Gradients as Features for Deep Representation Learning

Fangzhou Mu, Yingyu Liang, Yin Li

We address the challenging problem of deep representation learning--the efficient adaption of a pre-trained deep network to different tasks. Specifically, we propose to explore gra…