Subatomic, High Energy and Plasma Physics seminar series: Moses Chan

Abstract: In learning unknown parameters from physical experiments, uncertainty quantification plays an important role for understanding how confident one can be with their results, specifically, how the experimental data constrain the parameters.  In cases where numerical calculations (or computer simulation runs) are costly, the use of emulators has shown a promising direction.  Gaussian processes (GPs) are commonly considered as a "black-box" emulator method, non-intrusive, flexibile, good for interpolation, and provides natural uncertainty quantification. However, GPs also famously break down as the parameter dimension grows, or become intractable as the amount of data increases.  In this seminar, we take the perspective where GPs can be used as building blocks for constructing tractable emulator.  The main idea is to retain the advantageous properties of GPs while tackling its shortcomings.

Note: Attendees are welcome to bring along their laptops to run several computational examples and to explore the construction of GP-based emulators.