#overparameterization

Question 1: Single choice

In the "double descent" pattern reported for deep networks, how does test error change as model complexity (for example, network width) keeps growing?

It falls, peaks when the model can just fit the training data, then falls again1
It falls and then rises steadily, the classical U-shaped curve, with no second drop2
It decreases monotonically, because each parameter added lowers the test error a little3
It stays flat, while it is the training error that goes down twice instead4

#overparameterization

Timed test