Saturday, March 28, 2009

CMU talk: Empowering switching linear dynamic systems with higher-order temporal structure

The VASC/Disney Research Pittsburgh seminar

Title: Empowering switching linear dynamic systems with higher-order temporal structure

Sangmin Oh, Georgia Institute of Technology

Date: Friday 3/27

Abstract:
Automated analysis of temporal data is a task of utmost importance for intelligent machines. For example, ubiquitous computing systems need to understand the intention of humans from the stream of sensory information, and health-care monitoring systems can assist patients and doctors by providing automatically annotated daily health reports. Moreover, a huge amount of multimedia data such as videos await to be analyzed and indexed for search purposes, while scientific data such as recordings of animal behavior and evolving brain signals are being collected in the hope to deliver a new scientific discovery about life.

In this talk, we will describe a class of newly developed time-series models in Bayesian network formulation. In particular, we will focus on the extensions of switching linear dynamic systems (SLDSs) with higher-order temporal structure and inference methods thereof. SLDSs have been used to model continuous multivariate temporal data under the assumption that the characteristics of complex temporal sequences can be captured by Markov switching between a set of simpler primitives which are linear dynamic systems (LDSs). In particular, we will focus on the extensions of SLDSs which are developed to address problems such as continuous labeling, robust labeling for data with systematic global variations, and hierarchical labeling.

First, we will present a data-driven MCMC inference method for SLDS model. The distinctive characteristic of this approach is that it turns heuristic labeling methods into data-driven proposal distributions of MCMC where the outcome results in a principled approximate inference method. In other words, it is a methodology to turn a novice into an expert. We show the resulting MCMC method for SLDSs where an inference problem is now solved which could not be addressed efficiently by Gibbs sampling previously.

Second, parametric SLDSs (P-SLDSs) explicitly model the global parameters which induce systematic temporal and spatial variations of data. The additional structure of PSLDSs allows us to conduct the global parameter quantification task which could not be addressed by standard SLDSs previously in addition to providing more accurate labeling ability.

Third, segmental SLDSs (S-SLDSs) provide the ability to capture descriptive duration models within LDS regimes. The encoded duration models are more descriptive than the exponential duration models induced within the standard SLDSs and allow us to avoid the severe problem of over-segmentations and demonstrate superior labeling accuracy.

Finally, we introduce hierarchical SLDSs (H-SLDSs), a generalization of standard SLDSs with hierarchic Markov chains. H-SLDSs are able to encode temporal data which exhibits hierarchic structure where the underlying low-level temporal patterns repeatedly appear in different higher level contexts. Accordingly, H-SLDSs can be used to analyze temporal data at multiple temporal granularities, and provide the additional ability to learn a more complex H-SLDS model easily by combining underlying H-SLDSs.

The developed SLDS extensions have been applied to two real-world problems. The first problem is to automatically analyze the honey bee dance dataset where the goal is to correctly segment the dance sequences into different regimes and parse the messages about the location of food sources embedded in the data. We show that a combination of the P-SLDS and S-SLDS models has demonstrated improved labeling accuracy and message parsing results. The second problem is to analyze the wearable exercise data where we aim to provide an automatically generated exercise record at multiple temporal resolutions.

Bio:
Sangmin is a PhD candidate in Collge of Computing at Georgia Institute of Technology. He received his BS in computer science with cum laude from Seoul National Univ, in 2003. During his PhD thesis work, Sangmin has focussed on developing time-series models to address problems such as continuous labeling, robust labeling for data with systematic global variations, and hierarchical labeling, where he published his work at major conferences and journals in computer vision and AI. Additionally, he worked on problems in robotics, signal processing, and graphics, where he co-authored several academic publications. He was a recipient of Samsung Lee Kun Hee fellowship from '03 to '07. His research interests include computer vision, machine learning, robotics, computer graphics, data mining, time-series modeling and computational linguistics.

No comments: