Author(s): Wang, TX (Wang, Tengxiao); Shi, C (Shi, Cong); Zhou, XC (Zhou, Xichuan); Lin, YC (Lin, Yingcheng); He, JX (He, Junxian); Gan, P (Gan, Ping); Li, P (Li, Ping); Wang, Y (Wang, Ying); Liu, LY (Liu, Liyuan); Wu, NJ (Wu, Nanjian); Luo, G (Luo, Gang)

Source: NEUROCOMPUTING Volume: 425 Pages: 96-106 DOI: 10.1016/j.neucom.2020.10.100 Published: FEB 15 2021

Abstract: Brain-inspired spiking neural networks (SNNs) have become a research hotspot in recent years. These SNNs communicate and process information in a form of spatiotemporally sparse spikes, leading to high energy efficiency and low computational cost for object classification tasks. However, to reduce computational complexity while maintaining SNN classification accuracy still remains a challenge. Extracting representative and robust feature is the key. This paper proposes efficient spatiotemporally compressive spike features and presents a lightweight SNN framework that includes a feature extraction layer to extract such compressive features. Our experiments based on popular benchmark datasets demonstrated that the spatiotemporally compressive spike features are competent and robust in representing the input spike trains. The experimental results also suggest that our lightweight SNN framework with such compressive spike feature requires a small amount of processing time consumption while achieving comparable classification rate across many popular datasets: MNIST, MNIST-DVS, Poker-DVS, Posture-DVS and more challenging Fashion-MNIST datasets. The SNN framework has a potential to be applied in low-cost or resource-limited edge computing systems and embedded devices. ? 2020 Elsevier B.V. All rights reserved.

Accession Number: WOS:000632015900008

ISSN: 0925-2312

eISSN: 1872-8286

Full Text: https://www.sciencedirect.com/science/article/pii/S0925231220317136?via%3Dihub