AI4Research open seminar

Machine learning techniques are often stochastic and black box. In contrast, applications in (theoretical) Physics or Mathematics usually call for exact and verifiable results. This raises the question of how Machine Learning techniques, which have proven extremely successful in everyday life, can be adapted in these areas. In particular, there are many problems in Physics and Mathematics that are prohibitively expensive to compute since the number of possible outcomes is combinatorially large. I will explain how to adapt a machine learning technique called reinforcement learning to use neural networks in such cases while obtaining exact and verifiable results and illustrate the idea with examples from Physics and Mathematics.
Fabian