REPAIR

REPAIR数据集是由加州大学圣地亚哥分校的研究人员创建，旨在通过数据重采样技术减少现代机器学习数据集中的表示偏差。该数据集通过优化问题形式化，寻求一个权重分布，惩罚基于给定特征表示构建的分类器容易处理的数据示例。REPAIR通过交替更新分类器参数和数据重采样权重，使用随机梯度下降解决这一问题。数据集主要应用于动作识别等领域，旨在提高模型在处理真实世界任务时的泛化能力。

The REPAIR dataset was developed by researchers at the University of California, San Diego, to mitigate representation bias in modern machine learning datasets via data resampling techniques. Formulated as an optimization problem, it seeks a weight distribution to penalize data samples that are easily handled by classifiers constructed based on given feature representations. REPAIR solves this problem using stochastic gradient descent by alternately updating classifier parameters and data resampling weights. Primarily applied in fields such as action recognition, this dataset aims to improve the generalization ability of models when handling real-world tasks.