Lecture
Uncovering Shared Features for Learning Multiple Tasks
Speaker: |
Andreas Argyriou |
Date: |
Thursday, 28 June 2007 |
Time: |
13:00-14:30 |
Location: |
STEP-C Seminar Room, Building B, FORTH, Heraklion, Crete. |
Host: |
Ioannis Tsamardinos |
Abstract: |
We present a machine learning
method for integrating information from different sources and/or
of different forms. This situation arises often in biological
applications (e.g. different types of protein data), computer
vision (different visual features to be combined) and elsewhere.
We propose a solution based on the so called regularization methods
(e.g. support vector machines, logistic regression, least squares
regression). Our algorithm uses the data to learn optimal coefficients
which combine heterogeneous sources of information, according
to their relevance for the task at hand. In addition, we present a method for simultaneously learning multiple tasks. It is often the case that pooling data from different tasks helps in learning each task (in comparison to learning the tasks independently). Such examples are multiple medical databases, consumers' preferences, object recognition in vision etc. Specifically, we focus on learning features which are shared by all the tasks. We present algorithms which select or learn common features and lead to improvements in statistical performance. |
Bio: |
Andreas Argyriou is a PhD
candidate in the Department of Computer Science, University College
London. Born in Athens, Greece, he graduated on the top of his
secondary school class and won medals in international mathematical
olympiads. He then enrolled in MIT where he received a Bachelor's
and a Master's degree in Computer Science. Before pursuing a PhD,
he held positions in telecommunications companies in Greece. His
current research interests lie in the area of machine learning,
with a focus on kernel methods (such as SVMs and ridge regression),
multi-task learning and optimization. He has published his work
in NIPS, ICML, COLT and the Machine Learning journal. |