AMC: Adaptive Mixed Compression for ML Models based on Block-Wise Sensitivity

This is the authors' abstract. We don't add key points for this paper.

This paper proposes Adaptive Mixed Compression (AMC), a novel model compression algorithm designed to reduce computational costs and memory usage in neural networks while preserving performance. AMC integrates both mixed precision quantization (MPQ) and mixed sparsity pruning (MSP) at a block level, dynamically adjusting bit widths and sparsity rates based on sensitivity analysis. By continuously monitoring accuracy degradation, AMC effectively balances compression with minimal performance loss. Experimental results on various models such as YOLOv5, VGG16, and BERT show that AMC achieves up to 16.07% higher compression compared to traditional uniform methods, with superior accuracy retention across diverse tasks. This highlights AMC’s potential to enhance the efficiency of machine learning systems in resource-constrained environments.