#linear-probe

Question 1: Single choice

You adapt an ImageNet-pretrained ResNet to a new image task in two ways: fine-tuning all layers, or freezing the network and training only a new final linear layer (a linear readout). When should you expect the two to perform about equally well?

When the new images are close to the pretraining domain, so the frozen features already suit the task1
When the new dataset is very large, since the frozen layers then adapt to it on their own2
When the new images come from a very different domain, such as medical scans3
Never: training only the last layer is always clearly worse than full fine-tuning4

Timed test