High-Accuracy Deep Belief Network: Fuzzy Neural Networks Using MRAM-Based Stochastic Neurons

Abstract

Two-terminal magnetoresistive random access memory (MRAM) devices provide a recent approach to intrinsically-realizing stochastic neuronal behavior in cognitive architectures such as restricted Boltzmann machines (RBMs) for deep belief networks (DBNs). MRAM-based DBNs have achieved substantial energy and area improvements in comparison with the prior DBN hardware implementations. However, MRAM-based DBNs suffer from low accuracy in image classification. In order to resolve this problem, we present a new DBN-fuzzy system based on the combination of MRAM-based DBNs and fuzzy systems in the interest of improving the accuracy of MRAM-based neural networks. First, the MRAM-based DBN is employed to identify the top recognition results with the highest probability. Second, the fuzzy system is utilized to obtain the top-1 recognition results. We assess the accuracy of neural networks on MNIST dataset, finding that the top-1 accuracy of the DBN-fuzzy neural network is improved from 64% to 82% in comparison to the MRAM-based DBNs for a 784 × 10 network.