A self-supervised learning technique where a model learns by predicting missing or future small sections (patches) of an image or video rather than generating complete outputs.