Building Efficient Event-driven Networks from Frame-driven Networks with FIT

Spiking Neural Nnetworks (SNNs), the third generation Artificial Neural Networks (ANNs), are attractive because of their event-driven and sparsely spiking. An efficient train method is converting pre-trained ANNs to SNNs. There have been some approaches to decrease accuracy losses from ANNs to SNNs. However, their theoretical analyses are not enough and their effects still have room for improvement. In this paper, we analyze reasons of accuracy losses from ANNs to SNNs systematically. Then we propose an optimization method that can convert ANNs to SNNs with almost no accuracy loss, which is called finite-input-time (FIT). The MNIST database is employed to verify our analyses and optimization method. Simulation results are consistent with our analyses and show that our optimization method can convert ANNs to SNNs without accuracy loss.

Spiking neural networks (SNNs), artificial neural networks (ANNs), conversion, finite-input-time (FIT), accuracy loss

Yang Zhao, Haibo Wang, Yang Yang, Building Efficient Event-driven Networks from Frame-driven Networks with FIT. 2019 Scientific Conference on Network, Power Systems and Computing (NPSC 2019), 2019: 175-179. DOI: https://doi.org/10.33969/EECS.V3.040.