What changed
The site now includes an electrode ablation report for the featured 2-M16 recipe. The study retrains the model after changing which Muse 2 channels are available, then compares held-out action accuracy and non-REST finger accuracy.
This is a subject-specific retraining ablation. It is not yet deployment-gated replay, and it should not be read as anatomical localization.
The report is now included near the bottom of the deep dive.
Study setup
| Item | Value |
|---|---|
| Branch | ablation/electrode-channel-importance |
| Train/test config | Same featured 2-M16 derived dataset, split policy, model recipe, and deterministic test evaluation |
| Primary sweep | Seed 43, 60 epochs |
| Subsets | Full montage, single-channel, leave-one-out |
| Additional run | TP9+TP10 only, dropping both AF7 and AF8 |
The fresh full-montage retrain reached 90.96% action accuracy and 86.26% non-REST finger accuracy. It is close to the locked paper checkpoint, but it is a new training run and should be compared only against the ablation runs in this sweep.
Headline result
| Channels | Action accuracy | Non-REST finger accuracy | Action drop vs full | Finger drop vs full |
|---|---|---|---|---|
| TP9+AF7+AF8+TP10 | 90.96% | 86.26% | 0.00 pp | 0.00 pp |
| TP9+TP10 only | 88.31% | 83.90% | 2.65 pp | 2.36 pp |
Removing both frontal electrodes preserves most of the full-montage retrain. That supports the working hypothesis that, for this subject and split, the most useful signal is concentrated in the temporal electrodes TP9 and TP10.
Leave-one-out importance
| Omitted electrode | Kept channels | Action drop vs full | Finger drop vs full |
|---|---|---|---|
| TP10 | TP9+AF7+AF8 | 12.08 pp | 9.13 pp |
| TP9 | AF7+AF8+TP10 | 11.73 pp | 13.29 pp |
| AF7 | TP9+AF8+TP10 | 1.13 pp | -0.40 pp |
| AF8 | TP9+AF7+TP10 | 0.56 pp | -1.30 pp |
TP9 and TP10 are the important leave-one-out channels in this sweep. Dropping AF7 or AF8 causes little action degradation and slightly improves aggregate finger accuracy.
Single-channel sufficiency
| Electrode | Action accuracy | Non-REST finger accuracy |
|---|---|---|
| TP9 | 73.01% | 72.22% |
| TP10 | 72.66% | 67.80% |
| AF7 | 48.89% | 35.46% |
| AF8 | 48.33% | 36.16% |
TP9 and TP10 each retain substantial information alone. AF7 and AF8 perform much closer to weak single-channel models.
Frontal-channel action behavior
Dropping a frontal electrode did not reduce REST accuracy.
| Run | REST delta | OPEN delta | CLOSE delta |
|---|---|---|---|
| Drop AF7 | +0.65 pp | -2.39 pp | -0.56 pp |
| Drop AF8 | +2.28 pp | -3.37 pp | +1.03 pp |
The small action decrease from frontal-channel removal is mainly an OPEN decrease. It is not a REST failure and not a uniform action decrease.
Per-finger behavior
| Run | Overall delta | THUMB | INDEX | MIDDLE | RING | PINKY |
|---|---|---|---|---|---|---|
| Drop AF7 | +0.40 pp | +0.00 pp | -7.95 pp | -3.16 pp | +12.60 pp | +0.51 pp |
| Drop AF8 | +1.30 pp | +0.00 pp | -2.45 pp | -1.81 pp | +18.90 pp | -6.87 pp |
| TP9+TP10 only | -2.36 pp | -0.43 pp | -15.29 pp | -10.38 pp | +12.33 pp | +1.53 pp |
The frontal channels may still help specific digit boundaries, especially INDEX and MIDDLE. They may also add noise or conflicting information for RING in this subject-specific split.
Current interpretation
The ablation supports three practical conclusions:
- TP9 and TP10 are the strongest channel candidates for this subject-specific model.
- AF7 and AF8 are not required to preserve most held-out accuracy in this single-seed retraining sweep.
- Reduced-channel command-path claims still require repeated seeds and deployment-consistent replay.
The next validation step is to repeat the channel sweep across seeds, then replay the strongest reduced montages through the same conservative command gates used before robotic-hand actuation.