A Biological Population Threshold Coding with Robust Feature Extraction and Neuronal Jitter for SNN-based Speech Recognition

Chen Wang, Chengting Yu, Aili Wang

Abstract: The neuronal dynamics of brain-inspired spiking neural networks (SNNs) make them more suitable for processing dynamic signals. In SNN, neurons interact via discrete spikes. Neuronal coding is crucial to the advancement of neuromorphic computing. In the field of temporal coding, the population threshold coding (PTC) which uses multiple neurons to encode the trajectory of a time-varying signal attracts lots of research attention. It features noise robustness and spike sparsity. In this paper, we (1) evaluate the number of threshold levels and the number of filter banks in the PTC; (2) compare the Mel filter bank, the Gammatone filter bank, and the mix of two-based PTC; and (3) apply different levels of neuronal jitter to the encoding process using speech (TIDIGTS) and sound (RWCP) datasets. The classifications are performed using two types of classifiers: biologically plausible supervised Tempotron learning rule and backpropagation (BP)-based SNN learning rule. Our findings indicate that (1) the appropriate threshold resolution and number of filter banks are dependent on the datasets, and (2) PTC is robust to cochlear filter bank-based feature extractions and neuronal jitter.