Weak supervision is an approach to labeling that generates training labels from cheap, noisy sources, heuristics, rules, knowledge bases, or multiple imperfect labelers, instead of manual annotation. Programmatic labeling is its practice: writing labeling functions that apply labels at scale, then combining and denoising them.
Weak supervision trades perfect, expensive labels for abundant, imperfect ones. Rather than have humans annotate each example, you encode domain knowledge as sources of weak labels, rules and heuristics ("if the filename contains 'xray,' it's medical"), pattern matches, existing databases, or several mediocre models, each of which labels data cheaply but noisily.
The practice of doing this in code is programmatic labeling: you write labeling functions, run them over the data, and use a model to estimate each function's reliability and combine them into a single, denoised label. You get a large labeled set fast, at the cost of noise the combination step is designed to manage.
Key takeaways
Weak supervision creates labels from cheap, noisy sources instead of manual annotation.
Programmatic labeling implements it as labeling functions, code rules applied at scale.
The combination-and-denoising step is what turns many noisy signals into usable labels.
What weak supervision provides
How weak supervision compares to auto- and pseudo-labeling.
How weak supervision compares to auto- and pseudo-labeling.
Approach
Label source
Human role
Weak supervision
Rules, heuristics, existing sources
Write the rules, not the labels
Auto-labeling
A trained or foundation model
Review the labels
Pseudo-labeling
A model's own confident predictions
Usually none
How it works
The weak labels come from human-written heuristics and rules, distant supervision (existing databases or knowledge bases), pattern and keyword matching, or several imperfect models whose votes are combined. You write these as labeling functions, run them over the data, and a model estimates each function's reliability and merges them into one denoised label.
Why it matters
Weak supervision shifts human effort from labeling individual examples to expressing knowledge once, in rules, that label millions. Its quality hinges on the diversity and independence of the weak sources, not their individual accuracy. Many noisy-but-uncorrelated signals combine into a strong label because errors cancel, but many sources that share the same blind spot just amplify the same mistake confidently, and no denoising recovers a bias every source agrees on. So the skill is engineering varied, independent labeling functions, and weak supervision works best as a first pass that a curated, human-reviewed set then validates and corrects.
Frequently asked questions
What is weak supervision?
Creating training labels from cheap, noisy sources like rules and existing data instead of manual annotation.
What is programmatic labeling?
The practice of weak supervision in code, writing labeling functions that label data at scale and combining them.
How is weak supervision different from auto-labeling?
Auto-labeling uses a model to predict labels. Weak supervision combines rules, heuristics, and other noisy sources.