Weak supervision refers to training machine learning models using imperfect, noisy, or indirect labels instead of relying solely on hand-labeled ground truth. This approach helps scale supervised learning when labeled data is scarce, expensive, or time-consuming to obtain.

Sources of weak supervision include heuristic rules, distant supervision (e.g., using a knowledge base to label text), user interactions, label propagation, or outputs from other models. These weak labels may be noisy individually, but when combined intelligently—using methods like label models or confidence weighting—they can approximate high-quality supervision.

Frameworks like Snorkel and weak supervision pipelines in NLP or computer vision leverage this strategy to bootstrap models for tasks like classification, information extraction, or object detection. It’s especially useful in domains where expert annotation is slow or costly, such as medical imaging or legal text processing.

Weak supervision trades off label accuracy for scale and speed, often requiring robust model architectures and post-hoc validation to ensure generalization.