PASTA-dataset

PASTA-dataset是由U2IS, ENSTA Paris, Institut Polytechnique de Paris等机构创建的，用于评估计算机视觉中XAI技术的人类感知系统。该数据集包含四个多样化的子数据集（COCO, Pascal Part, Cats Dogs Cars, Monumai），共计100张图像，每张图像都有图像级和概念级的标注。数据集的创建过程涉及训练多种分类器模型并计算XAI解释，最终形成包含图像、标签和解释的三元组。PASTA-dataset主要应用于XAI技术的评估和比较，旨在解决深度学习模型在关键应用中的可解释性和透明度问题。

PASTA-dataset was developed by institutions including U2IS, ENSTA Paris, Institut Polytechnique de Paris, and others, to evaluate human perception systems for Explainable Artificial Intelligence (XAI) technologies in computer vision. The dataset includes four diverse sub-datasets: COCO, Pascal Part, Cats Dogs Cars, and Monumai, with a total of 100 images. Each image is annotated at both the image-level and concept-level. The construction process of PASTA-dataset involves training multiple classifier models and computing XAI explanations, ultimately forming triplets composed of images, labels, and explanations. PASTA-dataset is primarily used for the evaluation and comparison of XAI technologies, aiming to address the issues of interpretability and transparency of deep learning models in critical application scenarios.

搜集汇总

数据集介绍

构建方式

PASTA-dataset is constructed by integrating four diverse datasets—COCO, Pascal Parts, Cats Dogs Cars, and Monumai—each annotated at both image and concept levels. This dual-level annotation framework allows for robust evaluation across various XAI methods. The dataset includes 21 XAI methods across multiple model architectures, with explanations subjected to rigorous human evaluation along comprehensive criteria. The final dataset comprises a triplet of images, explanations, and labels, enabling quantitative assessment of XAI techniques.

特点

The PASTA-dataset is characterized by its large-scale and diverse composition, encompassing a wide range of XAI methods and model architectures. It includes both image-based and concept-based annotations, facilitating the evaluation of different explanation modalities. The dataset's human-centric evaluation protocol ensures that explanations are assessed not only for technical soundness but also for alignment with human cognitive processes and expectations.

使用方法

The PASTA-dataset can be utilized to benchmark and compare various XAI techniques across different modalities. Researchers and practitioners can use this dataset to evaluate the interpretability and usefulness of XAI methods by analyzing the human-annotated scores and the associated explanations. Additionally, the dataset supports the development of automated metrics that mimic human assessments, providing a scalable and reliable way to evaluate XAI techniques in a human-aligned manner.

背景与挑战

背景概述

PASTA-dataset, introduced in the paper 'BENCHMARKING XAI EXPLANATIONS WITH HUMAN-ALIGNED EVALUATIONS', represents a pioneering effort in the realm of Explainable Artificial Intelligence (XAI). Developed by a consortium of researchers from prestigious institutions such as ENSTA Paris, University of Trento, TU Berlin, and others, the dataset aims to address the critical need for human-centric evaluation of XAI techniques in computer vision. The creation of PASTA-dataset is rooted in the growing necessity to interpret the decision-making processes of Deep Neural Networks (DNNs), especially in high-stakes domains like legal and medical fields. The dataset integrates four diverse datasets—COCO, Pascal Parts, Cats Dogs Cars, and MonumAI—to form a comprehensive benchmark for evaluating XAI methods across various modalities. This initiative underscores the importance of perceptual evaluation in XAI, bridging the gap between computational metrics and human cognitive processes.

当前挑战

The development of PASTA-dataset presents several significant challenges. Firstly, the integration of multiple datasets with varying characteristics necessitates a robust framework to ensure consistent evaluation across different modalities. Secondly, the human-centric evaluation process, which involves extensive human annotation, poses challenges related to cost, time, and consistency due to potential biases and subjective factors. Thirdly, the dataset must address the challenge of standardizing the evaluation of diverse XAI techniques, which currently lack unified benchmarks. Additionally, the perceptual aspect of evaluation, which mimics human preferences, introduces complexities in automating the scoring process while maintaining alignment with human assessments. Finally, the dataset faces the challenge of scalability, ensuring that the evaluation framework can be applied to a wide range of XAI methods and future advancements in the field.

常用场景

经典使用场景

PASTA-dataset 在计算机视觉领域中被广泛用于以人为中心的可解释人工智能（XAI）技术评估。其经典使用场景包括对 XAI 解释方法在图像和概念层面的评估，通过结合 COCO、Pascal Parts、Cats Dogs Cars 和 MonumAI 四个多样化的数据集，构建了一个大规模的基准数据集。该数据集允许对各种 XAI 方法进行鲁棒评估和比较，特别是在解释的感知质量和人类偏好方面。

实际应用

PASTA-dataset 在实际应用中主要用于开发和验证新的 XAI 技术，特别是在需要高度透明度和用户信任的领域，如医疗、法律和金融。通过提供一个标准化的评估框架，该数据集帮助开发者在不同模型架构和数据集上测试其解释方法的有效性。此外，PASTA-dataset 还支持在多模态解释方法上的比较，这在之前的研究中尚未得到充分解决。

衍生相关工作

PASTA-dataset 的引入催生了一系列相关的经典工作，特别是在 XAI 评估方法的标准化和自动化方面。例如，基于 PASTA-dataset 的研究已经开发出多种自动化评分模型，这些模型能够模拟人类对解释的评估，从而加速了 XAI 技术的迭代和改进。此外，PASTA-dataset 还促进了在多模态解释方法上的研究，推动了 XAI 技术在不同应用场景中的广泛采用。