Expressivity and learnability: linear regions in deep ReLU networks - Naver Labs Europe