<ul><li>Learning paradigms like active learning, semi-supervised learning, bandits, or boosting self-select training data based on previously learned parameters.</li><li>Reciprocal learning unifies these paradigms, and this article focuses on the generalization ability of methods using self-selected samples.</li><li>The article presents universal generalization bounds for reciprocal learning, using covering numbers and Wasserstein ambiguity sets without assumptions on the data distribution.</li><li>Results are provided for both convergent and finite iteration solutions, offering anytime valid stopping rules for practitioners to ensure out-of-sample performance, illustrated through the semi-supervised learning case.</li></ul>

Generalization Bounds and Stopping Rules for Learning with Self-Selected Data

Discover more