Design of Adaptive Learning Quantitative Training Algorithm Based on Deep Neural Network Algorithm
Shu Xu, Chenxiao Li, Xinyue Cao, Huilin Huang · 2023
In the differentiated machine learning method, the number of constraints imposed by the training sample on the parameters is equal to the number of bits required for it to specify the label. Labels usually contain very little bit information, and if the number of parameters exceeds the number of training samples, it will usually lead to serious over-fitting. In this paper, the adaptive learning quantization training algorithm based on DNN (Deep Neural Network) algorithm is designed. For our algorithm, first, we train the floating-point model on the training set. When the floating-point model meets the precision requirement, we regard the floating-point model weight as the pre-training model of quantization convolution training, and train on the same training set. Finally, the quantization convolution model is used as a pre-training model to train the quantization activation layer. The results show that the accuracy of the proposed quantization model is 94.359% on the CIFAR-10 data set and 81.204% on the CIFAR-100 data set. The parameter of the quantization model in this paper is 6.159 MB. Experiments show that the algorithm proposed in this paper can achieve real-time reasoning performance on edge embedded devices with extremely limited computing and memory resources.