Perceptual Model Updating with Uncertainty and Inconsistency Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting target data from original data for training perceptual models include ineffective data, leading to overlearning of outliers and inappropriate label prediction, thus requiring inefficient use of time and resources.
Innovation Solution
A perceptual model update device that classifies target data based on uncertainty and inconsistency scores, manually labels uncertain data, and updates the model with labeled and unlabeled data to improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If target data is extracted from original data on the basis of uncertainty, then data effective in training a perceptual model can be included as much as possible and time and work for making training data can be reduced, but the extracted target data includes many outliers of the entire original data and the perceptual model may overlearn outliers
Solution Approach 1:
The patent segments target data into three categories based on uncertainty and inconsistency scores: (1) high uncertainty/high inconsistency data labeled manually, (2) low uncertainty/high inconsistency data labeled by labeling model, and (3) low uncertainty/low inconsistency data used as is. This segmentation resolves the contradiction by processing different data types differently, improving efficiency while maintaining reliability.
Solution Approach 2:
The patent introduces inconsistency score as an additional parameter alongside uncertainty score to filter and classify data. By changing the classification parameters from single-criterion (uncertainty only) to multi-criterion (uncertainty and inconsistency), the system can identify and handle outliers appropriately, preventing overlearning while maintaining training efficiency.
2Ease of manufacture
If target data is extracted from original data at random or on the basis of representativeness, then the extraction process is simple, but the target data will include a predetermined percentage of data ineffective in training a perceptual model and time and work for making training data cannot be reduced
Solution Approach 1:
The system uses the perceptual model's own uncertainty predictions to automatically identify and select effective training data without requiring manual inspection or complex external criteria. The model serves itself by indicating which data points need attention, achieving simple yet effective data extraction.
3Measurement precision
If manual annotation is performed on all target data, then labeling accuracy can be ensured, but much time and work is required
Solution Approach 1:
The patent applies different labeling strategies to different portions of data based on their characteristics: high-uncertainty data receives manual annotation for accuracy, while low-uncertainty data is processed automatically. This local differentiation ensures labeling accuracy where needed while minimizing time loss through automation where appropriate.
Solution Approach 2:
Instead of manually annotating all data (excessive action), the system performs manual annotation only on the necessary subset of high-uncertainty data. This partial action approach maintains labeling accuracy for critical data while significantly reducing overall annotation time and resources.
Data Source
AI summary
A perceptual model update device inputs each of pieces of target data and modified data obtained by making a predetermined modification to the piece of target data into a perceptual model to predict a label corresponding to the piece of target data and to calculate an uncertainty score and an inconsistency score; classifies the pieces of target data into first, second, and third data, based on the uncertainty score and the inconsistency score; gives an input label inputted via an input device to the first data as a label of the first data; gives a label to the second data by a predetermined labeling model; and updates the perceptual model by training the perceptual model with the first and second data given the label and the third data.