May the Forgetting Be with You: Alternate Replay for Learning with Noisy Labels

• Proceedings of the 35th British Machine Vision Conference

• November 2024

• BMVC

• Continual Learning Noisy Label Learning Rehearsal Sample Selection

This study investigates the impact of annotation errors on the performance of rehearsal-based methods, showing that current approaches are limited to the single-epoch scenario. Indeed, as the model is allowed to fit the data from each task, the buffer becomes poisoned by the mislabelled data. We thus propose AER and ABS, two methods that leverage forgetting to mitigate the impact of annotation errors.