2026-02-22

Second subject results: 1-M16

Published the second subject run and a side-by-side comparison with 2-m16.

Historical note: archived update posts preserve the figures published at that time. For the current verified run bundles, use the results page.

What a run means

A run is a full training + evaluation pass for one subject using fixed-length EEG windows. The model predicts both an action class and a finger class for each window, then we measure accuracy on a held-out test split. The finger accuracy highlighted here is reported on non-REST windows, meaning it excludes windows labeled as rest. This makes the finger metric more relevant to active movement identification. These results are window-level snapshots for a single run and are best interpreted alongside additional runs.

Comparison: 1-M16 vs 2-M16

Metric1-M162-M16
Test action accuracy77.38%89.31%
Test finger accuracy on non-REST windows86.27%89.90%
Test windows2,6482,040
Test non-REST windows2,5341,871
Train vs test action gap (pp)+5.98 pp+1.68 pp
Train avg loss0.74210.5174

What changed / what we learned

  • 1-M16 has lower test action accuracy than 2-M16.
  • 1-M16 has lower finger accuracy on non-REST windows than 2-M16.
  • 1-M16 evaluated more test windows and more non-REST test windows.
  • The train-test action gap is larger for 1-M16.
  • Train average loss is higher for 1-M16.

Since this update was first published, live robot-hand actuation testing has been completed for subject 2-M16. We are continuing to collect additional runs to better quantify variation across subjects and sessions.

Links