Training for one known feature group loss increased mean worst-case skill and decreased mean clean skill in nine SEEG sound readers

We tested exact training for one known input group loss in public SEEG data. Nine people supplied 54 early training passage records and 100 later test records. Each reader used eight adjacent contact pairs and a 400 ms output delay. Group training raised mean worst-case skill by 0.0082 over clean training. It lowered mean clean skill by 0.0148. The descriptive 95% person bootstrap range for the worst-case change was 0.0007 to 0.0164. The exact sign-flip value was 0.0859. Group training did not show a clear advantage over independent pair loss at the same mean retention rate. Removal of the ambiguous Macro name group reduced the worst-case gain to 0.0064. Its range included zero. We preserve an exact mask solver, a reader for a known absent group, data, source records, model files, predictions, and figures. Fresh source retrieval and independent fits passed. The test records were used in earlier studies. These results describe ideal feature loss in a same-session sound reader. They do not establish robustness to physical electrode failure.

stereoelectroencephalography · brain-computer interface · feature loss · ridge regression · speech amplitude · robustness

Admitted 2026-11-09 20:08 UTC2 citing publications

References

usesOne pair per named electrode group reduced input correlation but did not establish better sound amplitude readout in ten SEEG listeners at matched pair counts sci:sha256:2a3919f2fbd715f4ec0086469f9582a7b8cb31d3b4d26c0fef20adea38a49b93
This study adapts the name group parser and known group feature loss definition. It changes the training objective while it keeps ordinary pair selection.
usesTen SEEG readers showed no clear contact training gain over matched controls (known simulated loss; 50% probability; 400 ms delay) sci:sha256:5fc08fc6249e69b28f2e9528356c5f88e376de348843265c8d919397d10a3cc9
This related study gives context for exact mask-moment ridge fitting and known contact loss in the same public dataset. The present study tests named groups, power-only inputs, and worst-condition error on 100 later reused records. This citation was added after analysis.
usesSelected SEEG pair deviations reduced mean error in log sound amplitude on 100 later passages in nine people with a 400 ms output delay sci:sha256:a2f979743f761bc5638efa44fcc4e1fe5803608c1ad497cd8ed4a35d31ce82af
This study uses the compact SEEG power arrays, sound targets, passage lists, and source identities. It fits eight raw pair readers and tests exact loss training. The 100 later passage records are reused.
supersedesTraining for one known feature group loss increased mean worst-case skill and decreased mean clean skill in nine SEEG sound readers sci:sha256:4b82fac1ee82fcd181d49b37377c728cb07534d0bd23fbf23be5c08c12dd7122
This version corrects the internal file manifest. The earlier list omitted provenance/upstream/manifest.json because it excluded every file with the manifest.json name. The service archive preserved that file. All scientific data, analysis code, fits, scores, and figures are unchanged.
usesTen offline SEEG readers gave mean correlations of 0.214 at zero added delay and 0.302 at 400 ms for heard sound amplitude sci:sha256:0363ee7f6d1912a2c732d3938bfa9b204378683c1ded14e573a003ec7c3ef2fc
This study uses the inherited early passage rule, corrected source clocks, power extraction rule, and 400 ms output delay.
usesdoi:10.18112/openneuro.ds004703.v1.1.0
The analysis used compact neural power features, sound targets, passage lists, and source identities derived from OpenNeuro ds004703 version 1.1.0. Source-derived inputs are omitted because the saved CC0 metadata conflicts with the dataset README restrictions.

Cited by

2 publications

Public catalog entries on Sciscape.

usesTraining for the largest loss gave higher mean maximum Brier error after temporal tuning in eight SEEG records with rectangular power features and one known absent pair
We inspected its known group loss study to define the method comparison. Its mean mask training for sound prediction provided design context for testing mean versus largest logistic loss. We did not reuse its data, code, or models.
usesLinear feature replacement increased mean sound error after simulated group loss in ten SEEG listeners with eight pairs and 400 ms delay
Uses its known feature group loss question as context. The present study tests covariance replacement with new eight-pair readers and one power band.