Segmented Machine Learning Model for Incremental Feature Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing machine learning models require significant retraining and increased work costs when new targets are added or changed, especially in scenarios with numerous types of targets, leading to decreased identification accuracy due to the need for retraining the entire model and allocating numerous labels.
Innovation Solution
The approach involves using separate identifiers for each feature of a target object, such as shape and color, which are combined to identify the target, allowing for incremental updates to the model without retraining when new features are added, thereby reducing the computational burden and maintaining identification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire machine learning model is retrained when new targets are added, then identification accuracy is maintained, but work cost and computational load increase significantly
Solution Approach 1:
The machine learning model is divided into multiple independent identifier modules, each responsible for identifying a specific feature type (e.g., shape, color, texture). When new targets are added, only the relevant identifier modules need to be updated rather than retraining the entire model, thus reducing work cost while maintaining identification accuracy.
Solution Approach 2:
The system allows dynamic updating of individual identifier modules independently. Each identifier can be trained and updated separately when new data becomes available for that specific feature type, enabling flexible adaptation without requiring global model retraining.
2Adaptability or versatility
If numerous labels are allocated for numerous types of targets, then comprehensive identification capability is achieved, but device complexity and difficulty of operation increase
Solution Approach 1:
Instead of using a single complex model that handles all target types with numerous labels, the system segments the identification task into multiple independent identifier modules. Each module handles a specific feature type with its own simplified label set, reducing overall model complexity while maintaining comprehensive identification capability.
Solution Approach 2:
Each identifier module is designed to be universal for its specific feature type across all target categories. For example, the shape identifier handles shape features for all types of targets uniformly, rather than having separate identifiers for each target type, thus reducing complexity while maintaining versatility.
3Reliability
If the entire model is retrained when new targets are added, then identification accuracy is maintained, but computational load and processing time increase
Solution Approach 1:
The model is segmented into independent identifier modules that can be trained separately. When new targets are added, only the relevant modules undergo training, significantly reducing computational load and processing time compared to retraining the entire model, while still maintaining identification accuracy through coordinated use of all modules.
Solution Approach 2:
Instead of performing full model retraining (excessive action), the system performs partial training only on the necessary identifier modules. This partial action is sufficient to maintain identification accuracy for new targets while avoiding the unnecessary computational burden of retraining unrelated modules.
Data Source
AI summary
A non-transitory computer-readable recording medium storing a machine learning program causing a computer to execute a process including: generating vector information based on a first feature of a target included in an image, a second feature of the target, and conversion parameters; and executing training of a machine learning model and update of the conversion parameters by inputting the image and the vector information to the machine learning model, the machine learning model including a first machine learning model portion configured to identify the first feature and a second machine learning model portion configured to identify the second feature.