Label Smoothing
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A regularization technique that replaces hard 0/1 training targets with softened probabilities so a classifier cannot become fully confident in one class.
What is Label Smoothing?
Label smoothing is a regularization technique for classification that softens the training targets: instead of telling the model the correct class has probability 1 and every other class 0, it assigns the correct class slightly less than 1 and spreads the remainder across the rest.
The point is to curb overconfidence. A model trained on hard targets learns to push its predicted probabilities to extremes, which hurts generalization and makes its confidence scores untrustworthy.
How Label Smoothing Works
A smoothing factor sets how much probability mass to redistribute. With a factor of 0.1 in a ten-class problem, the correct class gets a target of 0.9 and each wrong class gets roughly 0.01, instead of 1 and 0.
The model still trains with an ordinary cross-entropy loss; only the targets change. Because a prediction of exactly 1.0 now overshoots the target, the loss penalizes extreme confidence, and the model settles on calibrated probabilities rather than certainties. That also blunts the effect of mislabeled examples, since no single label is treated as absolute truth.
Like dropout, label smoothing trades a little training-set accuracy for better behavior on data the model has not seen. The two are used together in many modern image and translation models.
Example of Label Smoothing
A sentiment classifier sorts reviews into three classes: positive, neutral, negative. A positive review carries the hard target [1, 0, 0]. With a smoothing factor of 0.1, that target becomes [0.9, 0.05, 0.05].
Trained on hard targets, the model learns to output near-certainties like [0.999, 0.0005, 0.0005] – even for a sarcastic review it should be unsure about. Trained on smoothed targets, the same review earns something closer to [0.85, 0.10, 0.05].
The predicted class is unchanged; the confidence attached to it is now honest. Downstream systems that act on that confidence, such as routing low-certainty reviews to a human, get a signal they can trust.
Related AI terms: Regularization · Dropout · Overfitting · Cross-Entropy Loss · Model Training
Did you like the Label Smoothing gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About