Quaternary synapses network for memristor-based spiking convolutional neural networks

Abstract

This paper proposes a method that renders the weights of the neural network with quaternary synapses map into the only four-level memristance of memristive devices. We show this method is capable of operating with a negligible loss in classification accuracy when the memristors utilized can store at least four unique values. Compared with other state-of-the-art methods, the method presented can achieve 98.65% accuracy under the 0.60M parameters. Systematic error analysis shows that the network can still reach over 95% accuracy under the condition of 95% yield of memristor crossbar array, 100 µV op-amp offset voltage and 0.5% Single-Pole-Double-Throw switches noise.