. What is the key idea behind transfer learning, as motivated in this lesson?
Transfer learning sidesteps data-starved target tasks by reusing generic low-level features learned on a data-rich source task, rather than learning everything from scratch.
. What is the key difference between the 'frozen backbone' and 'fine-tuning' strategies in this lesson?
Frozen backbone only trains the new head; fine-tuning also updates the backbone, but carefully, with a small backbone learning rate to avoid wrecking what it already learned.
. Why does fine-tuning use a backbone learning rate roughly 30x smaller than the classifier head's learning rate?
A much smaller backbone learning rate protects the pretrained features from being wiped out by noisy updates from a tiny target dataset.
. Why is the comparison between the toy frozen backbone and the real ImageNet-pretrained ResNet-18 not a perfectly controlled experiment, according to the lesson's caveat?
Since ImageNet already contains cat and vehicle photos, some of ResNet-18's advantage is direct class overlap, not purely richer general-purpose pretraining.