Concept标签平滑

label smoothing

A regularization technique that replaces hard labels with soft labels (for example, replacing 1 with 0.9) to prevent a model from becoming overconfident. It is commonly used when training on small datasets to reduce overfitting.

Timeline (last 90 days)

Oct 23

On small datasets, adding dropout (0.1–0.3) and label smoothing (0.1) can prevent overfitting.

Unverified50%

All Facts (1)

Unverified

On small datasets, adding dropout (0.1–0.3) and label smoothing (0.1) can prevent overfitting.

50%

Source Articles

Implementing the Transformer Architecture from Scratch: Training Acceleration and Performance Improvement in PracticeOct 23