Hints

MATH-C09-C-001

Compute the prediction error first.

Use w - learning_rate * gradient.

Divide dataset size by batch size.

Add the two objective terms.

Multiply learning rate by gradient.

Training loss and reported metrics can be different numbers.

This is the main clean guarantee of convexity.

Compute 8 + 0.25 * 4.

Momentum carries recent movement.

A symptom tells you something went wrong, not necessarily why.

Ask whether one training number can capture every part of useful behavior.

The gradient gives local direction. The learning rate decides how far to move.

Averaging more examples usually reduces noise, but those examples must be processed.

The clean local-minimum guarantee is a convexity guarantee.

Regularization changes the objective. Validation checks what happened outside the fitting data.