Skip to content
← Research
ML Theory

Self-Supervised Learning

Learning useful structure from unlabeled data, where labels are expensive, sparse, or simply don't exist yet.

Pretext tasksPretrainingSparse labelsAnomaly detection

Most real-world and scientific data is unlabeled by nature. This work explores self-supervised objectives that let models learn useful structure from raw input before any labels enter the picture — relevant both to physics data and to production ML systems with limited annotation.