Two phase-change memory (2-PCM) neurons for implementing a backpropagation algorithm

Abstract

In this paper, we proposed a neuron circuit, i.e. a two phase-change memory (2-PCM) neuron, to implement a backpropagation algorithm for hardware neural networks. PCM devices in the neuron are used for computing and storing signals. One of the PCM devices is used for storing the forward propagation (FP) signal, and the other is used for storing the backpropagation signal. This paper presents a new application of PCM devices in traditional artificial neural networks. With the proposed 2-PCM neuron circuits, the neuron circuits need not to remain powered during the entire process for temporarily storing the FP signals. And it does not require additional memory for storing backpropagation signals. In addition, the two PCM devices share a common read and write driver circuits.