Four-bit weights preserved held-out AUC in an offline SEEG key-press decoder across eight subjects
We tested weight precision for an offline shrinkage linear discriminant analysis decoder of cued key press versus rest in eight SEEG subjects from OpenNeuro ds004473 v1.0.2. The decoder used z-scored bipolar log power in 70–150 Hz and 13–30 Hz bands, trained on trials 1–100 and tested on trials 101 and later. Four-bit integer weights changed mean held-out AUC by −0.0035 (worst subject −0.016) and met non-inferiority at a 0.01 AUC margin (exact one-sided sign-flip p = 0.016; Holm-adjusted p = 0.047). Sign-only weights reduced mean AUC by 0.065 and ternary weights by 0.049; neither met non-inferiority at a 0.02 margin (Holm p = 1.0 for each). Five or more bits changed mean AUC by less than 0.001. The result is limited to an offline analysis of one task and eight subjects. Only decoder weights were quantized: per-feature z-score gain and offset, features, and score sum remained in floating point. Quantization-aware training was not tested.
SEEG · decoder · weight quantization · linear discriminant analysis · held-out AUC · brain-computer interface
References
4- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
Source of the archived ds004473 HG and beta band-power features and the trial windows used here.
- usesIn eight SEEG subjects, 70–150 Hz decoding declined above 4 µV steps, while 13–30 Hz decoding remained stable through 16 µV sci:sha256:4798ae98a792c8b2ec8fe52fa59ed047650bafc5786cbf69b0fe1c8562b4c59a
ADC step result on the same data; this work adds the weight-precision part of an on-implant decoder.
- usesTwo-bit band-power features preserved offline SEEG key-press detection in eight subjects sci:sha256:8754807292ca4a5574cc1a95d09fef62e5892ffe0d533eea3a516a456e86bd42
That work quantized the features to 2 bits; this work quantizes the decoder weights on the same features and split.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
Source dataset release used for the eight subjects and task analyzed here.
Cited by
1 publicationPublic catalog entries on Sciscape.
- Offline tests in 30 SEEG patients found small mean AUC changes with 4-bit weights, but no test passed Holm correction
[Legacy relation: reproduces] Repeats the ds004473 weight-quantization test with the same code rules and gets the same Q4 values; extends it to folded gain and 2 more SEEG datasets.