Implicit Task-Driven Probability Discrepancy Measure for Unsupervised Domain Adaptation
Implicit Task-Driven Probability Discrepancy Measure for Unsupervised Domain Adaptation
复制标题
DOI：
--
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Mao Li;Kaiqi Jiang;Xinhua Zhang
中科院分区：
文献类型：
--
作者：
Mao Li;Kaiqi Jiang;Xinhua Zhang
概率差异度量是许多机器学习模型的基本构造，例如弱监督学习和生成式建模。然而，大多数测量都忽视了这样一个事实：分布不是学习的最终产品，而是下游预测器的输入。因此，重要的是要翘曲的概率差异措施对最终任务，并朝着这个目标，我们提出了一个新的基于双层优化的方法，使两个分布比较不均匀对整个假设空间，但只相对于最佳预测的下游最终任务。当应用于边缘视差差异和对比域差异时，我们的方法显着提高了无监督域自适应的性能，并且具有更有原则的训练过程。
Probability discrepancy measure is a fundamental construct for numerous machine learning models such as weakly supervised learning and generative modeling. However, most measures overlook the fact that the distributions are not the end-product of learning, but are the input of a downstream predictor. Therefore, it is important to warp the probability discrepancy measure towards the end tasks, and towards this goal, we propose a new bi-level optimization based approach so that the two distributions are compared not uniformly against the entire hypothesis space, but only with respect to the optimal predictor for the downstream end task. When applied to margin disparity discrepancy and contrastive domain discrepancy, our method significantly improves the performance in unsupervised domain adaptation, and enjoys a much more principled training process.