Place: Duthie Center for Engineering Room 117
Date: Friday, Jan 22, 2010
Seminar: 3:30 p.m.
Reception: 3:00 – 3:30pm. 2nd Floor

Multi-stream Hidden Markov Model (HMM) framework

Speaker: Mr. Oualid Missaoui
PhD candidate in Computer Engineering and Computer Science
University of Louisville

Abstract

We propose a multi-stream Hidden Markov Model (HMM) framework. We hypothesize that raw temporal data can be interpreted better by multiple synchronous sequences representing different modalities or sources of information that capture different characteristics. Our work is motivated by the fact that individual modalities are not reliable. Also, different modalities have non-overlapping relevance. Thus, ideally different modalities, may be combined to achieve higher recognition. In order to fuse the different modalities, a multi-stream HMM that includes a stream relevance weighting component is developed. In particular, general discrete and continuous probability distributions that include stream relevance weighting components are integrated into the proposed multi-stream HMM. We argue that these novel structures alleviate the limitations of the existing multi-stream HMM. We generalize the Baum-Welch and MCE/GPD algorithms and derive the necessary conditions to update all model parameters simultaneously. Results on synthetic data set and a collection of ground penetrating radar signatures show that the proposed multi-stream continuous HMM framework outperforms the basic single stream HMM where all the streams are treated equally important. It also outperforms the existing multi-stream HMM structures.

Bio

Oualid Missaoui earned his engineering degree in Signal and Systems (2003) from Polytechnic School of Tunisia, where he also received his M.S. degree in Applied Mathematics (2005). In January 2006, he joined the University of Louisville where he is currently a PhD candidate in Computer Engineering and Computer Science. His research interests include machine learning, data mining, time series analysis, statistical analysis, and image processing.