Neural Network Model Compression for Edge AI Through Weight Approximation Exponent Sharing and Retraining

Abstract

The rise of edge AI (Artificial Intelligence) applications has brought exciting possibilities for tasks like computer vision and natural language processing on resource-constrained devices. These devices, often with limited memory and battery power, struggle to run large traditional neural networks. To address this challenge, model compression techniques have become a major focus of research. Existing methods like quantization, weight sharing, and pruning can achieve significant size reduction, but often at the cost of some accuracy loss. This trade-o! between size and accuracy becomes a critical bottleneck for deploying these powerful models on edge devices. newline