Deep Learning is Robust to Massive Label Noise

David Rolnick, Andreas Veit, Serge Belongie, Nir Shavit

Abstract: Deep neural networks trained on large supervised datasets have led to impressive results in recent years. However, since well-annotated datasets can be prohibitively expensive and time-consuming to collect, recent work has explored the use of larger but noisy datasets that can be more easily obtained. In this paper, we investigate the behavior of deep neural networks on training sets with massively noisy labels. We show on multiple datasets such as MINST, CIFAR-10 and ImageNet that successful learning is possible even with an essentially arbitrary amount of noise. For example, on MNIST we find that accuracy of above 90 percent is still attainable even when the dataset has been diluted with 100 noisy examples for each clean example. Such behavior holds across multiple patterns of label noise, even when noisy labels are biased towards confusing classes. Further, we show how the required dataset size for successful training increases with higher label noise. Finally, we present simple actionable techniques for improving learning in the regime of high label noise.

TL;DR: We show that deep neural networks are able to learn from data that has been diluted by an arbitrary amount of noise.

Keywords: label noise, weakly supervised learning, robustness of neural networks, deep learning, large datasets

Data: [CIFAR-10](https://paperswithcode.com/dataset/cifar-10), [CIFAR-100](https://paperswithcode.com/dataset/cifar-100), [ImageNet](https://paperswithcode.com/dataset/imagenet)