Method for training neural network model and apparatus

Patent metadata

Field	Value
Publication number	US-11521012-B2
Application number	US-202016910289-A
Country	US
Kind code	B2
Filing date	Jun 24, 2020
Priority date	Dec 29, 2017
Publication date	Dec 6, 2022
Grant date	Dec 6, 2022

Abstract

This application provides a method for training a neural network model and an apparatus. The method includes: obtaining annotation data that is of a service and that is generated by a terminal device in a specified period; training a second neural network model by using the annotation data that is of the service and that is generated in the specified period, to obtain a trained second neural network model; and updating a first neural network model based on the trained second neural network model. In the method, training is performed based on the annotation data generated by the terminal device, so that in an updated first neural network model compared with a universal model, an inference result has a higher confidence level, and a personalized requirement of a user can be better met.

First claim

What is claimed is:

1. A method executed by a terminal device for training a neural network model, the terminal device including a first neural network model and a second neural network model that are used to process a service, a precision of the first neural network model being lower than a precision of the second neural network model, the method comprising:
obtaining annotation data that is of the service and that is generated by the terminal device in a specified period;
training the second neural network model by using the annotation data that is of the service and that is generated in the specified period, to obtain a trained second neural network model; and
updating the first neural network model based on the trained second neural network model.

2. The method according to claim 1 ,
wherein after the trained second neural network model is obtained, before the updating of the first neural network model, and with a determination that the trained second neural network model is an initial update of a second neural network model corresponding to a first version, storing the first neural network model, and
wherein after the updating of the first neural network model, the method further comprises:
receiving a second version software package sent by a cloud server, the second version software package including a first neural network model corresponding to a second version; and
with a determination that the first neural network model corresponding to the second version is different from the stored first neural network model, updating the first neural network model corresponding to the second version.

3. The method according to claim 2 , wherein the updating of the first neural network model corresponding to the second version comprises:
obtaining annotation data that is of the service and that is generated in a plurality of historical specified periods;
training, by using the annotation data that is of the service and that is generated in the plurality of historical specified periods, a second neural network model corresponding to the second version, to obtain a trained second neural network model corresponding to the second version; and
updating, based on the trained second neural network model corresponding to the second version, the first neural network model corresponding to the second version.

4. The method according to claim 1 , wherein the terminal device generates the annotation data of the service in the specified period by:
performing an online inference on first input data of the service in the specified period using the first neural network model, to obtain an online inference result;
with a valid feedback of a user for the online inference result being received, generating the annotation data of the service based on the first input data and the valid feedback of the user for the online inference result; and
with the valid feedback of a user for the online inference result not being received, and after determining that a confidence level of the online inference result is greater than a first threshold, generating the annotation data of the service based on the first input data and the online inference result.

5. The method according to claim 1 , wherein the terminal device generates the annotation data of the service in the specified period by:
performing an offline inference on second input data of the service in the specified period using a third neural network model, to obtain an offline inference result, wherein a precision of the third neural network model is higher than a precision of the second neural network model, or the third neural network model is the second neural network model; and
with a determination that a confidence level of the offline inference result is greater than a second threshold, generating the annotation data of the service based on the second input data and the offline inference result.

6. The method according to claim 1 , wherein the training of the second neural network model by using the annotation data that is of the service and that is generated in the specified period comprises:
with the terminal device being in a charging state, training the second neural network model by using the annotation data that is of the service and that is generated in the specified period.

7. A terminal device, comprising:
a memory storing instructions, a first neural network model, and a second neural network model that are used to process a service, a precision of the first neural network being lower than a precision of the second neural network model; and
one or more processors in communication with the memory, configured to execute the instructions to:
obtain annotation data that is of the service and that is generated by the terminal device in a specified period;
train the second neural network model by using the annotation data that is of the service and that is generated in the specified period, to obtain a trained second neural network model; and
update the first neural network model based on the trained second neural network model.

8. The terminal device according to claim 7 , wherein the one or more processors are further configured to execute the instructions to:
with a determination that the trained second neural network model is an initial update of a second neural network model corresponding to a first version, store the first neural network model; and
receive a second version software package sent by a cloud server, the second version software package including a first neural network model corresponding to a second version, and with a determination that the first neural network model corresponding to the second version is different from the stored first neural network model, update the first neural network model corresponding to the second version.

9. The terminal device according to claim 8 , wherein the one or more processors are further configured to execute the instructions to:
obtain annotation data that is of the service and that is generated in a plurality of historical specified periods;
train, by using the annotation data that is of the service and that is generated in the plurality of historical specified periods, a second neural network model corresponding to the second version, to obtain a trained second neural network model corresponding to the second version; and
update, based on the trained second neural network model corresponding to the second version, the first neural network model corresponding to the second version.

10. The terminal device according to claim 7 , wherein the one or more processors are further configured to execute the instructions to:
perform an online inference on first input data of the service in the specified period using the first neural network model, to obtain an online inference result;
with a valid feedback of a user for the online inference result being received, generate the annotation data of the service based on the first input data and the valid feedback of the user for the online inference result; and
with the valid feedback of the user for the online inference result not being received, and after determining that a confidence level of the online inference result is greater than a first threshold, generate the annotation data of the service based on the first input data and the online inference result.

11. The terminal device according to claim 7 , wherein the one or more processors are further configured to execute the instructions to:
perform an offline inference on second input data of the service in the specified period using a third neural network model, to obtain an offline inference result, wherein a precision of the third neural network model is higher than a precision

Assignees

• Huawei Tech Co Ltd

Inventors

• Ma Tao
• Su Qing
• Jin Ying

Classifications

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06F18/22

  Matching criteria, e.g. proximity measures · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title