Date:
14. September 2018 - 10:00
Speaker:
Elena Sellentin (Université de Genève -> University of Leiden)
Famous founders of social media services or internet search engines
repeatedly explain in the media that theory is dead, and the future
consists of labelled data and machine learning algorithms. The mentioned
machine learning algorithms are increasingly used in astronomy as well.
In my goodbye-talk, I will therefore describe why and how machine
learning cannot replace theoretical physics. The different points of
view can of course be debated, but it is surprisingly difficult to
really show with equations where machine learning is limited, and where
theoretical physics must take over. To demonstrate it, I will fit all
cosmological data sets simultaneously, with a single parameter, and
derive the best-fitting solution analytically. I will show that this fit
achieves arbitrary precision. Most statistical tests therefore prefer
this universal parameter over any theoretical model. The universal
parameter can only be discarded, once randomness is fully incorporated
as a feature into conventional theoretical physics.