Analysis of Memristive Quantized Convolutional Neural Network Accelerator with Device Nonideality

Memristive convolutional neural network accelerator has attracted intensive interest in reducing time and energy consumption. In this article, we analysis the viability of the quantization method using LeNet-5 model on MNIST dataset. Low bit-precision quantization is achieved with slight accuracy loss. The fabricated bilayer AlO x memristor with 3-bit states is used to emulate the synapse, and an accuracy of 97.93% is accomplished within the device nonideality. Furthermore, the device requirements of variation and reliability on inference application are evaluated and proposed.