Semi-Supervised Learning
Learning from a small labelled set and a large unlabelled one — the setting almost every real project is actually in.
Reviewed July 16, 2026Stable
When not to use it
- When labelled and unlabelled data come from different distributions, where it commonly hurts.
- When a suitable pretrained model exists and fine-tuning is available, which is usually simpler and stronger.
- Under severe class imbalance, where pseudo-labelling amplifies the majority class.
Reach for something else instead
- Self-supervised pretraining followed by fine-tuning, which has absorbed most of this territory.
- Active learning, which reduces labelling cost by choosing what to label rather than by using unlabelled data.
- Weak supervision, where noisy programmatic labels replace manual ones.
Sources & further reading
- Chapelle, Schölkopf & Zien (2006), Semi-Supervised Learning — the standard reference and the assumption taxonomy.
- Oliver et al. (2018), Realistic Evaluation of Deep Semi-Supervised Learning Algorithms — shows reported gains often do not survive fair comparison.
- Sohn et al. (2020), FixMatch — consistency regularisation with pseudo-labelling.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Not verifying that the unlabelled pool matches the labelled distribution.
- Pseudo-labelling without a confidence threshold, which promotes errors into training data.
- Reporting gains against a weak supervised baseline, which is the flaw the realistic-evaluation work identified.
At a glance
FieldMachine Learning
Usessmall labelled set plus large unlabelled one
Main methodpseudo-labelling
Key riskconfident wrong labels compounding
Rests oncluster and manifold assumptions
DifficultyPractical
Where this sits
A destination. 3 concepts lead here, and nothing in the corpus depends on it.
2Levelsteps in
3Needs firstconcepts
0Opens upnothing further
1Areastays here
Computed from the prerequisite graph, not assigned. How this works