DistillSort: Distilling Autoencoders for Efficient Spike Sorting in Intracortical Brain—Computer Interfaces

As intracortical brain—computer interfaces (BCIs) scale to thousands of channels, separating overlapping neural spikes in real time becomes increasingly challenging under strict power and memory budgets. Deep autoencoders have recently emerged as powerful feature extractors for spike sorting, yet their large parameter counts and high inference costs preclude deployment on implantable devices. This work introduces DistillSort, a spike-sorting pipeline that compresses autoencoder-based feature extractors through knowledge distillation. A deep teacher autoencoder is trained on neural waveforms, and a shallow student is distilled to match both its latent codes and reconstructions via a weighted combination of self-reconstruction, code-distillation, and reconstruction-distillation losses. The distilled student produces latent representations that closely approximate those of the teacher while containing only half of the teacher's parameters and requiring$2.3 \times$fewer multiply-accumulate operations. On real multichannel neural recordings, the distilled model maintained the deep autoencoder's clustering quality and was consistently superior to a shallow autoencoder trained without distillation. Inference latency is reduced from$97.0 \mu ~\mathrm{s}$to$60.1 \mu ~\mathrm{s}$. These results demonstrate that knowledge distillation enables efficient and deployable neural feature extraction for spike sorting in BCIs.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.