Label Smoothing

Label smoothing is an effective technique used in classification training to prevent the model from becoming overly confident in the target label. Instead of using hard 0 and 1 targets, slightly softened probabilities are used. This can reduce the tendency to learn overly sharp decision boundaries and may improve both generalization and calibration. It is commonly used in large classification problems and modern deep learning architectures. Although it appears to be a small adjustment, it can meaningfully affect output behavior and confidence quality.