The paper titled ‘Decoupling Pixel Flipping and Occlusion Strategy for Consistent XAI Benchmarks’ explores the issue of feature removal in Explainable AI (XAI). It highlights how occlusion strategies can vary and how this ambiguity limits the usefulness of occlusion-based approaches. The study proposes two perspectives to address this issue. First, it introduces the R(eference)-Out-of-Model-Scope (OMS) score to measure reliability and provide a systematic comparison of occlusion strategies. Second, it combines the most influential first (MIF) and least influential first (LIF) measures into the symmetric relevance gain (SRG) measure to achieve consistent rankings irrespective of the underlying occlusion strategy. The paper verifies these proposals using 40 different occlusion strategies.
Publication date: 12 Jan 2024
Project Page: https://arxiv.org/abs/2401.06654v1
Paper: https://arxiv.org/pdf/2401.06654
