Self-Supervised Learning is a training approach where a model learns from unlabelled data by generating its own learning signal — for example, by hiding part of the data and learning to predict it. It is how large language models are trained.
What it means in plain English
Labelled data is scarce, but unlabelled data is everywhere. Self-supervised learning exploits this by creating its own training tasks from raw data: it hides a piece of the input and trains the model to predict the hidden part. No human labels are needed, because the data supplies its own answers. This is how models learn deep patterns from vast amounts of text or images at a scale labelling could never reach.
It is a big reason today’s largest models are so capable.
A simple example
A language model is trained by hiding words in sentences from a huge body of text and learning to predict them. The correct word is already there in the original text, so the model effectively supervises itself — no human labelling required.
Why it matters
Self-supervised learning unlocked training on internet-scale data without the bottleneck of manual labelling, and it is the key technique behind modern foundation models and large language models. It is one of the main reasons AI advanced so rapidly.
Related terms
- Pre-training — typically done via self-supervised learning.
- Large Language Model — trained this way.
- Supervised Learning — the labelled-data approach it contrasts with.
Frequently asked questions
What is self-supervised learning?
It is learning from unlabeled data by creating the labels automatically from the data itself — for example, hiding a word and training the model to predict it.
Why is self-supervised learning important?
It lets models learn from vast amounts of unlabeled text or images without costly manual labelling, and it underpins how large language models are pre-trained.
