[C81] Weight Sharing for Array-Efficient CNN Inference in Compute-In-Memory Architectures

Abstract

In this work, we propose a weight sharing method for compute-in-memory (CIM)-based convolutional neural network (CNN) inference, which reduces the number of mapped weights without relying on pruning. By sharing weights and layer-wise scale factors, our approach reduces overall CIM array usage while maintaining model accuracy. Evaluations on ResNet-20 with 32x32 arrays demonstrate that our method achieves a 28% reduction in array usage with less than 1% accuracy loss.