B1931
Title: Multilevel Bayesian deep neural networks
Authors: Neil Chada - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Ajay Jasra - KAUST (Saudi Arabia)
Kody Law - University of Manchester (United Kingdom)
Sumeetpal Singh - Cambridge University (United Kingdom)
Abstract: The application of multilevel Monte Carlo for Bayesian computation tasks in machine learning is considered. There has recently been a synergy of statistics and machine learning, promoting the application and development of new methodologies. Based on this, we promote the use of multilevel Monte Carlo, which is a technique used to reduce the cost to attain a particular order of MSE with trace-class neural network priors. We provide some theoretical insights, and demonstrate the performance of our methodology on different model problems, such as classification and reinforcement learning.