Machine Model Update via Hard Example Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning lifecycle tools are complex and require significant technical knowledge, making it difficult to efficiently update machine models in real-time, especially in applications like driver monitoring systems, where continuous improvement is necessary for performance enhancement.
Innovation Solution
A method and apparatus for updating machine models by obtaining hard example samples, determining category-based data distribution, selecting second hard example samples based on learning requirements, performing learning operations, and updating the model, thereby forming a closed loop for efficient machine learning lifecycle management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing machine learning lifecycle tools are used, then machine models can be updated, but the tools are complex and require significant technical knowledge
Solution Approach 1:
The system automatically identifies hard example samples from model inference results and performs self-updating through the closed-loop pipeline (identification → data augmentation → model retraining → validation), eliminating the need for complex manual intervention and reducing technical knowledge requirements
Solution Approach 2:
The machine learning lifecycle is segmented into distinct automated modules: hard example identification module, data augmentation module, model retraining module, and validation module. Each module handles a specific task automatically, simplifying the overall system while maintaining update capability
2Reliability
If existing machine learning lifecycle tools are used, then machine models can be updated, but the process is not efficient in real-time applications
Solution Approach 1:
The system performs preliminary data augmentation on hard example samples before model retraining, preparing enhanced training data in advance. This preliminary processing accelerates the overall update pipeline by having ready-to-use augmented data when retraining begins
Solution Approach 2:
The closed-loop pipeline enables continuous automated updates: model inference continuously identifies hard examples, which immediately trigger data augmentation and retraining processes. This continuous cycle eliminates idle time and maintains constant improvement in real-time applications
3Manufacturing precision
If hard example samples are used for model updating, then model performance can be improved, but the process requires manual intervention and is not automatic
Solution Approach 1:
The system uses feedback from model inference results to automatically identify hard example samples that the model struggles with. This feedback loop triggers the entire update pipeline automatically, connecting model performance evaluation directly to targeted retraining without manual intervention
Solution Approach 2:
The system merges multiple functions into an integrated automated pipeline: hard example identification, data augmentation, model retraining, and validation are combined into a single automated workflow that executes continuously without manual intervention, while maintaining focus on improving model performance through hard examples
Data Source
AI summary
Disclosed are a machine model update method and apparatus, a medium, and a device. The method includes: obtaining first hard example samples of a machine model and attribute information of the first hard example samples; determining category-based data distribution information of the first hard example samples according to the attribute information of the first hard example samples; determining second hard example samples of the machine model in current machine learning according to learning requirement information of the machine model for categories of the first hard example samples and the data distribution information; performing, according to learning operation information corresponding to the machine model, a learning operation on the machine model by using the second hard example samples; and updating the machine model based on a learning result of the machine model. The present disclosure helps efficiently implement a machine learning lifecycle, and reduce the cost of machine learning lifecycle.