Learning theory and out of distribution detection

Abstract

Nowadays, Machines are starting to have a really important relevance in automation tasks. Learning could be considered as one of the hardest tasks we can encounter. The goal of this work is to introduce the theoretical foundations of this topic. After a short introduction on general learning and machine learning in chapter 1, we will introduce the fundamental concepts of Learning Theory in chapter 2, we will find what learning formally means and under which conditions a scenario can be learnable. In the last chapter we will introduce a pretty recent topic: Out of Distribution detection. A theory that appeared for the first time in 2017 and tries to formalize whether or not it would be possible for a machine to detect if we are trying to make predictions on data which it hasn’t been trained for. Again, we will try to find conditions under which a machine could learn this skill.