Beyond the Veil of Similarity: Quantifying Semantic Continuity in Explainable AI

2024-07-17

xAI

AI Summary

Existing model explanations in explainable AI (XAI) suffer from semantic discontinuity and concept drift—i.e., inconsistent conceptual interpretations across semantically similar inputs—undermining interpretability and trust. Method: We propose the first formal framework for quantifying *semantic continuity*, defined as the robustness of explanation concepts to small input perturbations. Our approach introduces a differentiable metric grounded in concept-space mapping, integrating Concept Activation Vectors (CAVs), adversarial perturbation analysis, and semantic embedding alignment. We further design a continuity-aware gradient regularization strategy for training explanation models. Results: Evaluated on ImageNet and COCO-XAI benchmarks, our method improves explanation stability by 37.2%, significantly enhancing human interpretability and decision reliability. This work overcomes the limitation of conventional similarity-based evaluation metrics—which ignore semantic drift—and establishes a new paradigm for rigorous assessment and optimization of explanation continuity in XAI.

Technology Category

Humans and AI: Explainable AI (XAI) for Human UnderstandingComputer Vision: Interpretability, Explainability, and TransparencyPhilosophy and Ethics of AI: Accountability, Interpretability & Explainability

Application Category

Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactionsSearch and Retrieval-Augmented AI: Web evaluation methodologies and metricsUser Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Problem

• Artificial Intelligence
• Explanation Stability
• Semantic Coherence

Innovation

• Stability Assessment
• Explainable AI
• Semantic Coherence