Bias Adjustment for ML Models Using Accuracy-vs-Data Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional technology fails to appropriately execute processing related to training of a model by using a recognizer trained on adjusted data, making it difficult to generate information indicating the relationship between the accuracy of the model and the data, and to enable training of a desired model.
Innovation Solution
A bias adjustment device and method that adjusts the number of training data pieces to generate accuracy information, measuring and predicting the model's accuracy across different categories, and providing a graphical representation to guide data adjustments for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If conventional accuracy measurement is used, then the accuracy of the recognition target is obtained, but it is impossible to generate information indicating the relationship between model accuracy and data, or to enable training of a desired model by adjusting the data
Solution Approach 1:
The patent implements feedback by measuring accuracy for multiple different data amounts and using this measured information to guide subsequent data adjustment decisions. The accuracy measurement unit measures accuracy for each data amount in the plurality of different data amounts, and this measured accuracy information is fed back to determine the optimal data amount for training, enabling iterative improvement of the model training process.
Solution Approach 2:
The patent performs preliminary accuracy measurements for multiple different data amounts before final model training. By measuring accuracy for each data amount in advance and comparing results, the system determines the optimal data amount beforehand, allowing practitioners to prepare appropriate training datasets before committing to final model training, thus avoiding wasted computational resources.
2Measurement precision
If data amount is increased to improve model accuracy, then model accuracy may improve, but training time and computational resources increase
Solution Approach 1:
The patent applies partial action by measuring accuracy for a plurality of different data amounts that are less than the total available data. Instead of always using all available data, the system selectively determines the sufficient data amount needed to achieve desired accuracy, avoiding the excessive action of always training with maximum data and thus reducing unnecessary training time and computational waste.
Solution Approach 2:
The patent changes the data amount parameter systematically by preparing multiple datasets with different data amounts and measuring accuracy for each. This parameter variation approach allows identification of the optimal data amount where further increases would yield diminishing returns, enabling practitioners to select the most efficient data amount for training rather than defaulting to maximum data.
3Measurement precision
If comprehensive accuracy measurement for multiple data amounts is performed, then optimal data amount can be determined, but measurement and processing complexity increases
Solution Approach 1:
The patent segments the accuracy measurement process by dividing it into discrete measurements for multiple different data amounts. Instead of attempting to measure all possible data amounts continuously, the system divides the data amount range into distinct segments and measures accuracy at each segment point, making the complex measurement task manageable and systematic while still providing comprehensive information for optimization.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A bias adjustment device according to the present disclosure is a bias adjustment device for an identification model by machine learning using training data, and includes a calculation unit that calculates first identification accuracy of the identification model trained on first training data and second identification accuracy of the identification model trained on second training data acquired by an adjustment of the number of pieces of data of the first training data, a prediction unit that predicts a change in identification accuracy with respect to the number of pieces of training data on the basis of the first identification accuracy and the second identification accuracy, and a control unit that adjusts the number of pieces of data used for the training or changes the identification model, on the basis of the predicted change in the identification accuracy, in such a manner that the predicted change in the identification accuracy becomes a predetermined target value.