| Abstract: | The use of all samples in the optimization process does not produce robust results in datasets with label noise. Because the gradients calculated according to the losses of the noisy samples cause the optimization process to go in the wrong direction. In this paper, we recommend using samples with loss less than a threshold determined during the optimization, instead of using all samples in the mini-batch. Our proposed method, Adaptive-k, aims to exclude label noise samples from the optimization process and make the process robust. On noisy datasets, we found that using a threshold-based approach, such as Adaptive-k, produces better results than using all samples or a fixed number of low-loss samples in the mini-batch. On the basis of our theoretical analysis and experimental results, we show that the Adaptive-k method is closest to the performance of the Oracle, in which noisy samples are entirely removed from the dataset. Adaptive-k is a simple but effective method. It does not require prior knowledge of the noise ratio of the dataset, does not require additional model training, and does not increase training time significantly. In the experiments, we also show that Adaptive-k is compatible with different optimizers such as SGD, SGDM, and Adam. The code for Adaptive-k is available at GitHub. |
