[Research] postdoc candidate talk NSH 3001 - noon today

[Jeff Schneider]

Here is the title and abstract for Roman Garnett's talk today.
Jeff.

-------------------------------------------------

Time-Series Prediction in the Presence of Changepoints and Faults

We will present a Bayesian approach to time-series prediction on
functions that can undergo various types of sudden, drastic changes.
We will address both changepoints (when the underlying process changes
behavior) and faults (when our observation mechanism changes in
nature).  The inference will be powered by Gaussian processes.
Typical covariance functions used in Gaussian process inference are
stationary and therefore inappropriate for the functions we will
consider.  We will introduce nonstationary covariance functions to
model various types of changepoints.  To make predictions, we will
estimate the full posterior predictive distribution by approximately
marginalizing the hyperparameters of our model using Bayesian
numerical integration.  We can also perform probabilistic changepoint
detection as a natural byproduct of the algorithm.  For faults, we
will propose a flexible observation model that can model various types
of changes in the observation process.  We will illustrate these
methods on several real-world functions.