arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2311.17876cs.CV
收藏

提升图像分类中事后解释基准可靠性的方法

Enhancing Post-Hoc Explanation Benchmark Reliability for Image Classification

Nantes Université(南特大学);École Centrale Nantes(南特中央理工学院);CNRS(法国国家科学研究中心)

Tristan Gomez, Harold Mouchère

AI总结:

本文针对图像分类事后解释基准不可靠的问题,提出用 Krippendorf's alpha 量化可靠性,并通过扰动样本训练和 focal loss 提升模型鲁棒性与校准性,显著改善了多种指标、数据集和事后方法上的基准可靠性。

摘要:

深度神经网络虽然在图像分类中功能强大,但通常以“黑箱”方式运行,使得理解其决策过程变得困难。各种解释方法,尤其是生成显著性图的方法,旨在应对这一挑战。然而,忠实度指标的不一致性问题阻碍了对解释方法进行可靠的基准评估。本文采用一种受心理测量学启发的方法,利用 Krippendorf's alpha 来量化图像分类中事后方法的基准可靠性。研究提出了模型训练修改方案,包括输入扰动样本和使用 focal loss,以增强鲁棒性和校准性。实证评估表明,在不同指标、数据集和事后方法上,基准可靠性均有显著提升。这项开创性工作为事后解释方法领域建立更可靠的评估实践奠定了基础,强调了模型鲁棒性在评估过程中的重要性。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑