Machine learning techniques are increasingly used for recognizing pattern and devising optimal strategies: situations in which the machine is taught (or teaches itself) to learn a known correct answer, or the best use of known rules. In particle physics, machine learning has been used now for several years in order to determine an underlying physical law which is known to exist, but which is unknown. Furthermore, because elementary particles are quantum objects, this law is stochastic in nature: the machine has to learn a probability distribution, rather than a unique answer. I will discuss some classic results, used among others in the discovery of the Higgs boson, as well as recent developments, which raise the fundamental question of how to decide whether an answer is correct.