Improving Performance Flashcards
(9 cards)
What is early stopping and why do we use it?
Stop training when validation performance stalls to avoid overfitting.
How is early stopping applied?
Monitor validation loss; if it fails to improve for a set patience, halt and keep best weights.
What does dropout do during training?
Randomly disables some neurons so the network cannot rely on specific paths, acting as regularisation.
What changes with dropout at test time?
All neurons are active; no random dropping is applied.
What is the purpose of batch normalization?
It normalises activations in each mini‑batch to stabilise and speed up learning.
How does batch normalization handle inference?
Uses stored running averages of mean and variance instead of batch statistics.
Why do we use data augmentation?
It increases dataset variety with label‑preserving transforms, reducing overfitting.
Name three simple image augmentation transforms.
Horizontal flip, slight rotation, random crop.
What does momentum add to SGD?
A running average of past gradients that smooths updates and speeds convergence.
