Core ML Concepts · advanced · concept 30 of 176
Self-Supervised Learning
Learning from unlabeled data by inventing the labels: predict the next word, fill in the masked patch, decide whether two crops came from the same image. This is the idea that unlocked LLMs, since the internet is unlabeled and labeling it is impossible.
Key terms
Pretext taskMasked modelingNext-token predictionContrastive objectiveRepresentation learning
Learn these first
Where you meet it in the real world
LLM pretraining, image encoders like DINO, speech models like Whisper, protein models
Videos
▶ 8: Deep Learning for Natural Language – Transformers, Self-Supervised Learning ↗
MIT OpenCourseWare · YouTube
▶ Stanford CS231N | Spring 2025 | Lecture 12: Self-Supervised Learning ↗
Stanford Online · YouTube
Guides and articles
Courses, papers, and more