Year: 2023
Type: conference-paper
Abstract: Spiking neural networks (SNNs) obtain impressive good performance on various applications due to their powerful computing capacity for encoding spatio-temporal information. However, most existing spiking neural networks remain shallow structures, lacking effective structures to handle real-world tasks. In this work, we propose an autoencoder induced deep spiking neural network (AE-DSN) to improve the representative capacity. Specifically, AE-DSN consists of three coding modules shared the same structure. Each module contains an autoencoder and a spiking coding layer. In particular, we present a progressive training strategy to train these modules one-by-one. For one module, the spiking coding layer is trained using ReSuMe algorithm, guided by the encoded information from the autoencoder, which would be dropped when training the subsequent module. The entire training process terminates when the final spiking coding layer is trained well. Experimental results show that the proposed AE-DSN could effectively extract discriminative features for the input images to achieve superior classification performance.
Authors: Ruoshi Li, Xin Xu, Lihong Wan, Qingdu Li, Ye Yuan, Na Liu
Institution: University of Shanghai for Science and Technology
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4385282944
