Physics Colloquium - Prof. IIya Nemenman
December 2, 2019 - 3:00pm to 4:00pm
Marcus Nanotechnology Building
Conference Rooms 1116-1118
Modern machine learning has already proven itself as an indispensable tool incorporating experimental data into existing large-scale computational models of physical systems. However, using machine learning for discovering new physics from experiments is a much harder problem. This is especially true in biophysics, where traditional theoretical physics intuition based on symmetry and locality is hard to use, and hence automated approaches are likely to be the most impactful. In this talk, I will review a few instances where we were able to discover new physics this way, focusing on behavior of worms and song birds. That is, we were able to make interpretable inferences and generalizations from data, relate them to physical mechanisms, propose new useful experiments, and predict their outcomes, some of which have been confirmed since.