Why put the results of the entire network into ADDA instead of the feature extraction output？

[iamxiaoyubei]

I found that you directly adapt all of the entire network layers from ADDA instead of the feature extraction layers. I am confused, is this still feature adaptation?