Memristive Synaptic Circuits for Deep Convolutional Neural Networks

Menglin Cui, Yang Zhang

Abstract: The memristor is a novel promising component for applications in non-volatile memory, logic circuits, and neuromorphic computing systems. In this paper, deep learning systems using memristive synaptic circuits are presented. One key issue in memristive modeling is the device variation issue. It is very difficult to distinguish the intermediate memristive states in a multilevel memristor. A fuzzy modeling method, which defines the memristive states in a dynamic way, is proposed to improve the robustness of the memristive neural networks. In addition, memristive deep learning systems are presented for pattern recognition such as image recognition. The effectiveness of the proposed methods has been verified through comprehensive tests. The proposed memristive deep convolutional neural networks greatly reduce the number of memristors and increase the recognition rate as compared with recent examples in the literature.