Inpainting Improves Visual XAI by Generating Photorealistic Perturbations

A new research paper explores how to improve eXplainable Artificial Intelligence (XAI) for image-based machine learning models by refining perturbation-based explanation methods. Traditional perturbation techniques, such as replacing pixels with uniform colors, often create unrealistic, out-of-distribution samples that can mislead the model and compromise the quality of the resulting explanations. These artifacts make it difficult to accurately understand why a model makes a particular prediction. The researchers propose an enhancement to LIME, a widely used perturbation-based XAI method, by integrating generative inpainting. This technique allows for the creation of photorealistic perturbed samples that maintain a closer resemblance to the original data distribution. By generating more natural-looking alterations, the method aims to provide more accurate and reliable insights into the model's decision-making process, ultimately elevating the quality of visual XAI explanations.