A Double Regularization Loss Based on Long-tailed Noisy Labels
Lei Wang, Shao-Yuan Li · 2023
Extensive research has been conducted in recent years to solve the long-tailed distribution and achieved excellent results. However, in contrast to well-designed data, datasets with label noise are common in the real world, even in long-tailed datasets. Then, loss functions that rely on prior knowledge of correct labels for long-tailed distributions will fail. To solve the above problems, the robustness of different loss functions in long-tailed data containing noise is first analyzed. Then algorithmic improvements are made to the LDAM loss, which is focused on dealing with long-tailed problems. We propose a robust loss function DRL, which solves the noisy long-tailed problem with two regularisation terms: label regularisation and sample regularisation. Experiments on several datasets validate the effectiveness of the proposed loss function.