Spoken Digits Classification Based on Spiking Neural Networks with Memristor-Based STDP

Abstract

Spiking neural networks are commonly attributed to the third generation of neural networks. They mimic biological neurons more closely by processing information in the form of impulses (spikes) and are characterized by low power consumption and ease of hardware implementation. This paper shows two approaches to the task of classifying audio data represented by the spoken digits dataset using spiking neural networks with memristive plasticity. It is shown that both supervised and unsupervised learning methods based on local plasticity can be successfully used for audio classification. The models achieve accuracies ranging from 80% to 94% depending on the network topology, plasticity type and the way of decoding output neuronal activity. The results obtained in the paper can be a step towards creating neuromorhic devices for recognizing audio signals.