Convolutional neural network mixed-precision quantization method considering layer sensitivity

To address the problem of how to faithfully map neural networks to resource-constrained embedded devices, a mixed-precision quantization method for convolutional neural networks based on layer sensitivity analysis was proposed. The sensitivity of convolutional layer parameters was measured by calculating the average trace of the Hessian matrix, providing a basis for bit-width allocation. A layer-wise ascending-descending approach was employed for bit-width allocation, ultimately achieving mixed-precision quantization of the network model. Experimental results demonstrate that compared to the fixed-precision quantization methods DoReFa and LSQ +, the proposed mixed-precision quantization method improves recognition accuracy by 10.2% and 1.7%, respectively, at an average bit-width of 3 bit. When compared to other mixed-precision quantization methods, the proposed approach achieves over 1% higher recognition accuracy. Additionally, noise-injected training effectively enhances the robustness of the mixed-precision quantization method, improving recognition accuracy by 16% under a noise standard deviation of 0.5.