Noise is the fatal poison: A Noise-aware Network for noisy dataset classification

Abstract

Recently, many benchmark datasets have been found to contain noisy labels caused by unavoidable human mistakes. Many researchers propose new noise-aware loss functions to achieve robust classification performance. However, we discover that existing noise-aware loss functions cannot fully heal the damage caused by the noise. On the other hand, some methods filter out low confidence samples and train new models, whereas the filtered samples contain both noisy and hard samples that are critical for the robustness of models. Based on the above two discoveries, we devised the Noise-aware Network (NA-Net) for robust training with noisy labels. Each layer of NA-Net contains three groups of convolution kernels responsible for mix samples, clean samples, and noisy samples, termed as mix-kernels, clean-kernels, and noise-kernels, respectively. Mix-kernels are used for finding the clean samples with a newly devised noise-immune (NI) loss function; clean-kernels are targeted at learning better features without being misguided by noise; noise-kernels are trained by the remaining samples to rectify wrong labels for the next iteration. Meanwhile, for increasing the classification performance of mix-kernels, the extracted feature maps of clean-kernels without being poisoned are combined as the input of mix-kernels of the next layer. Also, the knowledge distillation strategy is adopted to distill the knowledge from clean-kernels to the noise-kernels. Extensive experiments demonstrate that the mutual promotion of three groups of kernels in NA-Net achieves state-of-the-art performance on both artificial noisy datasets and real-world datasets.