Recent Research on Deep Learning for AI

• Activity Number: 328
• Type: Invited
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Yoshua Bengio*+
• Companies: U. Montreal
• Keywords: statistical learning ; machine learning ; deep learning ; manifold learning ; MCMC ; latent variables

Abstract

Research in AI has progressed rapidly in recent years, with practical products being deployed in web search engines, object recognition on your cell phone, or speech recognition. Machine learning algorithms are at the heart of many of these developments: computers learn to capture statistical structure and discover underlying explanatory factors. In order to reduce the effect of the curse of dimensionality, deep learning algorithms were proposed that discover multiple levels of representation of the observed data, with more abstract features or latent random variables at higher levels of representation. This presentation will review major findings in this area and focus on the geometrical interpretation of representation learning, with the assumption (observed in AI tasks) that the unknown underlying data generating density concentrates in some lower-dimensional regions (called manifolds). Novel results will also be presented on the statistical interpretation of some of these procedures, and related issues that arise when trying to sample from such models using Monte-Carlo Markov Chain techniques.