Research Progress Meeting
Date: September 14, 2023
Time: 4:00- 5:00 pm
Location: Sessler Conference Room- 50A-5132 [In-Person and HYBRID]
Speaker: Uros Seljak (LBNL)
Title: AI for Physics, Physics for AI
Abstract: Artificial Intelligence (Machine Learning) is revolutionizing many aspects of our life, but its success stories in physics and astronomy are rare and limited to a few subfields only. I will argue that this is because physics applications require development of physics specific AI methods, rather than using off the shelf methods from the AI community. A few examples of physics specific nature of the data are large dimensionality of the data, stochastic nature of the data, and symmetries. I will argue that learning the data structures first using generative learning approaches such as Normalizing Flows enables not only better learning, but also provides additional information on robustness, such as anomaly detection. These methods applied to cosmology data show the promise of up to an order of magnitude improvement relative to traditional methods. Physics ideas have also influenced the development of AI, and many of these have been based on stochastic processes and sampling. I will discuss recently developed MicroCanonical Hamiltonian and Langevin Monte Carlo, which are a new class of sampling methods that outperform previous state of the art such as Hamiltonian Monte Carlo, in some cases by orders of magnitude. These new sampling methods will in turn enable solutions of physics problems that were not possible before, in a wide range of fields from cosmology to lattice QCD.
Join Zoom Meeting
https://lbnl.zoom.us/j/98854322464?pwd=K2tKUm1VZjRlV1J5RHE3cXdHQzRxdz09
Meeting ID: 988 5432 2464
Passcode: 142239