Explainable Multi-Scale Gaussian Mixture Model Distance for Saliency Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Explainable Artificial Intelligence (XAI) methods lack robustness in comparing model saliencies due to sensitivity to individual input and pixel anomalies, and poor convergence properties on smaller datasets, especially in complex machine learning models like deep neural networks.
Innovation Solution
The introduction of an explainable multi-scale Gaussian mixture model distance (XGMD) metric, which uses a weighted Gaussian mixture model to encode pixel-level intensity and calculate a 2-Wasserstein distance, providing a robust similarity measure by learning high-order probability distributions across multiple scales and being less sensitive to dataset size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current XAI methods are used to compare model saliencies, then the comparison process is simple, but the results are sensitive to individual input and pixel anomalies and show poor convergence on smaller datasets
Solution Approach 1:
The patent introduces an explainable multi-scale Gaussian mixture model distance (XGMD) metric as an intermediary to compare saliency maps. Instead of directly comparing pixel values which is sensitive to anomalies, the XGMD metric transforms saliency maps into Gaussian mixture models and computes distance between these probabilistic representations. This intermediary transformation makes the comparison robust to individual pixel anomalies and improves convergence on smaller datasets by capturing the overall distribution rather than individual deviations.
Solution Approach 2:
The patent moves the comparison from the pixel intensity dimension to the probability distribution dimension. By fitting Gaussian mixture models to saliency maps and computing distances in the parameter space of these distributions (means, covariances, mixing coefficients), the method operates in a different dimensional space that is less sensitive to pixel-level noise and anomalies while maintaining the ability to distinguish meaningful differences between saliency maps.
2Measurement precision
If traditional similarity metrics are used, then computational speed is fast, but sensitivity to noise and dataset size limitations reduces measurement precision
Solution Approach 1:
The patent applies Gaussian mixture models at multiple scales rather than a single scale. By computing XGMD metrics at different spatial resolutions and combining these partial results, the method achieves more precise and robust comparisons. The multi-scale approach captures both fine-grained and coarse-grained similarities, improving measurement precision without requiring excessively large datasets or computational resources at any single scale.
Data Source
AI summary
An example apparatus includes interface circuitry, machine-readable instructions, and at least one processor circuit to be programmed by the machine-readable instructions to access a first saliency map and a second saliency map associated with an image dataset, encode pixel-level intensity of the first saliency map, encode pixel-level intensity of the second saliency map, generate a saliency comparison metric based on the pixel-level intensity of the first saliency map and the pixel-level intensity of the second saliency map, and compare spatial properties of the first saliency map and the second saliency map using the saliency comparison metric.