Richard Wilkinson
Wednesday 21st September 2016
Time: 4.00pm
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Gaussian process accelerated ABC
Approximate Bayesian computation (ABC) is a class of algorithms used for
doing Bayesian inference when you do not have access to the likelihood
function. Instead, all simulation is done using realisations from the
simulation. These methods are widely applicable, easy to implement, and
consequently have become popular in many applied scientific disciplines.
One of the major challenges for ABC methods is dealing with the
computational cost that arises from needing to repeatedly run the
simulator.
In this talk, I will discuss several approaches for overcoming this
cost. There are two main approaches: inverse modelling, where regression
models (random forests, neural nets etc) are used to directly learn a
mapping from the (high dimensional) simulator output to the input
parameter; and surrogate modelling approaches, where we instead seek to
approximate the behaviour of the forward simulator (i.e. from parameter
to simulator output), before inverting it to learn the parameters using
Bayes theorem. I will concentrate on the latter, but discuss the pros
and cons of both approaches and discuss how approaches from machine
learning have led to advances in the field.