Partial Dataset Learning for Multi-Class Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning-based image recognition technologies face high costs and inefficiencies in data labeling, particularly in multi-class classification, where assigning correct answers to all recognition objects is labor-intensive and costly, especially as the number of classes increases.
Innovation Solution
A learning device and method that utilize a dataset composed of multiple partial datasets, where each partial dataset is assigned a responsible range of categories, allowing for the use of correct answer data indicating belonging or non-belonging categories, enabling the computation of an intersection matrix to calculate recognition loss and update model parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If correct answers are applied to all recognition objects in multi-class classification, then recognition accuracy is improved, but labeling cost and time increase significantly
Solution Approach 1:
The patent divides the complete dataset into multiple partial datasets, where each partial dataset contains only a subset of categories. This segmentation allows annotators to focus on specific categories within each partial dataset, significantly reducing the time and cost required for labeling while maintaining comprehensive category coverage across all partial datasets combined.
Solution Approach 2:
Instead of requiring complete labeling of all categories for every recognition object, the patent applies partial labeling where each object is labeled only for the categories relevant to its partial dataset. This partial action approach reduces the overall labeling workload while still enabling effective model training through the aggregation of multiple partial datasets.
2Measurement precision
If correct answers are applied to all recognition objects in multi-class classification, then recognition accuracy is improved, but labeling cost increases
Solution Approach 1:
The patent segments the labeling task into multiple partial datasets, each covering a subset of categories. This division reduces the per-dataset labeling cost by allowing specialized annotators to work on specific category subsets, while the combined cost of labeling multiple partial datasets is lower than labeling a single complete dataset with all categories.
Solution Approach 2:
The patent employs partial labeling where each recognition object receives labels only for categories within its assigned partial dataset rather than all categories. This partial action reduces the total quantity of labeling work required, thereby reducing labeling costs while maintaining sufficient training data quality through the aggregation of multiple partial datasets.
3Productivity
If multiple partial datasets are used for learning, then labeling workload is reduced, but loss computation becomes complex
Solution Approach 1:
The patent introduces an intersection matrix as an intermediary component that bridges multiple partial datasets. The intersection matrix encodes the relationships and overlaps between categories across different partial datasets, enabling the loss computation unit to efficiently aggregate losses from multiple partial datasets without requiring complex manual coordination. This intermediary structure simplifies the overall computation process while maintaining the benefits of partial dataset labeling.
Data Source
AI summary
The dataset supply unit supplies a learning dataset. The recognition unit outputs the recognition result for the recognition object data in the supplied learning dataset. Further, the intersection matrix computation unit computes the intersection matrix based on the learning dataset. The recognition loss computation unit computes the recognition loss using the recognition result, the intersection matrix, and the correct answer data given to the recognition object data. Then, the updating unit updates the parameters of the recognition unit based on the recognition loss.