Metric learning aims to learn a distance to measure the difference between two samples, and it plays an important role in pattern recognition tasks. Most of the existing metric learning methods rely on pairs of samples. However, the importance of sample pairs varies greatly because of possible noise and the difference between samples and the decision boundaries. In this paper, we propose a robust hierarchical metric learning (SPHML) framework based on self-paced learning, which can help gain knowledge about the weights of sample pairs and utilize them in an easy or hard manner. Hierarchical nonlinear functions are learned by back-propagation to map sample pairs into a more discriminative feature space. Experimentally, our method achieves very competitive performance when compared with state-of-the-art methods.

度量学习旨在学习一个距离来衡量两个样本之间的差异，它在模式识别任务中起着重要的作用。大多数现有的度量学习方法依赖于成对的样本。然而，样本对的重要性差异很大，因为可能存在噪声以及样本与决策边界之间的差异。在本文中，我们提出了一个基于自定进度学习的强大的分层度量学习（SPHML）框架，它可以帮助获得关于样本对权重的知识，并以简单或困难的方式利用它们。通过反向传播学习分层非线性函数，以将样本对映射到更具辨别力的特征空间中。在实验上，与最先进的方法相比，我们的方法实现了非常有竞争力的性能。