Maximizing Coding Rate Reduction for Robust Classification of Long-Tailed Data with Noisy Labels
Zhen Yuan, Jinfeng Liu · 2023
In numerous extensive classification scenarios, the prevalence of long-tailed data and the existence of noisy labels are customary rather than anomalous occurrences. However, most existing deep learning approaches assume that the data is balanced and clean, or consider only one of these cases. Consequently, proposed solutions often underperform when faced with long-tailed data and noisy labels. To tackle this challenge, we suggest employing the Maximal Coding Rate Reduction (MCR2) principle for learning from datasets characterized by long-tailed distributions and label noise. We utilize the information-theoretic measure, MCR2, to maximize coding rate differences between the total data and the cumulative sum of each category, addressing challenges in long-tailed data with label noise. Sub-sequently, a modified version of Double-Weighted Least Squares Twin Bounded Support Vector Machines (DWLSTBSVMs) is employed for image classification. Evaluating the effectiveness of our proposed method (MCR2-DSVMs), we conduct experiments on CIFAR-10 and CIFAR-100 datasets, augmented with synthetic imbalance and noise. Results highlight the superior robustness of our model compared to alternative methods, showcasing its competitive performance.