Multichannel Many-Class Real-Time Neural Spike Sorting With Convolutional Neural Networks

Abstract

Real-time in-sensor spike sorting is a forefront requirement in the development of brainmachine interfaces (BMIs). This work presents the characterization, design, and efficient implementation on a field-programmable gate array (FPGA) of a novel approach to neural spike sorting intended for implantable devices based on convolutional neural networks (CNNs). While the temporal features, the shape of the spike signals, could be highly mitigated from the ambient noise, the proposed classifier effectively extracts spatial features from the multi-channel neural signal to maintain high accuracy on the noisy data. The proposed classifier mechanism was tested on real data that is recorded from multi-channel electrodes, containing 27 neural units, and the classifier achieves 93.1% accuracy despite high temporal noise in the signal. For hardware synthesis, the CNN weights are quantized to reduce the model storage requirement by 93% compared to its floating point-precision version, and the model achieves an accuracy of 86.1%.