PRDML+: A Proxy-based Robust Deep Metric Learning versus Label Noise via Self-Supervised Label Refinement Process

Abstract

Label noise in real-world datasets can negatively impact the efficacy of deep learning models. Manually correcting labels is labor-intensive and impractical for large datasets. Consequently, various methods have been developed to improve the robustness of deep models against label noise. However, most existing methods are primarily designed for classification problems and are not readily applicable to similarity learning applications without adaptation. We tackle this issue by proposing a resilient representation learning approach that models observed labels as a mixture of clean and noisy label distributions. Then, our methods identify label noise data and a semantic embedding jointly using Expectation Maximization (EM) approach. Meanwhile, it progressively modifies targets of data by incorporating a modified self-adaptive training mechanism into the EM algorithm, thereby improving the quality of the semantic embedding using the refined labels. Comprehensive evaluations on datasets containing actual or artificial noisy labels demonstrate that the proposed approach consistently surpasses peer methods experiments on datasets containing real or synthetic label noise demonstrate that our approach consistently surpasses peer methods. Furthermore, it effectively refines noisy labels during the initial training epochs.