Label noise
Also known as: noisy labels
Label noise is when some training examples are mislabeled or inconsistent. The model does its job and faithfully learns the errors — so bad labels in, bad behavior out.
It’s a common, underappreciated cause of odd model behavior.
- Human labeling is imperfect at scale.
- Some “correct” labels are genuinely wrong.
- The model can’t tell a mistake from a pattern.
The cure is curation, not capacity: more parameters won't fix bad labels. Auditing and cleaning the data is what moves the needle.