Monday, March 21, 2011
11:00 AM
Computer Studies Bldg. 209
Marc Aurelio Ranzato
University of Toronto
Unsupervised learning of hierarchical generative models with applications to vision and audio processing
Unsupervised learning algorithms aim to discover the structure hidden in the data and they often produce representations that are more suitable than the raw input for subsequent discriminative tasks. Learning representations and, in particular, hierarchical representations is especially useful for tasks involving perceptual processes because high variability and ambiguities in the sensory inputs easily disguise the underlying structure of the data. Hierarchical representations are composed by a sequence of non-linear transformations that are useful to efficiently capture complex statistical dependencies and to factor out irrelevant transformations from the input signal.

In this talk I will focus on a class of models that discover the structure latent in the data by fitting a probabilistic model to the empirical observations. This model, dubbed gated Markov Random Field, is a very powerful generative model that uses its latent variables to modulate singleton and pair-wise interactions between input variables. This model can be easily integrated into a hierarchical model by using its latent variables as data to train the layer above. This procedure can be recursively repeated to build a deep hierarchy. While previous attempts to use similar hierarchical models have found success mainly when generating from binary data in fairly low dimensional spaces, I will show that this model can generate high-resolution samples from an unconstrained distribution of natural images. These samples are far more realistic and structured than those produced by previous parametric generative models. I will also show how to exploit the generative ability of the model to improve performance on several vision tasks, such as recognition of facial expressions from heavily occluded images. Similarly, the same model can be applied to acoustic modeling to learn representations of speech. These representations are particularly useful in discriminative tasks, achieving the lowest phone error rate on the widely used TIMIT dataset to date.


Marc'Aurelio Ranzato is a postdoctoral fellow working with professor Geoffrey Hinton in the department of computer science at the University of Toronto. He received a PhD in computer science in 2009 under the supervision of professor Yann LeCun at New York University. He is recipient of the 2008-2009 NYU Dean's dissertation fellowship. Marc'Aurelio's major interests are in the areas of machine learning, computer vision, audio and video processing, natural language processing, and computational neuroscience.

Refreshments will be provided at 11:00 AM