Self-Supervised Learning - Page 6

Self-supervised learning creates training signals from the data itself instead of relying entirely on human labels. A model may predict masked words, reconstruct a missing image region, identify whether two views came from the same object, or learn the next element in a sequence. These objectives let systems pretrain on enormous collections of text, images, audio, or video and develop representations that can later support specific tasks. The approach reduces annotation needs but does not remove data-quality concerns; the model can absorb bias, private information, duplication, and harmful patterns from its source material. Downstream evaluation is necessary because success on the pretraining objective may not translate evenly across applications or groups.