Method for operating a neural network model, medium, program product, and electronic device

By numerically converting and quantizing the floating-point input data of exponential function operations in the neural network model, a lookup table with higher accuracy is constructed, which solves the problem of low accuracy of fixed-point exponential function operations and improves the accuracy of the model's operation results.

CN114612996BActive Publication Date: 2025-03-21ARM TECH CHINA CO LTD
1 Cites 0 Cited by

Patent Information

Application Number
CN202210335080.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-31
Publication Date
2025-03-21
Estimated Expiration
2042-03-31

AI Technical Summary

Technical Problem

In neural network model, for the quantization of exponential functions, as the floating-point input data increases, the calculation results of fixed-point exponential functions are lower, resulting in lower accuracy of model operation results.

Method used

By numerical conversion of floating-point input data of exponential function operations, the range becomes less than or equal to zero, and then quantization and lookup tables are performed to build a lookup table with higher accuracy to improve the accuracy of the calculation results.

Benefits of technology

By adjusting the input data range and building a lookup table with higher accuracy, the problem of low accuracy of exponential function calculation results is solved, and the accuracy of operation results of neural network models is improved.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present application relates to the field of artificial intelligence technology, and in particular to an operation method, medium, program product and electronic device of a neural network model. The method comprises: during the operation of the neural network model, receiving first floating-point data to be subjected to exponential function operation; performing numerical conversion on the first floating-point data to obtain second floating-point data belonging to a preset floating-point data range, the preset floating-point data range being less than or equal to zero; quantizing the second floating-point data to obtain a first fixed-point number of the second floating-point data; using the first fixed-point number as an index value to search a lookup table for a first exponential fixed-point result value corresponding to the first fixed-point number. The second exponential fixed-point result value of the exponential function operation of the first floating-point data is determined based on the first exponential fixed-point result value. The lookup table constructed in the present application has high precision, so that the result of the exponential function quantization operation based on the lookup table with higher precision is more accurate, which can improve the accuracy of the model operation result.
Need to check novelty before this filing date? Find Prior Art