On-Chip Spike Detection and Classification using Neural Networks and Approximate Computing.

Neural ensembles control sensory, motor, and cognitive functions. Action potentials of neuronal cells (spikes) may signify such functions, or the presence of a pathology. In this paper we give the circuital implementation of an Artificial Neural Network, able to sort (detect and classify) spikes in real time. The system is synthesized targeting a 14nm FinFET technology. To partially alleviate the computational burden, approximate computing methods have been integrated during the inference stage, yielding up to 63% reduction in dynamic power. The different versions of the circuit reach an accuracy range from 65% to 93%, with silicon area and power that range from 2000μm ² , 0.1μW@30kHz to 6000μm ² , 0.7μW@30kHz. The electrical performances of the proposed circuit overcome the state of the art of spike detection circuits while providing the additional feature of spike sorting in a single integrated solution.