Effect of local excitation-inhibition ratio on word recognition in hierarchical spiking neural network

The auditory neural coding is flexible to acoustic variability and can recognize sounds from competing sound sources. However, it remains unclear how neural circuits encode complex time-varying patterns of sound sources. Here we predict the ascending auditory pathway based on optimized hierarchical spiking neural network (HSNN). Primarily, HSNN model is built and it is found that the local state of HSNN plays an essential role in information processing, which depends on excitation/inhibition (E/I) ratio between the layers. Besides, the effect of E/I ratio on spiking activity propagation and word recognition of recurrent neural network (RNN) is investigated based on HSNN model. It is shown that the performance of sound signal propagation differs dramatically with different local network states. Particularly, an optimal E/I ratio enables HSNN to effectively propagate the spiking activity and filter out information irrelevant to word recognition. Collectively, this work provides a feasible tool to study the neural mechanism of ascending auditory pathway and digital recognition, contributing the field of machine vision and brain-like computing.