Training
Regularization
Regularization changes training to discourage brittle or overly complex solutions.
How it works
Penalties, noise, data augmentation, and early stopping impose different preferences. The goal is better performance on unseen data, not merely higher training accuracy.
A concrete example
Illustrative: L2 coefficient 0.01 adds 0.01 times the squared-weight sum to loss.
Do not confuse it with Normalization
regularization targets generalization; normalization controls activation scale.