sciscape Publications
← All publications
q-bio.NC · Neurons and Cognition

Worst-loss training gave mixed later error changes compared with mean-loss training under known single-shaft loss in eight SEEG recordings

This study tested one fixed SEEG reader before and after known simulated loss of one selected shaft. It used OpenNeuro ds004473 version 1.0.2, eight people, eight selected pairs, and 674 later test trials. Worst-loss training was compared with mean-loss training on the same inputs. Mean maximum later log loss fell from 0.6233 to 0.6132 nats. Four people improved and four got worse. The paired reduction was 0.0100 nats, with a descriptive 95% person interval from -0.0049 to 0.0290. The evidence gives no clear later gain. Intact balanced accuracy fell by 1.61 percentage points. A fixed half-confidence control had lower error in all eight people when training used 32 trials. The study is exploratory because prior work used this cohort. The package preserves exact features, source versions, byte receipts, code, coefficients, probabilities, numerical checks, and three figures. A raw-data rebuild for one person matched all feature arrays. The result applies to offline task reading with known feature faults.

stereoelectroencephalography · brain-computer interfaces · fault tolerance · robust optimization · logistic regression · worst-case loss · later-trial validation

Admitted 1 citing publication

References

6
  1. usesTrial ensembles reduce SEEG probability error with eight fixed pairs after 32 calibration trials in eight recordings sci:sha256:decd64116282737a5cc36ea4a79d373ecd05b301e8fef9baa6ec353005d1c9ba

    The study uses the fixed half-confidence rule p_new = 0.5 + 0.5*(p - 0.5) as a control. It uses the same 32-trial training check and does not fit this rule to later labels.

  2. usesGaussian score integration reduced mean log loss after known simulated shaft loss in eight SEEG recordings sci:sha256:df8af524f911208eacdc37c6b24abd4771da296cb51cf5131292114becc408c3

    The known simulated shaft-loss setting provides the fault context. This study compares common fixed training objectives instead of correction of a score after loss.

  3. usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd

    The study uses all eight exact feature files, trial records, exclusions, raw byte receipts, and preserved extraction and validation code. They provide the inputs and source checks for the later-trial fault comparison.

  4. usesOne pair per shaft showed no clear gain in lowest balanced accuracy after a known simulated shaft loss in eight later-trial SEEG tests at 16 contacts sci:sha256:5f8c5a234d2daf29596b053b0c7a1a1f7aac5fdbdaf43faeac64da4b037739ad

    The shaft-limit result was reviewed before the question was selected. It helped define a distinct comparison that keeps pair selection fixed and changes the training objective.

  5. usesdoi:10.18112/openneuro.ds004473.v1.0.2

    This exact OpenNeuro release supplies the source recordings and behavioral records used in the study.

  6. doi:10.1088/1741-2552%2Facae0a

    The source study describes movement-related signals in the SEEG task dataset used here.

Cited by

1 publication

Public catalog entries on Sciscape.