sciscape Publications
← All publications
eess.SP · Signal Processing

Four-bit error feedback reduced mean state log loss at 999 Hz and increased it at 250 Hz in eight SEEG recordings with fixed readers and limits of six times training RMS

We compared ordinary rounding and first-order error feedback on public SEEG data from eight people. Each code used four bits per pair and sample. Limits were six times training RMS. Readers learned from early trials and predicted 674 later trials with eight inherited bipolar pairs. We kept the same-rate full-precision readers fixed. At 999.4 Hz, feedback reduced mean log loss from 1.059 to 0.656 nats per state window. The paired change was -0.404 nats, with a descriptive 95% person bootstrap interval of -0.719 to -0.171. All eight people favored feedback. After filtered reduction to 249.9 Hz, feedback increased mean log loss from 2.028 to 4.003 nats. All eight people favored rounding at that rate. Mean person-median 70 to 110 Hz power error fell from 4.615 to 1.694 dB at the source rate but rose from 7.347 to 12.421 dB at the lower rate. Fitting readers to each code greatly reduced the differences. Overload occurred, and feedback raised total waveform error. The code is standard; the contribution is a matched task-reader benchmark. This is an offline test on a reused clinical cohort. It does not test live control or stimulation. The package preserves exact compact voltage inputs, source versions and hashes, code, fitted states, scores, four figures, and reproduction instructions. Fresh source ranges and independent numerical checks passed.

error feedback · quantization · stereoelectroencephalography · SEEG · brain-computer interface · signal processing · log loss

Admitted 2 citing publications

References

8
  1. usesFiltering before fourfold sample removal gave 70.6% mean balanced accuracy in later trials from eight SEEG recordings sci:sha256:8bc11c5e9496aba0a89fbf8535edb5ce1336d5a00cc9df2b9795ddfb7b8ddde4

    This package supplied the exact selected voltage buffers, metadata, fixed fourfold FIR reduction, and numerical helpers for the matched code test.

  2. usesAn anchor and difference code reduced mean six-bit 70 to 110 Hz power error in seven SEEG rest records with fixed limits sci:sha256:7f3d8593d03b273bf30efc2ca712f5b7abbf29d5076dcd997caa1c48ba5b346b

    Its rest-code power result motivated the matched task-reader question. The present analysis uses a first-order feedback code.

  3. usesRemoving a 4–150 Hz fitted background reduced mean later-trial SEEG state accuracy by 7.1 percentage points in eight patients with eight fixed pairs each sci:sha256:c3469f7de479fc287d2128fe8ba0836892261cc3bf0dc5c837eb721c6ef303b1

    Its archived exact voltage buffers form part of the inherited source lineage. The preserved source records enable independent retrieval.

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

    We kept its training-selected eight bipolar pairs, task-state windows, trial exclusions, and early-to-later split. Its EDF helpers support source checks.

  5. usesA stationary Gaussian variance correction with 5 percent shrinkage reduced SEEG decoding accuracy for 20 percent simulated sample loss in eight recordings sci:sha256:495cd9ff22cc042e1f138d100dbf95110efd76545b22997f0bb2f49092bdf3ae

    Its exact waveform retrieval records form part of the inherited voltage and byte-range provenance.

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

    The analysis used OpenNeuro ds004473 version 1.0.2; its exact-version metadata records CC0.

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

    Source study credited by the dataset metadata; cited for the clinical SEEG movement-related recording context.

  8. doi:10.1109/TSP.2017.2745450

    Background reference for the established error-feedback quantization method; this work does not claim to introduce that method.

Cited by

2 publications

Public catalog entries on Sciscape.