Blaise Yvert, BrainTech Laboratory, Inserm et Université Grenoble Alpes
Neuromorphic spike sorting: Toward very low-power neural processing in cortical implants
Embedding real-time processing of large-scale neural signals in cortical implants is challenging due to the high power consumption and heat dissipation associated to classical approaches. Here we will discuss a new way to address spike sorting, a key step in the analysis of neural signals, using a neuromorphic approach. An artificial spiking neural network has been developed embedding local STDP and STP learning rules and an attention mechanism. It has been designed to process the online flow of neural signals and to output artificial spikes corresponding to the real spikes present in the input signal, with one active output artificial neuron for each active real neuron. This approach has been tested on single electrode data and found to give comparable performance than classical spike sorting methods with even more robust results at low signal-to-noise levels.