Self-Supervised Learning
Self-Supervised Learning
A machine learning method that creates its own questions from unlabeled data and learns by answering them.
In Simple Terms
Self-supervised learning is a way for a computer to get smarter by hiding part of its data and solving a fill-in-the-blank puzzle from what's left. It practices over and over, like covering a word in the middle of a sentence and guessing it from the surrounding context, or cutting a piece out of a photo and trying to restore it. Since humans don't have to attach correct labels one by one, it can learn features from huge amounts of data, which is why it's widely used to build the foundation of today's giant AI models.
Behind the Name
It's called "self-supervised" because, unlike "supervised learning," where humans supply the correct answers, the data itself plays the role of the "supervisor." Nothing is handed over from the outside. Instead, part of the data is hidden and the computer tries to guess it, a bit like studying on your own with a quiz you made yourself.
Take a Closer Look!
Self-supervised learning is a machine learning method where the computer creates both the questions and the correct answers from the data itself, so no human has to supply the answers.
It's an approach in which the computer finds features and patterns on its own, starting from data that has no labels prepared in advance.
In ordinary supervised learning, humans have to prepare the correct answers by hand, such as labeling each image as "dog" or "cat."
Self-supervised learning, on the other hand, uses the collected data itself to drive the learning.
For example, it might deliberately hide a single character in a sentence and make the computer guess it from the characters around it, or rotate an image and make it figure out the original orientation.
Because a huge number of questions can be generated from the data alone, you can train on massive amounts of text and images without the hassle of labeling them by hand.
A model that has picked up a broad understanding of the structure and features of data this way can then reuse what it learned in further training aimed at a specific job.
It's widely used as the basic training for large language models, which read enormous amounts of text to understand how words connect, and for image recognition models.