What is semi-supervised learning?

Semi-supervised learning is a type of machine learning that trains a model using a combination of labeled data and unlabeled data.

It is useful when collecting raw data is easy, but manually labeling every example is expensive, slow, or difficult.

Simple example

Suppose you have 100,000 images, but only 5,000 of them have labels such as:

cat
dog
car
tree

A semi-supervised approach can use the 5,000 labeled images together with the much larger set of unlabeled images to improve the model.

Why use unlabeled data?

Labeled examples tell the model what the correct outputs should be, while unlabeled examples can help the model learn the structure or distribution of the input data.

A simplified view is:

Small labeled dataset + Large unlabeled dataset → Better trained model

How it differs from related approaches

• Supervised learning mainly learns from labeled data.
• Unsupervised learning works without target labels.
• Semi-supervised learning combines both labeled and unlabeled data.

Common semi-supervised techniques include methods such as pseudo-labeling and consistency-based training, but those techniques are separate implementation topics.

In short, semi-supervised learning uses a small amount of labeled data together with a larger amount of unlabeled data to learn effectively when complete labeling is impractical.