Self-Learning from Pairwise Credal Labels

Overview & Synopsis

Self-learning is a popular machine learning approach to compensate for the lack of labeled data, in instances associated with “pseudo-labels” produced by the trained model are used. Such an approach can degrade the solution when the provided pseudo-labels are of a scarce quality. A solution is to take into account the labeling uncertainty. We advocate the use of rich uncertainty models, which make it possible to better account for the lack of information or imprecision attached with pseudo labels. We address the specific case where the model is obtained via a one-versus-rest decomposition of the set of classes.