Label noise

Label noise is the presence of incorrect, inconsistent, or imprecise labels in a dataset, the errors mixed into the ground truth. It includes wrong classes, sloppy boundaries, and mislabeled examples, and it degrades both how a model trains and how accurately it can be evaluated.

What is label noise?

Label noise is the slice of a dataset's labels that are simply wrong: a cat tagged as a dog, a box that misses half the object, a class applied inconsistently across annotators. Some noise is random, the occasional slip, and some is systematic, a misunderstood guideline repeated thousands of times.
Models trained on noisy labels learn the errors as if they were truth, and, just as damaging, noise in a test set makes evaluation unreliable, because a correct prediction can be scored wrong when the label it is compared against is itself the error.

Key takeaways

  • Label noise is the error in your labels, the wrong, imprecise, or inconsistent annotations.
  • It hurts twice: it misleads training and it corrupts evaluation.
  • Systematic noise, from bad guidelines, is far more harmful than random noise.

What label noise looks like

The common forms label noise takes in a dataset.
The common forms label noise takes in a dataset.
TypeWhat it looks like
Class noiseThe wrong category assigned
Boundary or localization noiseA box or mask that is off, loose, or imprecise
InconsistencyThe same case labeled differently by different annotators or over time
Missing labelsObjects that should be annotated but are not, teaching the model to ignore them

How it works

Label noise is found by surfacing suspicious labels, not by reading the whole dataset. In FiftyOne you can rank samples by model mistakenness, compare predictions to ground truth, and sort by high loss or low agreement to pull the likely-wrong labels to the top for review.

Why it matters

Label noise sets a hard, invisible ceiling on a project, so you can spend weeks debugging a model that is actually performing correctly on broken labels. Random noise mostly costs you sample efficiency, because the model still learns the pattern, just slower, but systematic noise changes what the model learns, encoding the mistake as a rule. That is why a small amount of consistent, biased mislabeling can do more damage than a large amount of random error, and why finding the systematic source, one annotator or one ambiguous class, beats trying to scrub every individual slip. Even curated benchmarks are not exempt: measured label-error rates average a few percent and reach an estimated 6% in the ImageNet validation set (Northcutt et al., 2021).

Frequently asked questions

What causes label noise?

Human error, ambiguous guidelines, rushed labeling, and unreliable auto-labels.

How do you detect label noise?

By surfacing labels the model strongly disagrees with, low-agreement samples, and high-loss examples for review.

What is the difference between label noise and annotation quality?

Annotation quality is the overall picture. Label noise is specifically the share of labels that are wrong.

Related terms

Last updated July 9, 2026

Building visual or physical AI?

Let's talk.