Label Noise#
Errors in training labels that can mislead a model.
Important
✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.
What it is#
Label noise is incorrect or unreliable labels in a dataset — training examples assigned the wrong class or target, arising from human labellers, weak heuristics or imperfect sensors (a tweet “I love this” marked negative, or a misrecorded diagnosis).
The three types#
Random (uniform) noise: labels flipped independently of the features (say 10% at random). Class-conditional noise: mislabeling depends on the class — “cat” is confused for “dog” more often than for “car”. Feature-dependent (systematic) noise: ambiguous or low-quality inputs are mislabeled more — blurry dog photos read as cats.
Why it hurts#
It degrades training (models learn wrong patterns), miscalibrates probabilities, and distorts evaluation (a noisy test set makes metrics meaningless). Tell-tale signs: training accuracy never reaching 100%, a stalling loss, memorisation of noise with low validation performance, and high disagreement among annotators.
Coping#
Clean the data (multiple annotators, keep high-agreement labels); use noise-robust losses (MAE, generalised cross-entropy) with regularisation and early stopping; model the noise with a transition matrix; combine a small clean set with a large noisy one (weak / semi-supervised); and always keep a clean gold-standard evaluation set with robust metrics like AUC.
Theme: Model Training & Optimization · All terminology
Hint
Mind map — connected ideas
Dataset Shift · Evaluation Set · Weak Supervision · Full Annotation · Discriminatory Power · Multiclass AUROC
Hint
More in Model Training & Optimization
Active Learning · Binary Cross-Entropy (BCE) · Deep Ensembles · Early Stopping · Ensemble · Epochs · FLOPs · Full Annotation · Hyperparameter · Log-Odds · Logit Space · Logits · Loss Functions · Model Distillation (Knowledge Distillation)
See also
Source article Adapted (context, re-expressed) in our own words from: Label Noise (insightful-data-lab.com).