#feature-generality

Question 1: Select all that apply

Yosinski et al. measured how well the features of each layer of an ImageNet-trained CNN transfer to another task. Which findings did they report? Select all that apply.

Transfer gets worse as the base and target tasks grow apart, but distant-task features still beat random ones1
First-layer features such as Gabor-like filters and color blobs are general and transfer almost perfectly2
After fine-tuning on the target data, a transferred initialization does no better than a random one3
The top layers are the most general, so they are the ones to reuse when the tasks differ4
Freezing copied middle layers hurt even when retraining on the original task, because co-adapted neurons were split5

Timed test