BCIwiki (bciwiki.com) — A 4.9K-parameter decoder picked without ever touching the held-out user's data reached 44.35±15.41% accuracy on four-class motor-imagery EEG (MI-EEG), researchers reported in a preprint posted to arXiv on September 27, 2026. The paper, by Abdul Basit, Saim Rehman and Muhammad Shafique, presents ThinkNet, a validation-controlled framework for selecting compact decoders and inference policies. It has not been peer reviewed; the authors note it has been accepted to IEEE-EMBS BHI'2026.
Assistive and rehabilitative brain-computer interfaces (BCIs) need subject-independent decoders that generalize to new users with limited data and constrained compute, the authors wrote. Held-out-subject results can be overstated when test-subject information influences preprocessing, model selection or ensemble selection. ThinkNet combines train-only normalization, validation-guided evolutionary search and validation-gated inference.
The team evaluated four-class decoding on BCI Competition IV-2a (session T) with 9 leave-one-subject-out folds, 3 seeds, 7 fixed decoder entries and a broader search over 10 decoder families. The held-out subject was never used for normalization, hyperparameter, architecture or ensemble-policy selection. The selected compact decoder has 19 KB of FP32 weights and ran batch-1 inference in 0.99 ms on Orin CUDA.
Across the broader search, compact models of up to 25K parameters averaged 40.10% held-out accuracy after selected retraining, versus 35.09% for mid-size and 34.78% for large models. Validation-gated ensembling beat validation-selected single-model inference, reaching 43.98±16.25% in the fixed benchmark and 43.31±15.88% for the compact six-family ensemble.
A non-deployable oracle analysis showed a 6.1-point family-selection gap and near-zero validation-test correlation, which the authors said makes validation reliability a key bottleneck under subject shift.