数据处理方法、介质及电子设备
By using the multiplication coefficients of the quantized multiplication operator as fixed-point scaling coefficients, the problem of large errors in the quantized multiplication operator is solved, thus improving the data processing accuracy of the neural network model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ARM TECH CHINA CO LTD
- Filing Date
- 2023-03-27
- Publication Date
- 2026-07-17
AI Technical Summary
When using quantized multiplication operators for data processing, there is a problem of large error in the results, which leads to a decrease in the accuracy of neural network models.
By obtaining the floating-point scaling factor and its reciprocal of the multiplication relation coefficients in the multiplication operator, and quantizing them into fixed-point scaling factors, the fixed-point multiplication relation coefficients are obtained and used to calculate the multiplication operator, thus avoiding the computational difficulties caused by infinitely large or infinitely small denominators in the floating-point domain.
This effectively ensures the data processing accuracy of the integration operator and improves the data processing accuracy of the neural network model.
Smart Images
Figure CN116339678B_ABST