P-Spiking Deep Neural Network Based on Adaptive SPCNN Temporal Coding

In this paper, we propose a P-spike deep neural network (P-SDNN) for image classification based on an adaptive simplified pulse coupled neural network (SPCNN) temporal coding. The proposed P-SDNN model introduces a SPCNN tem-coding layer into a spiking deep neural network (SDNN) with parameters adjusted by unsupervised STDP learning rule. The advantage of the proposed SPCNN temporal coding is to obtain adaptive time steps in terms of different input images. Each time step corresponds to a spiking-timing map which may contain a semantic segmentation of the input image. This is guaranteed by the working principle of SPCNN that the higher the neuron intensity is, the larger its internal activity will be and the earlier it will fire. And the adjacent neurons with similar intensity will pulse synchronously in a spiking-timing map. We evaluate the proposed P-SDNN model in the tasks of image classification on the Caltech face/motorbike and MNIST datasets. The experiments show that, under the same experimental conditions, the proposed P-SDNN model performs better than the SDNN model without SPCNN tem-coding.

I. Introduction

Image classification and object detection are the most important visual tasks in computer vision research, which are widely used in many fields, including face recognition [1], pedestrian detection [2], vehicle counting [3], license plate detection and recognition [4], etc.