Fall 2012 Seminar Series

DATE: September 20th
LOCATION: E51-376
TIME: 4:15pm
Reception immediately following

SPEAKER:
Sham Kakade

TITLE
Scalable Tensor Methods for Learning Latent Variable Models

ABSTRACT
In modern applications, we are often concerned with modeling the interactions of many observed variables and in understanding how these interactions arise through certain hidden (or latent) causes. The (unsupervised) learning problem is to understand these interactions with only samples of the observed data. In practice, these models are often fit with local search heuristics (such as the EM algorithm) or sampling based approaches.

 

The first part of this talk will discuss how generalizations of standard linear algebra tools (e.g. spectral methods) to tensors provide provable and efficient methods for learning various latent variable models under mild non-degeneracy assumptions, including models such as mixtures of Gaussians, hidden Markov models, topic models, latent Dirichlet allocation, and models for communities in social networks. We will focus on both the theory and practical implications of using such approaches.

 

The second part is concerned with structure learning, where the goal is to discover both the existence of certain hidden causes and underlying graphical structure between these hidden causes and the observed variables. Here, we show how tensor and matrix decomposition methods can be utilized to discover this structure.