RSSNI - Automatic architecture selection for hierarchical mixture of experts models.

Date: Thursday 14 March 2024, 1.00PM
Location: QUB & Online (MS Teams)
QUB, Belfast, NI (at 18 College Green, Ground Floor, Room 8)
MS Link and Map link in main text.
Local Group Meeting

Share this event

Third talk of 2024 in the RSSNI seminar series.
 
Talk on Thursday March 14th at 1pm GMT

RSSNI is pleased to have Dr. Nema Dean, from the School of Mathematics and Statistics, University of Glasgow, speak to us on the following topic:

Title:  Automatic architecture selection for hierarchical mixture of experts models.

Abstract:  Hierarchical mixture of experts (HMEs) are an example of a tree-structured modelling technique based on the divide and conquer principle. They present a powerful prediction methodology due to their flexibility and ability to adapt their complexity. They also allow for interpretability that can be attractive in practice. However, deciding on the architecture or form of the model/tree is not a trivial task. Rather than fitting large numbers of competing models and choosing via some performance metric, we propose growing trees during the model fitting process instead of selecting the architecture in advance. We present a flexible and adaptive way of performing automatic architecture selection using reversible jump in the Bayesian inference setting. The common issue of low acceptance rates for reversible jump algorithms is ameliorated with an adaptive smart proposal mechanism. The results of applying this and competing approaches is presented for Glasgow rental property prices data.

All welcome!

This is a hybrid event with Nema, in-person, speaking in QUB, Belfast, NI (at 18 College Green, Ground Floor, Room 8)  and on-line using MS Teams (link here). She will be only the second in-person speaker from the mainland since the pandemic and I hope the group will give her a typical NI welcome.

Best

Gilbert MacKenzie
Meetings Secretary &
Felicity Lamrock
Meetings Host
RSSNI
 
Dr. Nema Dean, from the School of Mathematics and Statistics, University of Glasgow,