A Novel Two-Layer Memristive Spiking Neural Network with Spatio-Temporal Backpropagation

Abstract

In this paper, a novel two-layer memristive spiking neural network (MSNN) with spatio-temporal backpropagation (STBP) algorithm is proposed. The MSNN is composed of a memristive crossbar array, ten leaky integrate-and-fire (LIF) neurons and a winner-take-all (WTA) module. The memristive crossbar array with one memristor (1M) synapse performs the multiplication and accumulation without additional storage units. LIF neurons accumulate input current and fire spikes. WTA module ensures that only one neuron fires for one input pattern. The MSNN consumes a little power because there are no amplifiers or digital CMOS elements in the circuit. In order to train the memristive conductance, the LIF model is discretized and the gradients are propagated by the STBP algorithm. Furthermore, we perform a 6 × 5 black and white image classification based on the MSNN in PSPICE. Results verify that the MSNN realizes high recognition rates even under severe random noise and stuck-at faults