A Reliability-Concerned Compute-in-Memory Behavior Model for Convolutional Neural Network

A reliability-concerned compute-in-memory (CIM) behavior model is developed to evaluate and optimize the performance of convolutional neural network (CNN), like LeNet, VGG. Non-ideal factors, such as cell delay variation and mismatch are concerned as parameters in image classification tasks. The inference accuracy will be evaluated under specific dataset, neural network architecture, mapping method with non-ideal factors. In this work, a quantized CNN is designed for image classification as a case study to show how to perform circuit-algorithm evaluation with the developed reliability-concerned CIM behavior model. The accuracy of the quantized neural network for MNIST achieves 97% with variation=0.1 in Gaussian distribution.