Machine Learning and Friends Lunch





home
past talks
resources
conferences

Discrimination with Generative Models


Sam Roweis
University of Toronto

Abstract


Logistic regression is a well understood statistical model for classification and is a simple example of a method that is trained discriminatively. In this talk I'll show how it is possible to train a fully generative model that has exactly the same classification performance as logistic regression but many other benefits, including the ability to learn a metric which compares feature vectors, to reject outliers even when they are far from the decision boundary, to sample typical feature vectors conditioned on a class, and to easily incorporate unlabelled data. Futhermore, I'll discuss how this trick of producing a generative model with the same discriminative performance as another model can be extended to more complex architectures such as Hidden Markov Models and Conditional Random Fields.

Back to ML Lunch home