Róisín Luo, James McDermott, Colm O'Riordan
This paper presents a semi-optimal sampling approach for image attribution analysis that improves explanation certainty by aligning sample distributions with natural image distributions, outperforming state-of-the-art methods.
In the field of image attribution analysis, researchers aim to understand which parts of an image are most important for a computer model's decision-making. This study reveals that current methods, which often add noise to images, may not provide the most reliable explanations. Instead, the authors propose a new method that adjusts the sampling process to better match natural image distributions, leading to more trustworthy and accurate explanations. Their approach was tested on a large dataset and showed improved performance over existing techniques.
