Neil Chada - On optimal annealing schedules for Langevin algorithms (Joint with work Huy Chau)
|Starts:||12:00 7 Apr 2020|
|Ends:||13:00 7 Apr 2020|
|What is it:||Seminar|
|Organiser:||Department of Mathematics|
|Who is it for:||University staff, External researchers, Current University students|
|Speaker:||Dr Neil Chada|
Join us for this research seminar, part of the SQUIDS (Statistics, quantification of uncertainty, inverse problems and data science) seminar series.
Abstract: Simulated annealing is a popular optimization method designed to find global minimizers through perturbing dynamics with Gaussian noise. The continuous analogue of it is a diffusion process arising from a Langevin stochastic differential equation. In this work we consider understanding optimal schedules for both the step size and temperature which can optimally decay the mean squared error. This motivation is taken from Bayesian statistics within machine learning, where for stochastic gradient Langevin dynamics similar results were derived. Our key difference is that we consider these processes for optimization and not sampling. We derive annealing schedules for a variety of Langevin algorithms which coincide with results from the Bayesian paradigm. Furthermore we prove that the algorithms are consistent and satisfy a central limit theorem. We verify these results on a number of test functions in optimization.
Dr Neil Chada
Role: Research Fellow
Organisation: National University of Signapore
Travel and Contact Information
Frank Adams 1
Alan Turing Building