Self-supervised learning is a training approach in which a model learns from data by creating supervision signals from the data itself.
It is widely used for pretraining language, vision, and multimodal models before task-specific adaptation.