[CTN] CTN Seminar: Dr. Graham Taylor (Guelph), 3:30 Dec 16, PAS 2464

[Bryan Tripp]

Hi everyone,

Dr. Graham Taylor will give the next seminar, next Tuesday December
16. The title and abstract follow below.

The time and place are as usual, 3:30 in PAS 2464.

There is time for one or two individual meetings with the speaker
during the day. Please let me know if you are interested in this
and/or would like to join us for dinner after the talk.

Hope to see you there,

Bryan



Dr. Graham Taylor
Assistant Professor
School of Engineering
University of Guelph

Learning Representations with Multiplicative Interactions

Representation learning algorithms are machine learning algorithms
which involve the learning of features or explanatory factors. Deep
learning techniques, which employ several layers of representation
learning, have achieved much recent success in machine learning
benchmarks and competitions, however, most of these successes have
been achieved with purely supervised learning methods and have relied
on large amounts of labeled data. In this talk, I will discuss a
lesser-known but important class of representation learning algorithms
that are capable of learning higher-order features from data. The main
idea is to learn relations between pixel intensities rather than the
pixel intensities themselves by structuring the model as a tri-partite
graph which connects hidden units to pairs of images.  If the images
are different, the hidden units learn how the images transform. If the
images are the same, the hidden units encode within-image pixel
covariances. Learning such higher-order features can yield improved
results on recognition and generative tasks. I will discuss recent
work on applying these methods to structured prediction problems.