Answers are grounded in this paper and cite their sources.
Objective Training deep learning neural motor decoders requires large datasets, which are often difficult to obtain in human experiments. With the advent of complex stimulation feedback and compact wireless recorders, additional data is needed to account for stimulation artifacts and other interference. Here, we propose a platform for synthesizing neural data under controlled adverse recording conditions to augment existing datasets and support the evaluation and development of neural decoding and artifact-mitigation algorithms. Methods We developed custom hardware to convert recorded neural signals into analog waveforms, replay them through a controlled noise environment, and then resample them with realistic physical distortions. We demonstrate this approach by corrupting a clean motor control dataset with targeted Wi-Fi interference and training a 1D U-Net to remove the resulting artifacts. Results Baseline resampling deviated original decoder accuracy by <1%, confirming minimal signal degradation. Under wireless interference, the U-Net denoiser improved SNR by 14.2–22.3 dB. This allowed the motor decoder to overcome severe classification failures, restoring accuracy to within 5%–8% of the clean baseline and achieving up to 98.1% agreement with original predictions. Significance This work provides a practical framework for neural data augmentation by enabling previously recorded biosignals to be replayed and re-recorded under controlled adverse conditions. The platform can be used to quantify the effects of recording artifacts and to support the development and evaluation of artifact-mitigation methods for more robust neural decoding in real-world applications.
No takes yet. Share an insight, caveat, or question.
Drealan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: