Binary Weighted Memristive Analog Deep Neural Network for Near-Sensor Edge Processing

Abstract

The memristive crossbar aims to implement analog weighted neural networks, but realistic implementation is constrained by the limited switching states of memristive devices. In this work, an analog deep neural network with binary weight updates using binary-state memristive devices is proposed via a backpropagation learning algorithm. It enables image processing tasks with lower power consumption and reduced on‑chip area compared to digital counterparts. The network benchmarked on MNIST achieved approximately 90% accuracy.