Multimodal AI Explainability via Modality-Specific Shapley Reweighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI explainability tools, such as LIME, fail to adequately explain predictions generated by multimodal AI models due to the challenge of handling multiple modalities like text, images, and numerical data, where distances between features can vary greatly, leading to one modality dominating others and obscuring feature importance.
Innovation Solution
The method generates perturbed instances of multimodal AI models by perturbing encoded features, determines distances using modality-specific metrics, converts these distances to weights with kernel functions, and applies modality-specific Shapley values to obtain final weights for an interpretable surrogate model, which is then tuned using hyperparameters to enhance explanation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing explainability tools like LIME are used to explain multimodal AI models, then the explanation process can be implemented, but the explanations become inaccurate due to modality dominance where one modality overwhelms others
Solution Approach 1:
The patent applies local quality by using modality-specific distance metrics and kernel functions for different data types (text, images, numerical data). Each modality receives tailored treatment in the explanation process, preventing any single modality from dominating the explanation while ensuring accurate feature importance attribution for each specific modality.
2Ease of manufacture
If distances between features are computed directly in multimodal data, then the calculation is simple, but features from different modalities cannot be compared fairly due to varying scales
Solution Approach 1:
The patent transforms the distance calculation process by introducing modality-specific kernel functions that map distances from different modalities to a common scale. This parameter transformation allows fair comparison across modalities while maintaining computational feasibility, converting the raw distance metric into a normalized weight that reflects true feature importance across diverse data types.
3Productivity
If a single distance metric is used for all modalities, then the computation is efficient, but the explanation fails to capture modality-specific contributions
Solution Approach 1:
The patent segments the explanation process into modality-specific components, with separate distance metrics and kernel functions for text, images, and numerical data. This segmentation allows each modality to be processed according to its specific characteristics, capturing modality-specific contributions while maintaining overall computational efficiency through parallel processing of independent modality components.
Data Source
AI summary
Explaining decisions or predictions of an AI model can include generating perturbed instances by perturbing one or more encoded features of a multimodal AI model instance. For each of the perturbed instances, a distance between one or more encoded features of each perturbed instance and a corresponding one or more of the encoded features of the multimodal AI instance can be determined. Each distance can be converted to a weight using a kernel function. For each weight, a modality-specific Shapley value can be determined, and each weight can be adjusted by post-weighting each weight with the modality-specific Shapley value associated with the weight to obtain final weights. An interpretable surrogate model based on the final weights can be output.