2021 Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS) | 2021
A novel spiking neural network with the learning strategy of biomimetic structure
Abstract
Spiking neural network (SNN) is a new artificial neural network computing model inspired by the brain, which is an appropriate tool for processing complex spatiotemporal information with low power consumption. However, due to the discontinuous, non-differentiable mechanism and complicated calculation model of SNN, there are some challenges such as the difficult training process and low learning efficiency of SNN. Inspired by the self-organization process in biological nervous system and based on the spike time difference between presynaptic and postsynaptic neurons, a kind of biomimetic structure learning algorithm is proposed in this paper. We use the leaky integrate-and-fire (LIF) neurons, the conductance-based synapses and based on the spike-timing dependent plasticity (STDP) learning rules construct a four-layer feedforward network in this paper. Different from the fully connected network, based on the precise spike time difference between presynaptic and postsynaptic neurons, we propose a method to establish new connections between neurons as the network is trained. And the winner-take-all (WTA) competition mechanism is introduced in the output layer to ensure the specificity of network learning features. Experiments results show that the classification ability of our network could achieve 92.1% with on the MNIST dataset. And compared with the fully connected SNN, the training time is significantly shortened by 2.6 hours, accounting for 82% of the fully connected network. Finally, combined with relevant experimental data, we also prove that the power consumption of SNN is much smaller than the ANN.