RI: Small: Probabilistic Latent Variable Models for Sparse Data

Toyota Technological Institute At Chicago, Chicago IL

Investigators

Abstract

This project focuses on learning parsimonious representations of complex high-dimensional data. In recent years, the PI has developed probabilistic latent variable models based on Gaussian processes that are able to capture complex interactions and perform state of the art prediction in diverse applications such as 3D human body tracking and collaborative filtering. However, in real-word applications such as the analysis of human motion or object recognition from images, the data is structured (e.g., the correlations in a video sequence can be captured with tensors), can come from different sources of information (e.g., video and audio), can be generated from a complex dynamical process or additional information might be available (e.g., labels). To address these real world problems, this project investigates extensions of the aforementioned probabilistic latent variable models to learn parsimonious representations of this complex data, focusing on the development of efficient learning algorithms that are able to handle large and sparse datasets. This research is strongly tied to an empirical performance goal, consisting of improving the state-of-the-art in both pose estimation and object recognition applications through the modeling of such complex interactions. The proposed research has broad impact in several areas of computer vision in particular human body motion estimation and tracking.