Information processing method and apparatus

By pre-deploying reference coding information and quickly finding approximate coding information, the problem of high memory access and computational requirements of feedforward networks in NLP models is solved, achieving more efficient information processing and generalization capabilities.

CN122197991APending Publication Date: 2026-06-12HUAWEI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HUAWEI TECH CO LTD
Filing Date
2024-12-12
Publication Date
2026-06-12

AI Technical Summary

Technical Problem

The excessive memory and computational requirements of the feedforward network in existing NLP models have become a bottleneck for training.

Method used

By pre-deploying multiple reference coding information, approximate coding information can be quickly found, reducing the memory access and computation requirements of the target feedforward network. Fast lookup operations are used to replace matrix multiplication operations, and linear mapping networks are only used when the differences between approximate coding information are large. Combined with vector databases and dynamic update mechanisms, the generalization ability is improved.

Benefits of technology

It effectively reduces the memory access and computation requirements of feedforward networks in NLP models, and improves the output accuracy and generalization ability of the models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122197991A_ABST
    Figure CN122197991A_ABST
Patent Text Reader

Abstract

The application provides an information processing method and device, and belongs to the technical field of artificial intelligence. The information processing method comprises the following steps: inputting to-be-processed coded information into a target feedforward network; wherein the target feedforward network comprises at least one layer of linear mapping network, and the to-be-processed coded information is output information of a previous attention network of the target feedforward network; determining approximate coded information of the to-be-processed coded information from candidate coded information of the target feedforward network; determining an information processing result of the to-be-processed coded information based on a processing result corresponding to the approximate coded information; and the processing result corresponding to the approximate coded information is obtained by performing feature transformation processing on the approximate coded information by the target feedforward network. The application can reduce the memory access demand and the calculation demand of the feedforward network part in the NLP model.
Need to check novelty before this filing date? Find Prior Art