Transfer Learning for Monte Carlo Methods
Lead Research Organisation:
University College London
Department Name: Statistical Science
Abstract
A key message from the UK government's Blackett review was that 'In the future, there will be a greater need for reliable, predictive models that are relevant to the large-scale, complex systems that we want to understand or wish to construct'. Modelling is primarily useful if the model can be used for prediction and if uncertainty around these predictions can be quantified, which typically requires advanced computational tools. As our models grow in scale, significant computational challenges are created. The novelty is not the computational tasks in and of themselves, but the sheer scale of our models, and hence the computational cost of tackling these tasks. For example, tsunami models require several hours of compute time per run, whilst large eddy simulation models for predicting the energy production of wind farms require thousands of hours of compute time per run. These large compute costs are severely limiting the practical use of these models for making accurate predictions about the future.
One set of techniques which can help tackle this challenge is 'transfer learning', a sub-field of machine learning which proposes to share data across similar tasks to improve predictions. For example, in tsunami modelling, we might be interested in the expected energy of a wave as it approaches two coastal cities in a given region. In this case, runs of the tsunami models to make predictions for city 1 could reasonably be re-used to improve the accuracy of our predictions for city 2.
In this project, propose to use transfer learning to transform one of the most widely used class of computational methods in statistics and machine learning: Monte Carlo methods. These are commonly use to compute the expected value of some quantity of interest, and have been applied not just to tsunami models, but virtually all areas of science, social science and humanities. The use of transfer learning in Monte Carlo methods is currently very limited, and this project will set the foundations needed to pave the way for its widespread adoption.
One set of techniques which can help tackle this challenge is 'transfer learning', a sub-field of machine learning which proposes to share data across similar tasks to improve predictions. For example, in tsunami modelling, we might be interested in the expected energy of a wave as it approaches two coastal cities in a given region. In this case, runs of the tsunami models to make predictions for city 1 could reasonably be re-used to improve the accuracy of our predictions for city 2.
In this project, propose to use transfer learning to transform one of the most widely used class of computational methods in statistics and machine learning: Monte Carlo methods. These are commonly use to compute the expected value of some quantity of interest, and have been applied not just to tsunami models, but virtually all areas of science, social science and humanities. The use of transfer learning in Monte Carlo methods is currently very limited, and this project will set the foundations needed to pave the way for its widespread adoption.
