Unsupervised Confidence Approximation: Trustworthy Learning from Noisy Labelled Data

Navid Rabbani, Adrien Bartoli; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 4609-4617

Training neural networks with noisy labels presents a challenge due to inherent errors in label annotations. Concurrently, selectively predicting outputs from neural networks involves identifying confidently predicted results. These challenges are particularly important in the medical domain, as they often occur jointly. Existing techniques address either the training of models with noisy labels or the task of selective prediction in isolation, often neglecting their intrinsic interdependence. We establish a relationship between these challenges and propose a novel framework called Unsupervised Confidence Approximation (UCA) to address them simultaneously. UCA facilitates the concurrent training of neural networks for a main task such as image segmentation and classification while also predicting confidence levels. This is all done while accommodating datasets containing noisy labels. Remarkably, UCA operates autonomously, eliminating the need for labelled confidence information and qualifying as an unsupervised solution. Furthermore, UCA is versatile, integrating with diverse network architectures. Our evaluation of UCA's efficacy covers the general CIFAR-10N dataset as well as the medical image datasets CheXpert and Gleason-2019. In our experiments, incorporating UCA into existing networks enhances performance in both aspects of noisy label training and selective prediction. Moreover, networks equipped with UCA demonstrate comparable performance to state-of-the-art methods for noisy label training when operating in the conventional full coverage mode. By design, these UCA-equipped networks incorporate a risk-management mechanism, as evidenced by flawless risk-coverage curves. Additionally, UCA-equipped networks outperform existing selective prediction techniques, leading to substantial performance improvements and reinforcing its utility and impact within the context of trustworthy medical deep learning.