Method and apparatus for training neural networks

By training low-precision neural networks on cloud servers, the problems of high computational cost, long processing time, high power consumption, and pseudo-texture phenomena of pixel-level deep neural networks on edge devices are solved, achieving low power consumption and high imaging quality.

CN115700598BActive Publication Date: 2026-07-10HUAWEI TECH CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
HUAWEI TECH CO LTD
Filing Date
2021-07-23
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Pixel-level deep neural networks have high computational cost, long processing time, and high power consumption on edge devices. Furthermore, low-precision quantization can easily lead to pseudo-texture phenomena, affecting imaging performance.

Method used

By training a low-precision neural network based on first-order and second-order information on a cloud server, and utilizing the powerful computing capabilities of the cloud, a low-precision neural network can be generated that can perform the same task on edge devices while reducing the phenomenon of false textures.

Benefits of technology

It reduces the power consumption of edge devices and effectively eliminates the pseudo-texture phenomenon caused by low-precision calculations, thereby improving the imaging quality of pixel-level tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115700598B_ABST
    Figure CN115700598B_ABST
Patent Text Reader

Abstract

The application provides a neural network training method and device. The neural network training method of the application comprises the following steps: a cloud server receives training data and a first neural network from a terminal device, the first neural network being a neural network for a pixel-level task; a second neural network is trained according to first first-order information extracted from the first neural network and second first-order information extracted from a second neural network to obtain a trained second neural network, the second neural network being obtained based on the first neural network, the operation precision of the second neural network being lower than that of the first neural network, and the first neural network and the second neural network performing the same pixel-level task; and the trained second neural network is sent to the terminal device. The application can reduce the power consumption of the terminal device, effectively solve the pseudo-texture phenomenon caused by low operation precision, and improve the imaging quality of the neural network for the pixel-level task.
Need to check novelty before this filing date? Find Prior Art