ML Katas

Write a function to apply label smoothing for classification:

• Replace one-hot targets with $$1-ϵ$$ for true class, $$ϵ/(K-1)$$ for others.
• Use it in cross-entropy training.

Show improved generalization on toy data.