Modified Dropout and Maxout based on the MNN for improving accuracy

Abstract

Memristor crossbar array is an emerged architecture suitable for matrix computation. Memristor based neural networks (MNN) address the speed and energy efficiency issues in computing hardware. However, there are still a lot of problems with memristor, and the limited size of memristor crossbar resulting in the accuracy of the MNN is lower than conventional neural networks (CNNs). In this paper, a modified Dropout and Maxout based MNN for improving the accuracy of the MNN is proposed. A three-layer memristor based multilayer Perceptron (MLP) in 64*128 crossbar is built to perform MNIST image recognition. The experiment results demonstrate that the in-situ training of the MLP achieves a high accuracy near 96.5% with Dropout and Maxout.