New Date April 18th 13:30 CET!
Physicist Miriam Klopotek (Uni Stuttgart/SimTech) will join our group seminar 13:30 CET to present some of her current research on the intersection of physics and ML.
What can we learn from and through machine learning if the physics of many-body systems is behind it?
Abstract:
In physics, the study of many-body systems hones in on how complex yet adaptive and even coordinated behavior emerges out of the motion of individual particles or agents. I am interested in how such physical emergence happens in artificial learning systems by finding specific analogies to many-body behavior. This has led me to develop several ‘physics-explainability’ techniques for ML that offer deepened insight into how algorithms work and what their limitations mean. At its heart, the problem of reducing the complexity of data may be akin to (time-dependent) coarse-graining in many-body systems. Moreover, some learning phenomena could be viewed as phase transformations. In a further step, I discuss how ML modeling – made intelligible in this way – could latch onto our cognition and lead to new insight in the realm of natural science. In fact, the analogy between ML and many-body dynamics is exact when information processing arises through physical dynamics – I discuss some results on reservoir computing through the non-equilibrium dynamics of swarm models.
The seminar will be hybrid. Joining online is possible using this link: https://unistuttgart.webex.com/unistuttgart/j.php?MTID=m64b57a647a4e91f475f67ad7d6ea6770