An Adaptive Convolution Auto-encoder Based on Spiking Neurons

Chuanmeng Zhu, Jiaqiang Jiang, Runhao Jiang, Rui Yan

Abstract: Neural coding is one of the central questions in neuroscience for converting visual information into spike patterns. However, the existing encoding techniques require a preset time window and lack effective learning. In order to overcome these two problems, we design an adaptive convolutional auto-encoder based on spiking neurons in this paper. We first exploit the spike pixel mapping decoding approach to find the optimal value of the time window automatically. Next, we design a deep convolutional neural network to adapt the learning parameters by reconstruction errors to realize the spike encoding process. Then we can naturally get coding pre-training parameters for unifying the convolutional spike coding layer with back-end deep spiking neural networks (SNNs) for recognition tasks. Simulation results demonstrate that the proposed method can achieve better performance compared with other encoding methods.