Learning Scene Structures Using Correlated Nonparametric Models
Bayesian nonparametric models (BNP) have emerged as an important tool for describing complex phenomena. Nonparametric models allow the model size to vary dynamically instead of being fixed in advance, making it particularly suited in unsupervised tasks. However, the use of BNP models in practice faces two challenging problems: (1) how to capture statistical dependencies in a nonparametric setting, and (2) how to efficiently estimate BNP models from massive amount of data.
In this talk, I will present some of my recent efforts in addressing these problems. The talk consists of two parts. In the first part, I will introduce a new topic model called Latent Correlated Allocation, which can express correlations in both document-level and topic-level, thus allowing to capture a variety of relations that may arise in real problems. In addition, I will talk about an efficient method for learning nonparametric mixture models, which can complete the estimation in a single-pass over a large dataset. In the second part of the talk, I will introduce an application that adapts this model to learn the structures of outdoor scenes and present experimental results on semantic segmentation and layout hallucination.
Dahua Lin is a Research Assistant Professor at Toyota Technological Institute at Chicago (TTIC). He received his Ph.D. from the department of EECS at Massachusetts Institute of Technology in 2012, M.Phil. from the department of Information Engineering at the Chinese University of Hong Kong in 2007, and B.Eng. from the department of Electrical Engineering and Information Science at the University of Science and Technology of China in 2004. He was a research intern at Microsoft Research Silicon Valley, Microsoft Research Redmond, and Microsoft Research Asia, respectively in 2010, 2009, and 2004. He received the Best Student Paper Award at NIPS 2010, and the Outstanding Reviewer Awards at ICCV 2009 and ICCV 2011.
Dr. Lin's research covers several areas in computer vision and machine learning. In particular, he is interested in developing probabilistic models and inference techniques to describe and analyze images, videos, and their applications to solve real world problems (e.g. scene understanding and video recovery). He is also interested in Bayesian nonparametrics, inference over graphical models, and large scale optimization, which provides theoretical foundation for much of this work.