The aim of this talk is to explain that behind machine learning problems there are in fact optimization and statistical problems. After having introduced the subject, we will focus on a concrete and rather intriguing example of failure of deep neural networks (an "adversarial" example).