Master and Custom Model Management for Accuracy-Preserving Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems for managing trained models in machine learning applications lack efficient mechanisms for updating models while maintaining recognition accuracy and user-specific custom models, leading to potential degradation in performance when updating master models.
Innovation Solution
A trained model management device and method that includes a first storage for master models, a second storage for custom models, and an update determination unit to decide whether to update the master model based on new training data, allowing for seamless integration of new recognition targets without degrading existing recognition accuracy, and enabling users to continue using custom models alongside updated master models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the master model is updated with new training data, then the recognition model can identify new target objects, but the recognition accuracy for existing target objects may deteriorate
Solution Approach 1:
The patent segments the model management system into master models (for new target objects) and custom models (for existing target objects). This segmentation allows independent management and updating of different model types, enabling the master model to be updated with new training data without affecting the performance of custom models for existing objects.
Solution Approach 2:
The patent extracts custom models from the master model framework. When a user needs to maintain recognition accuracy for specific existing target objects, the system extracts and stores custom models separately in the second storage, allowing these models to be preserved and reused even when the master model is updated with new training data for different objects.
2Adaptability or versatility
If the master model is continuously updated with new training data, then new recognition targets can be added, but the system complexity and computational resources increase
Solution Approach 1:
The patent implements preliminary action by pre-training master models with comprehensive training data that covers multiple target objects. This allows the master model to be updated with new training data for new objects without requiring complete retraining from scratch, reducing computational resources and system complexity.
Solution Approach 2:
The patent uses copying by storing custom models in the second storage that can be copied and reused across different users and applications. This eliminates the need to retrain models for existing target objects, significantly reducing computational resources and system complexity when deploying recognition systems for multiple users.
3Measurement precision
If custom models are stored for each user, then user-specific recognition accuracy is maintained, but storage requirements and data management complexity increase
Solution Approach 1:
The patent implements universality by designing the second storage to store custom models that serve multiple users and applications. A single custom model stored in the second storage can be copied and used by multiple users for the same target objects, reducing storage requirements and data management complexity while maintaining user-specific recognition accuracy.
Data Source
AI summary
A trained model management device and a trained model management method that update master models without degrading recognition accuracy in the use environment of a user. The trained model management device (10) includes: a first storage (12A) storing first training data and a first model that is a trained model that has been trained to recognize, based on the first training data, a target object included in input information; a second storage (12B) storing second training data and a second model that is a trained model generated based on the second training data and the first model; and an update determination unit (132) that makes a determination as to whether or not to update the first model based on the second learning data when the second model is generated.