#inference-cost

Question 1: Select all that apply

Which statements about the computational cost of dropout are true? Select all that apply.

Dropout usually slows down training, since the network needs more epochs to converge1
Standard dropout adds no cost at prediction time, because it is switched off in evaluation mode2
MC Dropout averaging 10 stochastic forward passes makes each prediction roughly 10 times slower3
At prediction time standard dropout must rescale every activation by the keep probability, which slows inference down4