Label Smoothing is a regularization technique that softens hard one-hot labels by distributing a small amount of probability to non-target classes — instead of training with labels $[0, 0, 1, 0]$, use $[epsilon/K, epsilon/K, 1-epsilon, epsilon/K]$, preventing the model from becoming overconfident.

Label Smoothing Formulation

• Smoothed Label: $y_s = (1 - epsilon) cdot y_{one-hot} + epsilon / K$ where $K$ is the number of classes.
• $epsilon$ Parameter: Typically 0.05-0.1 — small enough to preserve the correct class, large enough to regularize.
• Effect: The model learns to predict ~90% for the correct class instead of trying to reach 100%.
• Calibration: Label smoothing improves model calibration — predicted probabilities better reflect true confidence.

Why It Matters

• Overconfidence: Without smoothing, models become extremely overconfident — label smoothing prevents this.
• Generalization: Acts as a regularizer — improves generalization by preventing the model from fitting hard labels exactly.
• Standard Practice: Used in most modern image classification (ResNet, EfficientNet, ViT) and NLP (BERT, GPT).

Label Smoothing is humble predictions — preventing overconfidence by teaching the model that no class should be predicted with 100% certainty.