← Deep Learning learning pathDeep Learning · Module 5 of 10
Overfitting and regularisation
Help models generalise to new data.
📖 Key ideas
- Overfitting: great training accuracy, poor validation accuracy.
- Use a validation set to monitor generalisation.
- Regularisation: dropout, weight decay (L2), data augmentation.
- Early stopping halts training when validation loss rises.
- More data is often the best cure.
💡 Real-life example: Overfitting is a student who memorised answers; regularisation forces them to understand concepts.
OverfittingLearning noise instead of patterns
DropoutRandomly disables neurons during training
Early stoppingStop when validation stops improving
Data augmentationCreating varied training examples