pseudo-labeling

Semi-supervised learning exists to solve a specific, very common headache, that labeling data is expensive, so you want to train mostly on a small labeled set plus a much larger pool of unlabeled data. The standard playbook, popularized by a…