Deep Learning

Regularization for Deep Networks

Regularization reduces overfitting by making memorization harder or less attractive. In deep learning, the main tools are weight decay, dropout, data augmentation, early stopping, and capacity control.
  • Weight decay discourages large weights
  • Dropout prevents brittle co-adaptation
  • Data augmentation encodes invariances
  • Early stopping uses validation behavior
  • Smaller models regularize directly
  • Regularization follows diagnosis
Regularization toolkit
TechniqueWhat it doesBest when
Weight decayPenalizes large weightsModel overfits smoothly
DropoutRandomly removes activations during trainingDense networks over-rely on paths
Data augmentationAdds label-preserving variationsInput has known invariances
Early stoppingStops before memorization dominatesValidation curve turns upward
Smaller modelReduces capacityModel is too flexible for data
Overfit signal
Regularization is a response to a train-validation gap.
Sources
  • Deep LearningRegularization for Deep Learning
  • Dive into Deep LearningRegularization
  • Machine Learning Crash CourseOverfitting