone, so it applies the same way to training and test data.
2. Backbone — UnifiedEEGNet, a small EEGNet-style CNN (~34k parameters).
3. IM-TTA — at inference, the model adapts to the target subject's unlabelled trials by information maximization: it sharpens each prediction while keeping the overall class mix balanced, which stops it from collapsing onto a single class. This is the main driver of the cross-subject improvement.

ARCHITECTURE.md walks through the whole pipeline in detail.

uv sync
# both protocols at once, the RA + IM-TTA setup
python main.py --dataset iv2a --protocol both --align ra --tta_steps 5 --tta_div 1.0 --seed 42
python main.py --dataset iv2b --protocol both --align ra --tta_steps 5 --tta_div 1.0 --seed 42

Flags: --dataset {iv2a,iv2b} · --protocol {within,loso,both} · --align {none,ea,ra} · --tta_steps N · --tta_div {1 = info-max, 0 = entropy-only} · --seed.

• model.py — UnifiedEEGNet and the alignment-model variants
• dataset.py — MOABB loader, dataset registry (iv2a/iv2b), session-holdout splits, EA/RA alignment
• train.py — training loop, validation-based checkpointing, tent_adapt (the test-time adaptation)
• main.py — entry point: datasets, protocols, alignment, TTA, multi-seed runs
• ARCHITECTURE.md / RESULTS.md — method writeup and full results
• results/ — per-run CSVs and training logs (see results/README.md)