Weak supervision is a training approach that uses imperfect, indirect, noisy, or limited labels instead of fully curated ground-truth labels. It is closely related to semi-supervised learning and can reduce the cost of creating labeled datasets.
See semi-supervised learning.