arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2303.06167cs.CVcs.CYcs.LG
收藏

用微调数据覆盖预训练偏见

Overwriting Pretrained Bias with Finetuning Data

Princeton University(普林斯顿大学)

Angelina Wang, Olga Russakovsky

AI总结:

本文研究预训练模型偏见向微调模型传播的问题,发现通过对微调数据进行轻微干预即可纠正虚假关联和群体代表性不足两类偏见,且几乎不影响性能。

摘要:

迁移学习具有优势,因为它允许在大规模数据集上预训练的模型所具备的表达性特征,被微调用于规模更小、更具领域特定性数据集的目标任务。然而,令人担忧的是,这些预训练模型可能带有自身的偏见,并会传播到微调后的模型中。在本研究中,我们考察了两类概念化的偏见:一是目标任务与敏感属性之间的虚假关联,二是数据集中特定群体的代表性不足。在这两种偏见概念下,我们发现:(1)在预训练模型基础上进行微调的模型确实可能继承其偏见;但(2)通过对微调数据集进行相对轻微的干预,这种偏见可以得到纠正,而且通常对性能的影响可以忽略不计。我们的研究结果表明,精心整理微调数据集对于减少下游任务中的偏见非常重要,而且这样做甚至可以弥补预训练模型中的偏见。

英文摘要:

Transfer learning is beneficial by allowing the expressive features of models pretrained on large-scale datasets to be finetuned for the target task of smaller, more domain-specific datasets. However, there is a concern that these pretrained models may come with their own biases which would propagate into the finetuned model. In this work, we investigate bias when conceptualized as both spurious correlations between the target task and a sensitive attribute as well as underrepresentation of a particular group in the dataset. Under both notions of bias, we find that (1) models finetuned on top of pretrained models can indeed inherit their biases, but (2) this bias can be corrected for through relatively minor interventions to the finetuning dataset, and often with a negligible impact to performance. Our findings imply that careful curation of the finetuning dataset is important for reducing biases on a downstream task, and doing so can even compensate for bias in the pretrained model.

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。

↑