Loss function
The function that measures how wrong the model's output is — what gradient descent tries to minimise.
Classics: Mean Squared Error (regression), Cross-Entropy (classification). Choice of loss determines what the model optimises for. Wrong loss = the model learns the wrong things.