Weak Supervision
Label a mountain of data cheaply with rough rules instead of paying humans to do it all by hand.
Think of it like
Sorting mail by zip-code stamps instead of reading every envelope — fast, mostly right, occasionally wrong.
Example
To flag toxic comments, you write a few keyword rules and treat their output as labels, then train on millions of examples.
How it actually works
Instead of one careful label per example, you combine many noisy signals — heuristics, keyword rules, existing models — and let their agreement stand in for truth. You get far more labeled data than humans could ever produce, at the cost of noise you have to model or tolerate. It shines when data is abundant but human labeling is the bottleneck.
For product teams
A way to bootstrap a labeled dataset in days instead of months when perfect labels aren’t worth the wait.
For engineers
Programmatic labeling from multiple noisy sources whose outputs are aggregated into training signals.
Related
- Gold Labels — The cheap alternative to hand-made gold labels.
- Labeling — Trades quality for scale.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome