Label Smoothing
Definition
Label smoothing replaces hard one-hot target values with slightly smoothed target probabilities.
Instead of assigning probability 1 to the correct class and 0 to all others, a small portion is distributed. This can reduce overconfidence but is not beneficial for every task.
Simply explained
Like giving a solution a little less absolute certainty.
Why it matters in practice
Relevant for classification and calibration of deep networks.
Example from business
An image classifier is trained with and without label smoothing and compared on accuracy and calibration.
Limits and pitfalls
- Adopt smoothing value without verification; ignore interactions in distillation or label noise; equate better loss values with better calibration.
Origin & history
Szegedy et al. used label smoothing in 2015 in Inception research as regularization.
Use in mid-sized companies
- Possible application in SMEs: Compare as an experimental training option for overconfident classifiers.
Frequently asked questions
Does label smoothing make labels uncertain?
It artificially changes the training targets. Is it always better? No.
Sources
- Rethinking the Inception Architecture · Szegedy et al. (2015), Rethinking the Inception Architecture
Sources checked on 10.10.2026.
Editorially reviewed on: