Active Contrastive Learning With Noisy Labels in Fine-Grained Classification
Byeongil Kim, Byung Chul Ko · 2024
In real-world scenarios mirroring fine-grained datasets, labeling may result in noisy labels due to ambiguous data characteristics. This study introduces a new classification approach that integrates active and contrastive learning to address the issue of highly noisy labels in fine-grained datasets. We strategically employ active learning to select boundary points near class-dependent decision boundaries, effectively expanding the distances between classes in the data space. By combining contrastive learning and active learning, we successfully extract robust features from the data. Our experiments consistently achieve high average performance across various levels of label noise, highlighting the potential of our approach in fine-grained datasets. This underscores the method's capability to maintain reliable performance even in the presence of label noise, contributing to the field of fine-grained classification and noisy label mitigation. This innovative approach provides promising insights for enhancing the performance of machine learning models in fine-grained classification tasks.