Terminal Neural Network Personalization via Local Annotation Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing offline neural network models trained on cloud cannot meet personalized user requirements due to being universal and lack of adaptation to individual user needs.

Innovation Solution

A method for training a neural network model on a terminal device using a high-precision second neural network model to generate annotation data, updating a lower-precision first neural network model based on the high-precision model's parameters, and incorporating user feedback to enhance personalization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a universal offline neural network model is trained on cloud, then the model can be deployed on terminal devices with optimized memory and power consumption, but the model cannot meet personalized user requirements

Engineering Contradiction:
Improvepersonalization capabilityVSAvoidmodel training complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent divides the training process into two segments: cloud-based pre-training of a universal model, and device-based fine-tuning using locally collected annotation data. This segmentation allows the model to achieve personalization without requiring complete retraining on the cloud, thus reducing complexity while improving adaptability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by pre-training the neural network model on the cloud with general data before deploying it to terminal devices. This preliminary training provides a solid foundation that enables subsequent local fine-tuning with minimal computational resources, facilitating personalization without excessive complexity.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If a high-precision neural network model is used for offline inference, then the inference accuracy improves, but the computational resources and power consumption increase

Engineering Contradiction:
Improveinference accuracyVSAvoidpower consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent changes the precision parameter of the neural network model based on usage context. The model uses high precision (e.g., FP32) for offline inference tasks requiring accuracy, and can switch to lower precision for online inference tasks where speed is prioritized. This dynamic parameter adjustment balances accuracy requirements with power consumption constraints.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If the neural network model is updated frequently to meet user requirements, then the model performance improves, but the upgrade frequency increases system complexity

Engineering Contradiction:
Improvemodel performanceVSAvoidupgrade management complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent implements self-service by enabling terminal devices to automatically collect annotation data from user interactions and perform local fine-tuning of the model. This self-updating mechanism allows the model to adapt to user requirements without requiring frequent manual updates from the server, thereby improving performance while reducing upgrade management complexity.

Inventive Principle:
Principle #25Self-service

4Measurement precision

If more annotation data is collected for model training, then the model accuracy improves, but the data processing time and storage requirements increase

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential annotation data from user interactions that are most valuable for fine-tuning the model. Rather than processing all collected data, the system selectively extracts high-value annotation samples, reducing data processing time and storage requirements while maintaining model accuracy improvement.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3716156B1Neural network model training method and apparatus
Publication Date: 2025.06.25 HUAWEI TECH CO LTD
  • EP3716156B1 patent drawingFigure 1a~1b
  • EP3716156B1 patent drawingFigure 2
  • EP3716156B1 patent drawingFigure 3

AI summary

This application provides a method for training a neural network model and an apparatus. The method is applied to a terminal device, the terminal device includes a first neural network model and a second neural network model that are used to process a service, precision of the first neural network model is lower than precision of the second neural network model, and the method includes: obtaining annotation data that is of the service and that is generated by the terminal device in a specified period (301); 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 (302); and updating the first neural network model based on the trained second neural network model (303). 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.