A way to train AI using cheap labels with some mistakes.
You teach an AI with the messy names people gave their photo files. Many names are wrong, but enough of them still point to what is in the picture.
It uses cheap, messy clues instead of perfect labels. You see it with huge sets of text or pictures.
Supervised Learning
Weak Supervision needs less exact labels. This lowers labeling costs.
Data Labeling
Weak labels can come from rules. Old data and crowds can help too.
Semi-Supervised Learning
Both methods save labeling work. They use data in different ways.