Memristive Neural Network with Efficient In-Situ Supervised Training

Abstract

Artificial Neural Networks (ANNs) are trained and tested with empirical data in a supervised and unsupervised manner. During the learning of an ANN in hardware, matrix vector multiplication and the outer product of vectors are costly operations in terms of time and space. Memristor-based crossbars are one of the most prominent candidates for implementing these operations efficiently. Here, a memristive neural network is proposed where a synapse is realized by employing two memristors only. No transistor as a controlling element is associated at the cross point of the memristive crossbar. An in-situ training architecture to train a memristive multilayered neural network (MNN) is proposed where online supervised gradient descent back propagation is implemented and all the synaptic weights are updated in (1) time. The proposed training algorithm has been tested on simulated memristive MNN and single layer neural network (SNN) on the IRIS data set and Breast Cancer Wisconsin data set with a classification accuracy of 99.11% and 90.14%, respectively. Further, with 10% of the memristors stuck at maximum conductance state, a classification accuracy of 94.67% on the IRIS data-set shows the robustness of the proposed method.