TWO4TWO

TWO4TWO数据集是由柏林自由大学创建的，专门设计用于用户研究，特别是评估解释性视觉方法在发现模型偏差任务上的表现。该数据集包含两个抽象动物类别，Peeky和Stretchy，以及多种可控制的属性，如动物颜色、背景、形状和身体姿势。数据集的生成过程允许研究人员精确控制和量化模型对不同属性的依赖性，从而为评估解释技术的效果提供了一个基准。TWO4TWO数据集的应用领域主要集中在机器学习模型的可解释性和偏差检测，旨在帮助用户更好地理解和信任模型的决策过程。

The TWO4TWO dataset, developed by the Free University of Berlin, is specifically designed for user studies, particularly for evaluating explanatory visual methods in model bias detection tasks. It contains two abstract animal categories, Peeky and Stretchy, along with multiple controllable attributes such as animal color, background, shape and body posture. The dataset's generation process enables researchers to precisely control and quantify a model's dependence on different attributes, thereby providing a benchmark for assessing the effectiveness of explanation techniques. The main application fields of the TWO4TWO dataset focus on machine learning model interpretability and bias detection, aiming to help users better understand and trust the decision-making processes of models.

原始信息汇总

Two4Two 数据集概述

数据集描述

Two4Two 数据集是用于研究用户是否从可解释的视觉模型中受益的用户研究、基线和数据集。该数据集包含偏倚和无偏倚两种版本。

下载链接

• 偏倚数据集
• 无偏倚数据集

使用方法

数据集的生成、模型训练和用户输入生成可以通过以下命令进行：

bash python -m two4two --download_blender config/dataset_config.toml dubfiv_train --device cuda --output_base_dir ./models --experiment config/model.toml dubfiv_export_user_study --model ./models/<model_dir>

代码位置

• 数据生成代码：dubfiv/data_generation.py
• 训练循环代码：dubfiv/train.py

引用

@inproceedings{ sixt2022do, title={Do Users Benefit From Interpretable Vision? A User Study, Baseline, And Dataset}, author={Leon Sixt and Martin Schuessler and Oana-Iuliana Popescu and Philipp Wei{ss} and Tim Landgraf}, booktitle={International Conference on Learning Representations}, year={2022}, url={https://openreview.net/forum?id=v6s3HVjPerv} }