weak supervision
/ WEEK soo-per-VIZH-un /
Hand-labeling a million examples one by one is slow and expensive. But often you can write quick, rough rules that get most of them right: "if an email contains the phrase 'wire transfer urgent,' it's probably a scam." The rule is imperfect, but cheap, and you can apply it to everything instantly. Weak supervision is the practice of training models from labels that are cheap and noisy rather than scarce and perfect.
Weak labels come in several flavors: rules and heuristics ("contains the word 'invoice' → finance"), existing databases used as a stand-in for ground truth, labels that are imprecise (we know a photo contains a cat somewhere, but not where), or labels from non-experts or other models. Each source is individually unreliable. The key idea is to combine many weak, overlapping, disagreeing sources statistically — modeling how much to trust each and how they correlate — to produce probabilistic "best-guess" labels good enough to train a real model on.
It is a backbone of practical, data-centric machine learning, because it lets a small team label enormous datasets in hours instead of months, and lets domain experts inject knowledge as rules rather than as endless manual tagging. The candid limits: the labels really are noisy, so blind spots and systematic errors in your rules get baked into the model. Weak supervision works best when many independent sources can correct each other; when they all share the same bias, the model inherits it wholesale.
To build a spam filter without hand-labeling, a team writes a dozen rough rules — links to known bad domains, all-caps subject lines, certain phrases. Each rule is wrong sometimes, and they disagree. A weak-supervision model weighs and merges them into probabilistic labels for a million emails, and a classifier trained on those rivals one built from costly hand labels.
Many cheap, imperfect rules combine into usable labels.
Weak supervision trades label quality for label quantity. It shines when many independent sources correct each other — but if they all share the same blind spot, the model inherits that error wholesale.