Deep Face Decoder: Towards understanding the embedding space of convolutional networks through visual reconstruction of deep face templates.