Graph-Based Spatio-Temporal Backpropagation for Training Spiking Neural Networks
- Resource Type
- Conference
- Authors
- Yan, Yulong; Chu, Haoming; Chen, Xin; Jin, Yi; Huan, Yuxiang; Zheng, Lirong; Zou, Zhuo
- Source
- 2021 IEEE 3rd International Conference on Artificial Intelligence Circuits and Systems (AICAS) Artificial Intelligence Circuits and Systems (AICAS), 2021 IEEE 3rd International Conference on. :1-4 Jun, 2021
- Subject
- Bioengineering
Components, Circuits, Devices and Systems
Computing and Processing
Power, Energy and Industry Applications
Robotics and Control Systems
Signal Processing and Analysis
Backpropagation
Training
Energy consumption
Firing
Network topology
Neural networks
Encoding
spiking neural network (SNN)
spike sparsity
graph-based spatio-temporal backpropagation (G-STBP)
leaky integrate-and-fire (LIF)
recurrent network
- Language
Dedicated hardware for spiking neural networks (SNN) reduces energy consumption with spike-driven computing. This paper proposes a graph-based spatio-temporal backpropagation (G-STBP) to train SNN, aiming to enhance spike sparsity for energy efficiency, while ensuring the accuracy. A differentiable leaky integrate-and-fire (LIF) model is suggested to establish the backpropagation path. The sparse regularization is proposed to reduce the spike firing rate with a guaranteed accuracy. GSTBP enables training in any network topologies thanks to graph representation. A recurrent network is demonstrated with spike-sparse rank order coding. The experimental result on rank order coded MNIST shows that the recurrent SNN trained by G-STBP achieves the accuracy of 97.3% using 392 spikes per inference.