A Unified Approach to Learning with Label Noise and Unsupervised Confidence Approximation

Noisy label training is the problem of training a neural network from a dataset with errors in the labels. Selective prediction is the problem of selecting only the predictions of a neural network which have sufficient confidence. These problems are both important in medical deep learning, where they commonly occur simultaneously. Existing methods however tackle one problem but not both. We show that they are interdependent and propose the first integrated framework to tackle them both, which we call Unsupervised Confidence Approximation (UCA). UCA trains a neural network simultaneously for its main task (e.g. image segmentation) and for confidence prediction, from noisy label datasets. UCA does not require confidence labels and is thus unsupervised in this respect. UCA is generic as it can be used with any neural architecture. We evaluated its performance on the CIFAR-10N and Gleason-2019 datasets. UCA’s prediction accuracy increases with the required level of confidence. UCA-equipped networks are on par with the state-of-the-art in noisy label training when used in regular, full coverage mode. However, they have a risk-management facility, showing flawless risk-coverage curves with substantial performance gain over existing selective prediction methods.