DNN Model Parameter Integration for Risk-Aware Misclassification Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network (DNN) parameter correction methods do not adequately consider the risk levels of different types of misclassification, leading to insufficient or unintended increases in safety risks in safety-critical applications.
Innovation Solution
A machine learning system that identifies suspicious parameters for each type of misclassification using defect localization and metaheuristic optimization, generating integrated models by adjusting parameters based on risk levels through an objective function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional parameter correction methods are used to reduce misclassification, then prediction accuracy improves, but safety cannot be improved because risk levels of different misclassification types are not considered
Solution Approach 1:
The patent segments misclassification errors into different types based on risk levels (e.g., safe misclassifications vs. unsafe misclassifications). Instead of treating all misclassifications uniformly, the system identifies and corrects parameters specifically for high-risk misclassification types while preserving correct classifications, thereby improving safety without sacrificing overall accuracy.
Solution Approach 2:
The patent applies local quality by correcting parameters locally for specific high-risk misclassification cases rather than globally across all parameters. The correction process targets specific parameters that cause unsafe misclassifications while maintaining other parameters that contribute to correct classifications, achieving selective improvement in safety-critical areas.
2Measurement precision
If parameters are corrected to reduce all types of misclassification, then classification accuracy improves, but unintended safety risks may increase due to uniform correction approach
Solution Approach 1:
The patent segments misclassification errors into different types based on risk levels (e.g., safe misclassifications vs. unsafe misclassifications). Instead of treating all misclassifications uniformly, the system identifies and corrects parameters specifically for high-risk misclassification types while preserving correct classifications, thereby improving safety without sacrificing overall accuracy.
Solution Approach 2:
The patent changes parameters selectively based on the type of misclassification. Rather than uniformly adjusting all parameters to reduce any misclassification, the system modifies specific parameters only when they contribute to high-risk misclassifications, thereby avoiding unintended safety risks while improving classification accuracy for critical cases.
3Measurement precision
If defect localization technique is used to identify parameters affecting misclassification, then parameter identification becomes more precise, but the complexity of the correction process increases
Solution Approach 1:
The patent segments the correction process into distinct stages: (1) identifying high-risk misclassification types, (2) localizing parameters affecting these specific misclassifications using defect localization, and (3) correcting only those identified parameters. This segmentation makes the complex process more manageable and systematic.
Solution Approach 2:
The patent performs preliminary classification of misclassification types by risk level before applying defect localization. By pre-identifying which misclassification types are high-risk, the system narrows the scope of subsequent parameter analysis, reducing the overall complexity of the correction process while maintaining precision.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A model generation apparatus includes a parameter identification unit that identifies a parameter affecting misclassification of misclassification data incorrectly classified by a trained classification model, among parameters of the classification model for each type of the misclassification, based on the misclassification data incorrectly classified by the trained classification model; a parameter correction unit that generates a correction parameter obtained by correcting the parameter identified by the parameter identification unit for each type of the misclassification; and a parameter integration unit that generates an integrated model including an integrated parameter obtained by integrating the correction parameters for each type of the misclassification.