This paper proposes an index-based kernel transformation to reduce the memory requirements and energy consumption of convolutional neural networks (CNNs). This transformation is implemented on networks with quantized weights represented by an index while maintaining fixed-point precision. The proposed algorithm eliminates redundant operations by extracting common index patterns from different kernels, performing identical operations only once. A specifically designed hardware is implemented to perform the convolution and rebuild the correct results from the extracted patterns in parallel. The experiment shows that deploying an extensive network like VGG-16 on the proposed hardware only requires 2.620KB of on-chip memory, and importantly, it exhibits superior energy efficiency compared to the state-of-the-art. This promising result suggests a practical solution to the growing demands on memory and operations in large CNNs.
Track ID:
9.6
Track Name:
Efficient and Reliable AI Accelerator Design and Optimization for Deep Neural Networks
