Unsupervised learning
In short
Unsupervised learning is training an AI model on data that has no labels saying what the right answer is. The model looks for patterns in the data, for example by grouping similar items together.
- Also called
- Unsupervised machine learning
- Topic
- How models are trained
- Level
- Beginner
Unsupervised learning is machine learning from data that nobody has marked with a correct answer. NIST defines it as learning from patterns in unlabelled data, such as learning a function that clusters or groups data points1. Sutton and Barto describe the same idea as finding structure hidden in collections of unlabelled data2.
It is not supervised learning. In that setup an external supervisor supplies the correct category or action for each example, and the model is trained to predict it2. Unsupervised learning has no such answer key, so there is nothing to score the model against in the same direct way1.
It is also not reinforcement learning. A reinforcement learner tries actions in order to increase a reward. It does not have labels either, but its job is to maximise that reward, not to uncover structure in a fixed pile of data. Sutton and Barto treat the two as separate paradigms2.
The boundary with self-supervised learning is disputed. Google's 2018 BERT paper said it pre-trained on unlabelled text using two 'unsupervised tasks', one of which hid words and asked the model to fill them in3. Yann LeCun and Ishan Misra later argued that this kind of training is self-supervised, and that 'unsupervised' is a misleading name because it suggests there is no teaching signal at all4. NIST's 2025 glossary keeps unsupervised learning for tasks such as clustering, and gives the hide-and-predict setup its own entry as self-supervised learning1.
Example
A shop might group customers by what they buy, without anyone first deciding the names of the groups. Unsupervised learning is that search for groups. A person still has to decide whether the groups are useful.
Why it matters
It is the setup NIST describes for grouping data when there is no label to predict1. The groups still need a person to decide whether they are useful.
Questions people ask
Is unsupervised learning the same as self-supervised learning?
Not in current official usage. NIST keeps unsupervised learning for finding patterns such as clusters in unlabelled data, and defines self-supervised learning separately as predicting a hidden part of the data1. Some earlier papers, including BERT, called that hide-and-predict setup unsupervised3. LeCun and Misra argue the older name is misleading4.
Does unsupervised learning need labelled data?
No. NIST defines it as learning from patterns in unlabelled data, for example by grouping similar points1. If each example already has a correct answer attached, the setup is supervised learning instead2.
Is unsupervised learning the same as reinforcement learning?
No. Neither uses a list of correct actions, but they do different jobs. Unsupervised learning looks for structure in a collection of data. Reinforcement learning tries actions so as to increase a reward2.
Related terms
Sources
- NIST AI 100-2e2025 (March 2025), Adversarial Machine Learning, glossary entries for unsupervised learning and self-supervised learning
- Sutton and Barto (2018), Reinforcement Learning: An Introduction, 2nd edition, MIT Press, section 1.1
- Devlin et al. (2018), BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, arXiv:1810.04805
- LeCun and Misra (4 March 2021), Self-supervised learning: The dark matter of intelligence, Meta AI