A Neural Signal Codec with Resource Efficient Encoder for Implantable Brain Machine Interface Systems

ICLR 2026 Conference Submission15966 Authors

Keywords: Codec, neural signal compression, hardware resource efficient, implantable, brain machine interface

TL;DR: We proposed a hardware resource efficient codec with light weight encoder used for neural signal compression in brain machine interface systems, maintaining high fidelity as well holds the cluster information.

Abstract: In this paper, we present a neural signal codec (NSC) with a resource-efficient encoder for implantable brain machine interface (iBMI) systems. The proposed codec has a multiplication-free encoder with only 124-bit lightweight parameters, which is suitable for deployment at the edge of an iBMI system. To reduce the parameter size, a dynamic weight generation mechanism for parameter sharing within the window is implemented in the encoder design. On the decoder side of the codec, a conventional multilayer convolutional neural network with a specially designed loss factor – Energy Aware Loss (EAL) is adopted, which adds adaptive attention to the total loss function to improve reconstruction performance by emphasizing the signal energy intensive regions of the input data section. The parameter storage is reduced by 97% on the encoder side, compared to a conventional FC-based autoencoder with INT8-quantized weights. Large-scale evaluations show that NSC is capable of restoring high-fidelity neural signals and preserving the biological features across diverse neural signal datasets, making it a promising data compression approach for high-throughput iBMI systems. Furthermore, preliminary generalization experiments on other biomedical signals such as ECG (MIT-BIH) further demonstrate the potential of NSC as a general resource-efficient compression framework for streaming biosignals.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 15966