Mini toolsDefinition
A regularization technique that softens hard target labels to improve model generalization and prevent overconfidence. This involves mixing target labels with a uniform distribution.
Detailed Explanation
Label smoothing works by replacing one-hot encoded labels with slightly softened distributions typically by allocating a small probability ε to non-target classes. This prevents the model from becoming too confident in its predictions and helps it learn more robust features. The technique has been shown to improve model calibration and generalization performance.
Use Cases
Image classification tasks Speech recognition systems Language model training Multi-class classification problems
