semi-supervised learning
/ SEM-eye-soo-per-VIZED LER-ning /
Suppose a teacher works through ten problems on the board with full solutions, then hands you a thick stack of unsolved ones to ponder. The few worked examples anchor you, while the many blanks let you practice and sharpen your instincts. Semi-supervised learning is exactly this blend: a small pile of labeled examples plus a large pile of unlabeled ones, learned from together.
The reasoning is practical. Labeling data — a doctor marking tumors, a person transcribing speech — is slow and costly, so you usually have only a little of it. But unlabeled data is cheap and plentiful. Semi-supervised methods use the few labels to get a rough idea of the task, then lean on the structure of the unlabeled mass to firm up the answer. A common trick is "pseudo-labeling": the model labels the easy unlabeled cases itself, treats its confident guesses as if they were real labels, and keeps training.
When it works, you get accuracy close to fully-labeled training for a fraction of the labeling effort. But it leans on an assumption — that the unlabeled data really shares the structure the labels imply (similar things cluster, decision boundaries fall in sparse regions). When that assumption breaks, or when the model confidently pseudo-labels its own mistakes, semi-supervised learning can quietly drift worse rather than better, so it needs careful checking.
A team has 500 chest X-rays labeled by radiologists and 50,000 unlabeled ones. They train on the 500, let the model pseudo-label the unlabeled scans it is most sure about, fold those in, and retrain. The result lands close to what 50,000 hand-labeled scans would have given — at a tiny fraction of the doctors' time.
A few real labels guide; many unlabeled examples reinforce.
Pseudo-labeling can backfire: if the model is confidently wrong, it teaches itself its own mistakes and the errors compound. Semi-supervised learning helps only when the unlabeled data genuinely matches the labeled structure.