RT @RandomlyWalking: Fragile co-adaptation in deep learning. Sometimes fixing early layers is bad. Yosinski et al. #NIPS2014