1) Definition
- Label noise = incorrect or unreliable labels in your dataset.
- In supervised learning, it means some training examples are assigned the wrong class or target value.
- Happens because labeling is done by humans, weak heuristics, or imperfect sensors.
Example:
- Sentiment dataset: a tweet “I love this” accidentally labeled as negative.
- Medical dataset: diagnosis labeled incorrectly due to human error.
2) Types of Label Noise
- Random (Uniform) Noise
- Labels are flipped randomly, independent of features.
- Example: 10% of labels are just assigned at random.
- Class-conditional Noise
- Mislabeling depends on the class.
- Example: “cat” misclassified as “dog” more often than as “car”.
- Feature-dependent Noise (systematic)
- Harder cases (ambiguous or low-quality inputs) are mislabeled more frequently.
- Example: Blurry dog photos often mislabeled as “cat.”
3) Why it’s a problem
- Training degradation: models learn from wrong examples → worse accuracy.
- Calibration issues: probability estimates become unreliable.
- Evaluation distortion: test set errors reduce metric validity.
4) Symptoms of Label Noise
- Training accuracy never approaches 100%.
- Loss stops decreasing properly.
- Model memorizes noise (overfits) while validation performance stays low.
- High disagreement among annotators.
5) Coping Strategies
- Data cleaning
- Manual review or crowdsourcing multiple annotators.
- Rule: keep examples only if label agreement is high.
- Robust algorithms
- Use loss functions tolerant to noise: MAE, generalized cross-entropy.
- Regularization + early stopping prevent overfitting noisy labels.
- Noise modeling
- Explicitly estimate a noise transition matrix (probability that true label → observed noisy label).
- Train model to account for this.
- Weak supervision / semi-supervised learning
- Combine small high-quality labeled set with large noisy set.
- Evaluation safeguards
- Keep a clean evaluation set (gold standard).
- Use robust metrics (AUC instead of raw accuracy).
6) Example in Practice
Suppose a dataset of 1,000 images:
- 900 labeled correctly
- 100 mislabeled (dog as cat, etc.)
- If you train a classifier, it might memorize the 100 wrong examples, hurting generalization.
- If you apply noise-robust training or manually relabel, test accuracy improves.
Summary
- Label noise = mislabeled training/test data.
- Types: random, class-conditional, feature-dependent.
- Consequences: lower performance, overfitting, misleading evaluation.
- Fix: clean data, robust losses, noise-aware modeling, semi-supervised methods.
