Researchers have developed ConceptSMILE, a new model-agnostic framework designed to audit the trustworthiness of concept-based explanations in artificial intelligence. This framework extends existing perturbation-based logic to evaluate concept reliability by measuring shifts in concept responses after perturbing input regions. ConceptSMILE assesses explanations through metrics like attribution accuracy, surrogate fidelity, faithfulness, stability, and consistency. Initial evaluations on retinal fundus images, comparing MedSAM and vision-language model (VLM) derived concepts, indicate varying reliability across different concepts and pathways. AI
IMPACT Provides a new method for evaluating the reliability of AI explanations, potentially improving trust in AI systems.
RANK_REASON The cluster contains a research paper detailing a new framework for auditing AI explanations.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
