Colloquium Series 2014-2015

Graham Taylor
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.