Hybrid Macro/Micro Level Backpropagation for Training Deep Spiking Neural Networks

Spiking neural networks (SNNs) are positioned to enable spatio-temporal processing and ultra-low power event-driven neuromorphic hardware. , SNNs are yet to reach the same performances of conventional deep neural networks (ANNs), a long-standing challenge due to complex and non-differentiable spike events encountered in training. The SNN error backpropagation (BP) methods are limited in terms of, lack of proper handling of spiking discontinuities, and/or between the rate-coded loss function and computed gradient. We present hybrid macro/micro level backpropagation (HM2-BP) algorithm for training-layer SNNs. The temporal effects are precisely captured by the proposed-train level post-synaptic potential (S-PSP) at the microscopic level. The-coded errors are defined at the macroscopic level, computed and-propagated across both macroscopic and microscopic levels. Different from BP methods, HM2-BP directly computes the gradient of the rate-coded function w. r. t tunable parameters. We evaluate the proposed HM2-BP by training deep fully connected and convolutional SNNs based on the MNIST 14 and dynamic neuromorphic N-MNIST 26. HM2-BP achieves an level of 99. 49% and 98. 88% for MNIST and N-MNIST, respectively, the best reported performances obtained from the existing SNN BP. Furthermore, the HM2-BP produces the highest accuracies based on for the EMNIST 3 dataset, and leads to high recognition accuracy for the16-speaker spoken English letters of TI46 Corpus 16, a challenging-temporal speech recognition benchmark for which no prior success based on was reported. It also achieves competitive performances surpassing those conventional deep learning models when dealing with asynchronous spiking.