标签平滑

With K classes, one common label-smoothing convention assigns 1 − ε to the correct class and distributes ε across the remaining classes; implementations may instead distribute ε across all classes, so check the loss definition. This prevents the target from demanding probability one for the labeled class and can discourage extreme logits. The Inception-v3 work and the Transformer paper are prominent documented uses of the method. Research has found calibration and generalization benefits in some settings, while also showing that smoothing can remove information about similarities among incorrect classes and weaken a model as a knowledge-distillation teacher. It is therefore not a universal accuracy or calibration improvement. Choose ε and evaluate calibration, task quality, and downstream distillation on representative validation data.