Research on Mixed-Precision Quantization Optimization Method for Neural Networks Based on Quantization Sensitivity and Hardware Constraints

In mixed-precision quantization of neural network models, determining the quantization bit-width for each layer is a critical challenge. This paper proposes an optimization method that comprehensively considers quantization sensitivity and hardware constraints. By establishing quantitative relationships between quantization bit-width and hardware resources (computational overhead and bandwidth requirements), the problem is transformed into an optimization problem and solved using a hierarchical greedy algorithm. Experimental results demonstrate that the proposed method effectively reduces quantization-induced accuracy loss while satisfying hardware resource constraints.