Proxy-based robust deep metric learning in the presence of label noise

Abstract

Real-world datasets contain label noise data that can deteriorate the performance of a deep learning model. Cleaning annotations manually requires substantial efforts from experts and is not practical in large datasets. Therefore, many methods are proposed to enhance the robustness of deep models in the presence of label noise. However, these methods are primarily developed for classification tasks and cannot be directly applied to Deep Metric Learning (DML) applications. To bridge this gap, we present a proxy-based deep metric learning method to detect wrong labels through the estimation of the posterior distribution for observed and true labels. Specifically, we model the distribution of the observed annotations as a mixture of two components: one component represents the clean label distribution and the other denotes the noisy labels. Besides, we present an EM-like algorithm that precisely identifies label noise data jointly with the similarity learning method. We also exploit information of the identified noisy labeled data by utilizing state-of-the-art semi-supervised learning (SSL) techniques. The results of the extensive experiments on datasets with real or synthetic label noise indicate that our method consistently outperforms the state-of-the-art methods by a large margin. Moreover, the proposed method accurately detects noisy labeled data and generates correct pseudo labels for them after only a few epochs.