The QUCE method addresses the challenge of diminishing interpretability in Deep Neural Networks by minimizing path uncertainty. It quantifies uncertainty in explanations and generates more certain counterfactual examples. By comparing with competing methods, QUCE showcases superior performance in both path-based explanations and generative counterfactual examples. The method relaxes straight-line path constraints, providing a more flexible approach to generating paths towards alternative outcomes. QUCE utilizes single and multiple-paths approaches to offer generalized explanations over all paths for an instance, enhancing interpretability and reliability.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Jamie Duell,... kl. arxiv.org 03-15-2024
https://arxiv.org/pdf/2402.17516.pdfDybere Forespørgsler