Machine Learning Concept Images for Stage-Wise Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for visualizing the determination basis of machine learning models, such as Grad-CAM and LIME, are difficult for humans to understand and not suitable for real-time processing, while methods like TCAV may indicate different bases for varying recognition targets, making it challenging to correct errors.
Innovation Solution
Presenting a determination basis on a multi-stage concept basis, generating concept images for each stage of a machine learning model, identifying contribution degrees, and correcting errors at specific stages using a concept correction unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If Grad-CAM or LIME is used to visualize determination basis, then visualization of determination basis is achieved, but human understanding becomes difficult and real-time processing is not suitable
Solution Approach 1:
The patent introduces concept images as an intermediary between the machine learning model's internal features and human understanding. Instead of directly visualizing complex feature maps or using black-box explanation methods, the system generates concept images that represent activated concepts at each stage, making the determination basis interpretable while preserving the model's decision-making information
Solution Approach 2:
The patent segments the machine learning model into multiple stages and generates concept images for each stage separately. This segmentation allows users to understand the determination basis at different processing levels, from early feature detection to final classification, making the overall complex process comprehensible in manageable parts
2Measurement precision
If TCAV is used to calculate concept importance, then concept contribution is identified, but different bases are indicated for varying recognition targets making error correction challenging
Solution Approach 1:
The patent performs preliminary error analysis by examining concept images at each stage before final classification. By identifying where concept activation deviates from expected patterns at intermediate stages, the system can correct errors early in the processing pipeline rather than attempting to correct them after final classification, simplifying the correction process
Data Source
AI summary
Provided is an information processing apparatus that presents a determination basis of a machine-learned model on a concept basis. An information processing apparatus includes: a generation unit that generates a concept image identified in each stage of a machine learning model including a plurality of stages; an identification unit that identifies a contribution degree of a concept in each stage when the machine learning model processes an input image on a basis of activation of the concept image; and a presentation unit that presents a concept serving as a determination basis in each stage of the machine learning model on a basis of an identification result by the identification unit.