Enhancing Post-Hoc Explanation Benchmark Reliability for Image Classification
This work addresses the problem of unreliable evaluation for researchers and practitioners using post-hoc explanation methods in image classification, though it is incremental as it builds on existing benchmarking practices.
The paper tackled the inconsistency of faithfulness metrics in benchmarking post-hoc explanation methods for image classification by applying Krippendorf's alpha to quantify reliability and proposing model training modifications like perturbed samples and focal loss, resulting in significant improvements in benchmark reliability across various metrics, datasets, and methods.
Deep neural networks, while powerful for image classification, often operate as "black boxes," complicating the understanding of their decision-making processes. Various explanation methods, particularly those generating saliency maps, aim to address this challenge. However, the inconsistency issues of faithfulness metrics hinder reliable benchmarking of explanation methods. This paper employs an approach inspired by psychometrics, utilizing Krippendorf's alpha to quantify the benchmark reliability of post-hoc methods in image classification. The study proposes model training modifications, including feeding perturbed samples and employing focal loss, to enhance robustness and calibration. Empirical evaluations demonstrate significant improvements in benchmark reliability across metrics, datasets, and post-hoc methods. This pioneering work establishes a foundation for more reliable evaluation practices in the realm of post-hoc explanation methods, emphasizing the importance of model robustness in the assessment process.