What is 'Label Smoothing'?

What is 'Label Smoothing'?

Approach: Describe modifying target labels to prevent the model from becoming too confident.

A regularization technique that introduces noise into the labels by replacing the 'one-hot' binary targets with a mixture of the label and a uniform distribution over all labels. This prevents the model from being overly confident.

Follow-up: How does it affect the cross-entropy loss?

Key concepts: overfitting, probabilities, regularization.