I think the idea is that, for example, you start with some pre-trained visual object-identification network, remove the final logistic-regression layer, and rep

I think the idea is that, for example, you start with some pre-trained visual object-identification network, remove the final logistic-regression layer, and replace with a layer which you train your problem on.