1. Introduction
The success of an intelligent system depends on having the right representation of data. Computer vision community has gradually moved from engineering features to letting the deep neural networks learn data representations, given a differentiable task objective [1], [18].