BP-Triplet-Net

三重损失是深度度量学习 (DML) 方法之一，用于学习嵌入，其中来自同一类的示例比来自不同类的示例更接近。在DML的激励下，我们从贝叶斯学习的角度为无监督域适应 (UDA) 提出了一种有效的BP-三重损失，并将该模型命名为BP-三重网络。在以前的UDA基于度量学习的方法中，跨域的样本对被平等对待，由于域偏差，这是不合适的。在我们的工作中，考虑到成对样本对特征学习和域对齐的重要性不同，我们从贝叶斯学习的角度推导了有效UDA的BP-三重损失。我们的BP-三重损失调整了域内和域间成对样本的权重。特别是，它可以自行处理硬对 (包括硬正对和硬负对)。加上常用的领域对齐的对抗性损失，目标伪标签的质量也逐步提高。我们的方法实现了理想源和目标假设的低联合误差。然后，可以按照Ben-David定理将预期的目标误差上限。对四个基准数据集的综合评估证明了该方法对UDA的有效性Triplet loss is one of the deep metric learning (DML) methods for learning embeddings, where examples from the same class are closer than those from different classes. Motivated by DML, we propose an effective BP-triplet loss for unsupervised domain adaptation (UDA) from the perspective of Bayesian learning, and name the corresponding model BP-Triplet Network. In previous metric-learning-based UDA methods, cross-domain sample pairs are treated equally, which is inappropriate due to domain bias. In this work, considering that paired samples exhibit differing importance for feature learning and domain alignment, we derive the BP-triplet loss for effective UDA from the Bayesian learning perspective. Our BP-triplet loss adjusts the weights of intra-domain and inter-domain paired samples. Specifically, it can automatically handle hard pairs, including hard positive pairs and hard negative pairs. Coupled with the commonly used adversarial loss for domain alignment, the quality of target pseudo labels is gradually enhanced. Our method achieves a low joint error of the ideal source and target hypotheses. Then the expected target error can be upper-bounded according to the Ben-David Theorem. Comprehensive evaluations on four benchmark datasets demonstrate the effectiveness of our method for UDA.

搜集汇总

背景与挑战

背景概述

BP-Triplet-Net是一个用于无监督域适应的深度学习数据集，提出基于贝叶斯学习的BP-三重损失方法，通过调整域内和域间样本权重及处理硬对来优化特征学习和域对齐，结合对抗性损失提升目标伪标签质量。该数据集由安徽大学、阿里巴巴和重庆大学于2022年联合发布，相关论文和代码已公开。