Uncertainty Calculation for AI Determination Reliability

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques for evaluating the reliability of artificial intelligence determination results in image classification are inadequate, as they do not effectively assess the uncertainty of the determination model.

Innovation Solution

A determination evaluation device and method that calculates the uncertainty of the determination result using a machine learning model and evaluates the reliability of the result based on this uncertainty, allowing for improved assessment of the model's performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual assignment of correct labels to images is performed to prepare accurate train data, then the accuracy of the determination model is improved, but the time consumption and work load increase significantly

Engineering Contradiction:
Improveaccuracy of determination modelVSAvoidtime consumption for manual labeling
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The determination model performs self-evaluation by calculating uncertainty values for its own determination results. The model automatically identifies regions where its confidence is low and flags these for potential review, eliminating the need for comprehensive manual labeling while maintaining accuracy in high-confidence regions.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system introduces uncertainty as a new parameter to evaluate determination results. By calculating uncertainty values and comparing them against threshold values, the system automatically adjusts which regions require manual verification, reducing overall labeling workload while maintaining model accuracy.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If insufficient train data is used due to high manual labeling costs, then the productivity increases, but the determination accuracy decreases in certain regions

Engineering Contradiction:
Improvetrain data preparation efficiencyVSAvoiddetermination accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system implements a feedback mechanism where uncertainty calculations from the determination model feed back into the data preparation process. Regions with high uncertainty are automatically identified and prioritized for labeling, creating an efficient feedback loop that maximizes the value of limited manual labeling resources.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

By introducing uncertainty threshold values as a new parameter, the system can automatically filter and prioritize which regions need manual labeling. This parameter-driven approach ensures that limited labeling resources are allocated to the most critical regions, maintaining determination accuracy while improving productivity.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If comprehensive manual labeling is performed to ensure high determination accuracy, then the reliability of determination results improves, but the ease of operation decreases due to complex manual work

Engineering Contradiction:
Improvereliability of determination resultVSAvoidease of creating train data
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The determination model automatically evaluates its own results by calculating uncertainty values, eliminating the need for complex manual verification processes. The system self-identifies unreliable regions and flags them appropriately, making the workflow much easier while maintaining reliability.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system uses uncertainty threshold values as an automated quality control parameter. By comparing uncertainty against these thresholds, the system automatically determines which results are reliable enough without manual review, significantly easing operations while maintaining the necessary reliability standards.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4564243A1Determination evaluation device, method, and program
Publication Date: 2025.06.04 HITACHI LTD
  • EP4564243A1 patent drawingFigure 1~2
  • EP4564243A1 patent drawingFigure 3~6
  • EP4564243A1 patent drawingFigure 7

AI summary

It is evaluated how reliable a determination result by a determination model is. A determination evaluation device that evaluates a determination result for input data based on a determination model includes an uncertainty calculation unit that calculates uncertainty of the determination result, and an analysis evaluation unit that evaluates the determination result on the basis of the uncertainty.