November 24th, 2010 @ 11:30am : Marc'Aurelio Ranzato (University of Toronto)

On the quest for good generative models of natural images

Marc'Aurelio Ranzato (University of Toronto)

The study of the statistical properties of natural images has a long history and has influenced many fields, from image processing to computational neuroscience. In the literature there is a myriad of generative models that have been proposed to explicitly capture these properties. However, none of them is able to generate realistic samples. In fact, samples have statistics that are more similar to random images than to natural images.

In this talk, I will present a very powerful generative model of high-resolution natural images, which is a Deep Belief Network with a gated MRF at the lowest layer. This model is able to generate much more realistic samples than previous models. These samples typically exhibit long range structures and smooth regions separated by sharp boundaries. We can use the generation ability of the model to gain understanding on the structure learned by the model, and also to better cope with missing values in the input. For instance, by using the model to fill-in occluded pixels we can extract features that are more useful for discrimination of expression categories from face images, yielding better accuracy than state-of-the art methods on that task.

This is joint work with V. Mnih, J. Susskind and G. Hinton at University of Toronto.

seminars/seminaritems/2010-11-24.txt · Last modified: 2011/02/27 17:08 by yann