Quantisation and pooling method for low-inference-latency spiking neural networks

Spiking neural network (SNN) that converted from conventional deep neural network (DNN) has shown great potential as a solution for fast and efficient recognition. A layer-wise quantisation method based on retraining is proposed to quantise the activation of DNN, which reduces the number of time steps required by converted SNN to achieve minimal accuracy loss. Pooling function is incorporated into convolutional layers to reduce at most 20% of spiking neurons. The converted SNNs achieved 99.15% accuracy on MNIST and 82.9% on CIFAR10 by only seven time steps, and only 10-40% of spikes need to be processed compared with networks using traditional algorithms. The experimental results show that the proposed methods are able to build hardware-friendly SNNs with ultra-low-inference latency.