Regularization
Any constraint that keeps a model from fitting its training data too closely.
Overview
Regularization is the family of techniques that deliberately make fitting the training set harder: dropping units, penalising large weights, stopping early, limiting capacity. It solves the problem that a model with enough parameters will memorise rather than generalise. It exists as a trade made on purpose - worse training performance in exchange for better performance on data the model has not seen.
Technology Graph
How Regularization connects to the rest of the stack.
Applied to
- Fine-tuning technique Techniques Regularization used for Fine-tuning High ✓ source
- Pretraining technique Techniques Regularization used for Pretraining High ✓ source
- Overfitting concept Algorithms Regularization solves Overfitting High ✓ source
Tools
Software that implements or integrates with Regularization.