ACCS Weekly meeting
Place: Room 621/622, GP South (Building 78)
Time: Thursday 22 June, 10:00am morning Tea, 10:30am seminar
Speaker: Flora Yu-Hui Yeh
Title: Practical Techniques in Probabilistic Active Learning
Machine learning is an important and active field in computer science research that is concerned with the development of algorithms and models that learn to perform functions based on data. Active learning is a class of supervised machine learning techniques where the learning algorithm is able to select the data to be used in the learning process. The correct output for any selected data point is provided when the active learning algorithm chooses to query that point. Active learning focuses on improving the performance of the learner through actively selecting data that is most informative, with the goal of learning an accurate model using a minimal amount of data. Different approaches to active learning have previously been developed based on statistical studies, information theory and heuristic approaches. While it is possible to theoretically formulate optimal active learning in some learning situations, in practice a trade off between the active learning algorithm's performance and computational expense cannot be avoided.
The aim of this project is to develop and evaluate practical active learning algorithms based on two kinds of probabilistic models: Gaussian Processes and Gaussian Mixture models.
These models have a number of useful properties and have been widely used in machine learning. While GPs and GMMs have been previously proposed for active learning, they have only had preliminary experimental evaluation, leaving open the question of their practical usefulness mainly open. Furthermore, their implementations use simple approximations to probability distributions (random sampling) and greedy/or exhaustive search to determine query points. This project will develop algorithms that aim to improve on these approximations using Markov chain Monte Carlo methods. The use of MCMC techniques will also be investigated in the implementation of Bayesian model averaging using GPs and GMMs, to take account of model uncertainty during active learning. The project will provide comprehensive experimental results on large datasets to assess the effectiveness of the techniques developed.
Dr. Ariel Liebman
Research Fellow
ARC Centre for Complex Systems
School of ITEE, University of Queensland
email:
a.liebman@uq.edu.au
tel: +61-7-3365-1623
mobile: +61-(0)414-226-336
room: 78-414, GP South, St Lucia Campus
web:
