New XAI evaluation framework focuses on real-world CNNs with few classes

Researchers have introduced a new framework for evaluating the faithfulness of Explainable Artificial Intelligence (XAI) techniques, specifically tailored for Convolutional Neural Network (CNN) classifiers used in real-world applications with limited classes. The proposed method generates in-distribution, uncertainty-provoking perturbations to more accurately measure how well XAI methods reflect the model's decision-making process. This evaluation framework was demonstrated on both medical and natural imaging datasets, highlighting the importance of domain, data curation, and XAI method selection for validating new CNN models. AI

IMPACT Provides a more robust method for assessing the reliability of AI explanations in practical applications.

RANK_REASON This is a research paper proposing a new evaluation framework for XAI techniques. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Wistan Marchadour

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →