Trustworthy Machine Learning

Abstract

Representation learning has driven the field to develop pretrained models that generalize and transfer to new datasets. With the rising demand of reliable machine learning and uncertainty quantification, we seek pretrained models that output both an embedding and an uncertainty estimate, even on unseen datasets. To guide the development of such models, we propose the uncertainty-aware representation learning (URL) benchmark. It measures whether the uncertainty predicted by a model reliably reveals the uncertainty of its embedding. URL takes only four lines of code to implement but still has an information-theoretical backbone and correlates with human-perceived uncertainties. We apply URL to study ten large-scale uncertainty quantifiers that were pretrained on ImageNet and transfered to eight downstream datasets. We find that transferable uncertainty quantification is an unsolved open problem, but that it appears to be not at stakes with classical representation learning.