Uncertainty Index Selection for Learning Model Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing learning model accuracy improvement techniques often select data with the highest degree of uncertainty, leading to biased re-learning and reduced accuracy due to excessive noise, which can result in decreased model performance.
Innovation Solution
An information processing apparatus and method that calculates an index value indicating the degree of uncertainty of predictions from multiple learning models, selects data within a predetermined selection range that balances uncertainty, and outputs these selected data for annotation, thereby preventing bias and efficiently improving model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data with the highest degree of uncertainty is selected for re-learning, then the learning model can be re-trained to improve accuracy, but the selection becomes biased and introduces excessive noise that reduces model performance
Solution Approach 1:
The patent changes the selection criterion from extreme uncertainty (maximum index value) to moderate uncertainty (within a predetermined range). By adjusting the parameter of uncertainty selection from 'highest' to 'moderate', the system avoids selecting data that is either too easy or too difficult, thereby reducing bias and noise while maintaining improvement effectiveness.
Solution Approach 2:
Instead of selecting all data with the highest uncertainty (excessive action), the patent applies partial action by selecting only data with moderate uncertainty within a predetermined range. This partial selection approach filters out extreme cases that would introduce noise while still capturing enough uncertain data to improve model accuracy.
2Productivity
If data with the highest degree of uncertainty is selected for annotation, then more learning data can be added to improve the model, but the amount of useful learning data increases less efficiently due to bias
Solution Approach 1:
The patent changes the uncertainty index selection parameter from 'maximum' to 'within predetermined range'. This parameter adjustment ensures that a broader range of moderately uncertain data is selected, increasing the quantity of useful learning data added per cycle while maintaining high improvement efficiency by avoiding extreme outliers.
Data Source
AI summary
An information processing apparatus includes: a calculation unit that adds an index value indicating a degree of uncertainty of a prediction, to each of a plurality of instances, on the basis of the prediction for each of the plurality of instances respectively outputted from a plurality of learning models; a selection unit that selects at least one instance, of which the added index value is included in a predetermined selection range, from the plurality of instances; and an output unit that outputs the selected at least one instance.