What Is Semi-Supervised Learning? A Plain-Language Explanation
Semi-supervised learning is a machine learning approach that combines a small amount of labeled data with a much larger amount of unlabeled data during training, sitting between supervised learning and unsupervised learning and helping reduce the cost of labeling.
The simple analogy
Imagine a teacher who fully works through a few example problems with a student, then hands over a large set of unsolved problems. The student uses the worked examples to make sense of the rest. Semi-supervised learning uses a few labels to guide learning from many unlabeled examples.
Where it fits
| Approach | Labeled data | Unlabeled data |
|---|---|---|
| Supervised | All examples labeled | None required |
| Semi-supervised | Small set | Large set |
| Unsupervised | None | All data |
How it works
- Train an initial model on the small labeled set.
- Use it to make tentative predictions on unlabeled data.
- Treat confident predictions as pseudo-labels.
- Retrain combining labeled and pseudo-labeled data.
Why it is used
Manual labeling can be slow and costly, especially when experts are needed, such as medical images. Unlabeled data is often abundant and cheap, so semi-supervised learning aims to get more value from a few labels.
Honest caveats
Pseudo-labels can be wrong, and incorrect ones may reinforce errors. Results depend on the unlabeled data reflecting a similar distribution to the labeled data.
FAQ
What is a pseudo-label?
A confident prediction on unlabeled data that is treated as if it were a label.
How is it different from self-supervised learning?
Semi-supervised learning still uses some real labels, while self-supervised learning constructs its training targets entirely from the data.
When is it useful?
When labels are expensive but unlabeled data is plentiful.
Can pseudo-labels be wrong?
Yes, so confidence and validation are important.
Key takeaways
- Semi-supervised learning mixes a few labels with many unlabeled examples.
- It sits between supervised and unsupervised learning.
- Pseudo-labels extend the labeled set.
- Label quality and data distribution matter.
Continue with what is supervised learning and what is unsupervised learning.
Sources
Learn by doing
Read a guide, then practice with bite-sized AI lessons you can take offline.
Start learning free