Self-Supervised Learning in Computer Vision

How to train models with only a few labeled examples

A large portion of the value delivered by AI so far comes from supervised models trained on increasingly large datasets. Many of these datasets have been labeled by humans, a job that is mundane, time-consuming, error-prone, and sometimes expensive. Self-supervised learning (SSL) is a different learning paradigm allowing machines to learn from unlabeled data. In this article, we will discuss how SSL works and how to apply it to computer vision. We will compare simple approaches to state-of-the-art and see SSL in action for medical diagnosis, a domain that can benefit a lot from it but at the same time requires a deep understanding of the approach to be able to implement it correctly.

What is self-supervised learning?

According to Yann LeCun, Chief AI Scientist at Meta, self-supervised learning is “one of the most promising ways to build background knowledge and approximate a form of common sense in AI systems”. The idea behind the self-supervised approach is to train models on data without annotations.

Self-supervised learning is one of the most promising ways to build background knowledge and approximate a form of…