Efficient Nonparametric Bayesian Structured Prediction

Trung V. Nguyen (ANU, Research School of Computer Science)

CS HDR MONITORING

DATE: 2013-07-03
TIME: 12:00:00 - 12:30:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
Structured prediction is the class of machine learning problems where the output space has structure. A well-known instance is multiple output regression where the outputs are high-dimensional and correlated. Various techniques have been proposed for structured prediction, among which nonparametric Bayesian methods have proved effective due to the ability to accommodate uncertainty, for example in modelling the dependencies in the output space. However computation and inference in these methods are often intractable. In this talk, I will first describe a nonparametric model that significantly improves the scalability of standard Gaussian process regression. The model can be up to 5 times faster and can deal with problems of size not manageable by existing methods, while retaining comparable predictive performance. I will then discuss several future research directions to address other issues in using nonparametric Bayesian methods for structured prediction.