Representational Difference Explanations

RDX enables direct comparison of how different AI models "think": The method visualizes differences between learned representations, allowing researchers to see what makes one model's internal decision-making process fundamentally different from another's—something existing explanation tools struggle to do effectively.

The method recovers meaningful distinctions that other explainability techniques miss: When tested on models with known conceptual differences, RDX successfully identified important distinctions where standard post hoc explainable AI (XAI) methods failed, suggesting it provides insights unavailable through current tools.

Practical applications reveal both obvious and subtle model differences: Testing on challenging subsets of ImageNet and iNaturalist datasets showed RDX can uncover both clear representational differences between models and more subtle patterns in how data influences different systems.

Addresses a critical gap in responsible AI evaluation: While scientific analysis fundamentally relies on comparison, machine learning currently lacks effective tools for this purpose. RDX helps fill this gap, making it easier for practitioners to understand and audit how different models differ in their decision-making processes.

Supports more transparent model development and selection: By making model comparisons more interpretable and direct, this work enables stakeholders to make more informed decisions about which models to deploy based on their actual representational differences, rather than relying solely on performance metrics.