DATE: 4pm Monday December 5th

LOCATION: 7th Floor Computer Science Boardroom, X736 ICICS Building

SPEAKER: Vinod Nair – Machine Learning Group, University of Toronto

TITLE: Learning To Do Vision by Inverting a Graphics Model

ABSTRACT: In this talk I’ll describe a new way of learning a generative representation for an image class of interest given a graphics model that can generate any instance from this class. Our approach is to learn a mapping from an image to the inputs of the graphics model that would generate that image. The key difficulty is that in a typical application we only have the images available for learning, and not the graphics inputs that generated them, so supervised learning is not possible. I’ll present a new unsupervised learning algorithm that gets around this problem. The algorithm is demonstrated on the task of modelling images of handwritten digits. Using a graphics program that can generate realistic digit images, a neural network is trained for each digit class to infer the graphics inputs from images. Such a representation can then be used in many ways. Digit classification can be done by seeing how well the representation extracted by each class-specific network reconstructs a test image. It can also be used to create new synthetic training examples for improving the classification performance of other, discriminative, learning algorithms. The basic idea behind this work is more generally applicable to other types of images, such as faces, and even other types of data, such as speech.

Joint work with Geoff Hinton.

BIO:
Vinod Nair is currently a Ph.D. student with the Machine Learning group at the University of Toronto’s Department of Computer Science. He received his master’s and bachelor’s degrees from the Dept. of ECE at McGill University in 2004 and 2002, respectively. His main research interests are in machine learning and computer vision.

Hosted by Dr. Sidney Fels.

Relevance: Although this is not my research area, it is being hosted by the director of my lab. It should be a good talk.

Technorati Tags: , , ,