#pretrained-models

Question 1: Single choice

d2l fine-tunes an ImageNet-pretrained ResNet-18 on a small hot-dog dataset, with a learning rate of $5 \times 10^{-5}$ for the pretrained layers and 10 times that for the new output layer. What do you expect if you keep increasing the learning rate of the pretrained layers?

Accuracy keeps rising, since the pretrained layers can then adapt to hot dogs more quickly1
Accuracy stays the same, since pretrained layers have converged and receive no gradient anymore2
Accuracy stays the same, since only the new output layer's learning rate matters for the result3
Accuracy eventually drops, since large updates destroy the useful pretrained features4

Timed test