#warmup

Question 1: Single choice

What problem does a learning-rate warmup (a short phase in which $\eta$ rises from a small value to its maximum) address?

Early updates from random weights are large and poorly aimed, so a full-size $\eta$ at the start can diverge1
Without warmup the final learning rate stays too large for the model to converge2
Warmup lets the optimizer skip the gradient computation during the first epochs3
Warmup reduces overfitting by showing the model the easiest examples first4