Efficient Reservoir Encoding Method for Near-Sensor Classification with Rate-Coding Based Spiking Convolutional Neural Networks

This paper proposes a general and efficient reservoir encoding method to encode information captured by spike-based and analog-based sensors into spike trains, which helps to realize near-sensor classification with rate-coding based spiking neural networks in real applications. The concept of reservoir is proposed to realize long-term residual information storage while encoding. This method has two configurable parameters, integration time and threshold, and they are determined optimal based on our analysis about encoding requirements. Trough different setting we proposed, reservoir encoding method can be configured as compression mode to compress sparse spike trains obtained from spike-based sensors, or conversion mode to convert pixel values captured by analog-based sensor into spike trains respectively. Verified on MNIST and SVHN dataset, the mapping relationship of information before and after encoding are linear, and the experimental results prove that rate-coding based spiking neural networks with our reservoir encoding method can realize high-accuracy and low-latency classification in two modes.