Advancing Neuromorphic Computing Through Temporally Coded Spiking Neural Network Learning

Abstract

Standard deep learning architectures require a lot of power during inference, as all neurons operate in a dense and synchronous manner. Neuromorphic computing is a new paradigm built using Spiking Neural Networks (SNNs), where the input is represented by the exact time of sparse neuronal spikes instead of the value of its activation function. The learning of temporally coded SNNs—especially the Time-To-First-Spike (TTFS) and Weight-Temporally Coded Representation Learning (W-TCRL) schemes—is a promising area in the field of energy-efficient machine intelligence. This work presents a new learning algorithm called TempoNeuro to push forward the field of neuromorphic computing using biologically plausible temporal coding, surrogate gradient-backpropagation through time, and hardware-aware spike sparsity regularization. TempoNeuro achieves an accuracy of 93.4% on the CIFAR-10, DVS-Gesture, and N-MNIST benchmarks while only consuming 19.7% of the power of an equivalent ANN baseline. This gives a 5.1 power reduction with a loss of less than 1% accuracy and reduces the synaptic operations by 3.2 against the rate coded SNN and sets a new Pareto frontier between accuracy and energy for neuromorphic vision tasks.