Investigation of Analog Memristor Characteristics for Hardware Synaptic Weight in Multilayer Neural Network

Abstract

Analog conductance switching characteristics of memristor devices have been studied to be utilized for constituent elements of synaptic weight matrix in neural networks, related to system design of hardware-level parallel neuromorphic computing architecture for the artificial intelligence application. In this manner, it is important to systematically investigate the specific requirements of memristor characteristics associated with the capability to emulate plenty of synaptic weight elements linked between constituent layers in neural networks. Here, the learning capabilities of analog conductance state of memristor device for the perceptron of unstructured complex dataset in multilayer neural network are analyzed in terms of the number of analog state, nonlinearity, and conductance error. It is found that the requirable number of analog state is analyzed in about ≈50 states and conductance deviation of each analog state is until ≈5% of original value with nonlinearity of ≈0.142 according to constant programming pulse scheme. With the memristor characteristics enough to mimic synaptic weight to be learnt and infer the Fashion-mnist dataset, the classification accuracy is satisfied as ≈84.36% with the loss of ≈16.8% to original level. Owing to this investigation, applicability of novel memristor device could be conveniently examined for the utilization as synaptic weight in multilayer neural networks.