CNN Inference Error Analysis via Superimposed Attention Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current analysis techniques fail to specify image sections causing incorrect inference with adequate precision in image recognition processes using convolutional neural networks (CNNs).
Innovation Solution
An analysis method that generates a refine image with a maximized correct label score from an incorrect inference image, superimposes maps indicating pixel changes and attention degrees, and specifies pixels causing incorrect inference by calculating pixel values in a map adjusted based on appearance frequency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current analysis techniques (activation maximization, BP, GBP) are used to identify image sections causing incorrect inference, then the inference process can be analyzed, but the precision in specifying the image sections is insufficient
Solution Approach 1:
The patent combines multiple analysis techniques by superimposing the first map (from activation maximization showing pixel changes) and the second map (from BP/GBP showing attention degrees) to create a third map that integrates both types of information. This merging allows the system to leverage the strengths of different methods while achieving higher precision in identifying causative image sections.
Solution Approach 2:
The patent introduces a refine image as an intermediary element that is generated by replacing image sections identified in the third map with corresponding sections from a correctly inferred image. This intermediary refine image serves as a bridge to verify and improve the accuracy of the identified causative sections by comparing inference results before and after the replacement.
2Measurement precision
If the third map is generated by superimposing first and second maps without adjustment, then the processing is simpler, but the accuracy in identifying important pixels is reduced
Solution Approach 1:
The patent adjusts the second map by applying appearance frequency information as a weighting parameter. Pixels that frequently appear in attention regions across multiple images are given higher weights, while rare pixels are downweighted. This parameter adjustment enhances the reliability of the third map by emphasizing consistently important pixels and reducing noise from incidental attention patterns.
Data Source
AI summary
A method includes: generating a refine image having a maximized correct label score of inference from an incorrect image from which an incorrect label is inferred by a neural network; generating a third map by superimposing a first map and a second map, the first map indicating pixels to each of which a change is made in generating the refine image, of a plurality of pixels of the incorrect image, the second map indicating a degree of attention for each local region in the refine image, the each local region being a region that has drawn attention by the neural network; and specifying a set of pixels that cause incorrect inference in the incorrect image by calculating a pixel value of the third map for each set of pixels, wherein the map generating processing adjusts the second map based on appearance frequency of each degree of attention.