Regularization

Preventing overfitting

This unit is about one big problem: a model looking brilliant on training data but falling apart on new data. We'll start with the diagnosis (bias vs. variance), then move through the main fixes: weight penalties, dropout, and honest validation.

8 lessons · 90 min total · Prereq: Framing the Problem , Gradient Descent , Backpropagation

Start unit

Lessons

1

The bias-variance tradeoff

12 min

2

L2 regularization (Ridge)

12 min

3

L1 regularization (Lasso)

Dropout

Train/val/test splits

Data augmentation

Label smoothing

Ensemble methods