Training method and device of neural network, equipment and storage medium

CN117010461BActive Publication Date: 2026-05-26GUANGDONG OPPO MOBILE TELECOMMUNICATIONS CORP LTD
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
GUANGDONG OPPO MOBILE TELECOMMUNICATIONS CORP LTD
Filing Date
2022-04-26
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing technologies cannot effectively restore the accuracy of compressed neural networks when training them, resulting in low inference results.

Method used

By combining the output of the original neural network with the differences in network parameters, the training method of the compressed neural network is adjusted. The neural network is trained using the first loss value, the second loss value, and the third loss value, focusing on restoring the accuracy of the original neural network after parameter sharing.

Benefits of technology

This improved the accuracy of the compressed neural network and enhanced the inference accuracy of the neural network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117010461B_ABST
    Figure CN117010461B_ABST
Patent Text Reader

Abstract

This application discloses a training method, apparatus, device, and storage medium for a neural network, belonging to the field of machine learning technology. The method includes: determining a first loss value for the current iteration of training based on the first output result of a first neural network on training samples and the labeled target output result in the training samples; the first neural network is obtained by compressing a second neural network that has already been trained; determining a second loss value for the current iteration of training based on the first output result and the second output result of the second neural network on training samples in the current iteration of training; determining a third loss value for the current iteration of training based on the difference in network parameters between the first and second neural networks; and training the first neural network based on the first, second, and third loss values. This method improves the accuracy of neural network inference.
Need to check novelty before this filing date? Find Prior Art