Attribute-Verified Classification for Explainable Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification systems using deep learning struggle with misrecognition issues, lack of transparency in recognition results, and the inability to systematically explain the reasoning behind classifications, limiting their practical utility and necessitating compensation for errors.
Innovation Solution
A classification system that learns not only correct classes but also attributes, incorporating a verification mechanism to check against a class-specific attribute database, allowing for systematic explanation of classifications and estimation of unseen classes through parallel classification means.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning classification systems are used to improve recognition accuracy, then classification accuracy is improved, but the system cannot provide transparent explanation of recognition grounds and is prone to misrecognition
Solution Approach 1:
The patent segments the classification system into multiple independent classification means (first classification means and second classification means), each trained with different training data. This segmentation allows the system to process information through multiple pathways and aggregate results, thereby maintaining accuracy while reducing the risk of misrecognition and providing more robust classification grounds.
Solution Approach 2:
The patent merges the results from multiple classification means by aggregating their output values. The aggregation process combines the strengths of different classification approaches, producing a final classification result that is more accurate and reliable than any single classification means alone, while also providing comprehensive grounds for the classification decision.
2Productivity
If deep learning systems process large amounts of data to improve classification performance, then classification performance is improved, but the systems lack the ability to systematically explain classification reasoning
Solution Approach 1:
The patent implements a feedback mechanism where classification results from multiple classification means are aggregated and used to determine the final classification. This feedback loop allows the system to systematically evaluate and combine results, providing transparent grounds for classification decisions while maintaining high performance through comprehensive data processing.
3Device complexity
If single classification means are used to simplify the system, then device complexity is reduced, but misrecognition occurs and user accountability is compromised
Solution Approach 1:
The patent employs multiple classification means as a form of prior cushioning against misrecognition. By having redundant classification pathways trained on different data, the system creates a buffer that protects against errors from any single classification means, thereby improving reliability without requiring overly complex integration mechanisms.
Data Source
AI summary
A classification system includes classification training data-specific classification data correct attribute setting configured to set not only the correct classes of pieces of classification training data but also the correct attributes of the pieces of classification training data and classification/attribute training configured to train classification by comparing the class/attribute estimation result outputs with the correct classes set by correct class setting and comparing the attribute estimation result outputs with the correct attributes set by the classification training data-specific classification training data correct attribute setting. Pieces of classification target data from classification target data acquisition are newly inputted to the trained classification means. Classification/attribute result output outputs both the classification results and attribute results of the pieces of classification target data so that validity of the classification results can be considered.