Terminal Device Pattern Recognition Model Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing terminal devices face high power consumption and insufficient computing capacity when running pattern recognition algorithms, and trained models lack personalization, requiring large computing resources and accuracy improvements.

Innovation Solution

A method and apparatus that optimize a universal pattern recognition model by collecting local samples on a terminal device, using these samples to adjust the model's features, and applying specific training algorithms to create a personalized model without uploading data to the server, thereby reducing server computing load and enhancing user-specific recognition accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Power

If algorithm instructions are run concurrently by using a plurality of threads on a general-purpose computing unit, then the computing capability meets the requirement of large computing amount, but power consumption is extremely high

Engineering Contradiction:
Improvecomputing capabilityVSAvoidpower consumption
Core Design Contradiction:
PowerVSUse of energy by moving object

Solution Approach 1:

The patent segments the model training process into two distinct phases: cloud-based pre-training using massive parallel computing resources, and terminal-based fine-tuning using localized small-sample learning. This segmentation allows the computationally intensive tasks to be performed in the cloud while the terminal device only performs lightweight optimization, thereby reducing power consumption while maintaining computing capability requirements.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by performing comprehensive model pre-training in the cloud before deploying to terminal devices. The universal model is pre-trained on diverse data in the cloud, and then terminal devices only need to perform incremental fine-tuning on local small samples, significantly reducing the computing burden and power consumption on terminal devices.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If model training is performed in cloud using hundreds of millions of samples, then recognition accuracy is relatively high, but the model lacks personalized features for specific users

Engineering Contradiction:
Improverecognition accuracyVSAvoidpersonalization capability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies local quality by maintaining the universal model's broad recognition capabilities while adding localized personalization through small-sample fine-tuning on terminal devices. The model has different quality characteristics: cloud-trained universal features for general accuracy and terminal-trained local features for user-specific personalization, combining both advantages.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent introduces an intermediary mechanism where the terminal device acts as a bridge between the cloud-based universal model and user-specific personalization needs. The terminal collects local samples, performs fine-tuning, and adapts the universal model to specific user characteristics, serving as an intermediary that reconciles general accuracy with personalized adaptability.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If the terminal device collects local samples and optimizes the universal model locally, then personalized recognition is achieved and server computing load is reduced, but the terminal device requires additional computing resources

Engineering Contradiction:
Improvepersonalization efficiencyVSAvoidterminal computing resources
Core Design Contradiction:
ProductivityVSPower

Solution Approach 1:

The patent applies parameter changes by transitioning from full-model training with massive datasets to fine-tuning with small local samples. The training parameters are fundamentally changed: sample size reduced from hundreds of millions to thousands, training duration shortened, and computational complexity reduced, enabling personalization on terminal devices with limited resources.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses a disposable approach where the terminal device performs lightweight fine-tuning using minimal local samples that are processed and discarded after extracting personalized features. This avoids the need for expensive, long-running training processes on terminal devices, achieving personalization with minimal resource consumption.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Data Source

PatentEP3460792B1Optimization method and terminal device suitable for model of pattern recognition
Publication Date: 2021.10.27 HUAWEI TECH CO LTD
  • EP3460792B1 patent drawingFigure 1
  • EP3460792B1 patent drawingFigure 2~3
  • EP3460792B1 patent drawingFigure 4~5

AI summary

Embodiments of the present invention relate to a method and an apparatus for optimizing a model applicable to pattern recognition, and a terminal device. The terminal device receives a universal model delivered by a server, where the universal model includes an original feature parameter; recognizes target information by using the universal model, and collects a plurality of local samples; when a model optimization condition is met, corrects the original feature parameter by using a first training algorithm to obtain a new feature parameter; and optimizes the universal model according to a second training algorithm and the new feature parameter, to obtain an optimized universal model. That is, in the present invention, the terminal device further optimizes, according to the collected local samples, the universal model received from the server to obtain a relatively personalized model applicable to pattern recognition. Therefore, not only user experience is improved, but also a problem of a large computing amount of the server that exists when the server optimizes the universal model is resolved.