Spiking neural networks (SNNs) enable power-efficient implementations due totheir sparse, spike-based coding scheme. This paper develops a bio-inspired SNNthat uses unsupervised learning to extract discriminative features from speechsignals, which can subsequently be used in a classifier. The architectureconsists of a spiking convolutional/pooling layer followed by a fully connectedspiking layer for feature discovery. The convolutional layer of leaky,integrate-and-fire (LIF) neurons represents primary acoustic features. Thefully connected layer is equipped with a probabilistic spike-timing-dependentplasticity learning rule. This layer represents the discriminative featuresthrough probabilistic, LIF neurons. To assess the discriminative power of thelearned features, they are used in a hidden Markov model (HMM) for spoken digitrecognition. The experimental results show performance above 96% that comparesfavorably with popular statistical feature extraction methods. Our resultsprovide a novel demonstration of unsupervised feature acquisition in an SNN.
展开▼