#hypergradient

Question 1: Single choice

You want to tune the learning rate by gradient descent on a validation metric measured after one epoch of training. Why should the metric not be the validation error, the fraction of misclassified examples?

It is always higher than the training error, so its gradient points in the wrong direction1
It can only be measured after training, while gradients must be computed during training2
It does not depend on the learning rate at all, because the architecture alone fixes the error3
Its gradient in the learning rate is zero almost everywhere; use the validation cross-entropy4

Timed test