Feature Importance Network for Stable Image Classification Visualization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for image classification, particularly deep neural networks, are 'black box' solutions that lack interpretability, leading to difficulties in explaining classification decisions and requiring unstable visualization techniques, slow performance, and loss of classification accuracy.
Innovation Solution
A computing system that generates feature maps and importance vectors through a feature extraction and importance network, combining them to create an attention map for improved visualization and classification outputs, allowing for stable visualization without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network models are used for image classification, then classification accuracy is improved, but interpretability and visibility into inner layers deteriorate
Solution Approach 1:
The patent introduces visualization techniques as intermediary tools that bridge the gap between the black-box neural network and human interpretation. These techniques generate visual representations of feature importance and activation patterns, allowing users to understand model decisions without modifying the underlying network architecture or training process.
Solution Approach 2:
The patent extracts and visualizes specific features and activation patterns from the inner layers of the neural network. By taking out and highlighting important features through visualization techniques, the system makes the hidden decision-making processes visible while preserving the original model's classification accuracy.
2Loss of information
If visualization techniques are added to explain classification decisions, then interpretability is improved, but computational performance deteriorates
Solution Approach 1:
The patent performs feature extraction and visualization computations during the model training phase or as pre-computed caches. By preparing visualization data in advance rather than computing it in real-time during inference, the system minimizes the impact on computational performance while maintaining interpretability capabilities.
3Loss of information
If existing visualization techniques are used, then interpretability is improved, but stability and optimality of visual mappings deteriorate
Solution Approach 1:
The patent incorporates feedback mechanisms where the visualization system continuously refines its mappings based on the relationship between input images, feature activations, and classification outputs. This feedback loop ensures that visual mappings remain stable and optimal by adjusting to the actual behavior of the neural network rather than relying on fixed, pre-defined mappings.
Data Source
AI summary
Systems, methods and computer readable media provide technology to perform image classification and produce visualization using a machine learning architecture. The disclosed image classification and visualization technology includes a feature extraction network to generate a feature map, a feature importance network to generate a feature importance vector, an attention map generated based on a weighted sum of the feature importance vector and the feature map, a classification output determined based on a combination of the attention map and the feature map, and a feature visualization image generated by overlaying the attention map onto an input image. Each of the feature extraction network and the feature importance network can include a neural network.