Login
Talks & Seminars
Title: Scalable machine learning algorithms: Applications in vision, speech and geostatistics
Balaji Vasan,
Date & Time: August 24, 2009 14:25
Venue: SIC 205 classroom, 2nd floor, K. R. Bldg.
Abstract:
With improved sensors, the amount of data available in many learning problems has increased tremendously. This coupled with the fact that many modern learning algorithms have quadratic or cubic time complexity has necessitated the need for scalable learning algorithms. In this talk, I shall focus on algorithms to accelerate kernel based learning approaches. I shall first introduce two fast algorithms for summation of kernel functions, Gaussian in particular. The first approach, FIGTREE, is based on the ideas from computational physics (Fast Multipole Methods) and computational geometry. FIGTREE reduces the quadratic complexity to linear time complexity. The second approach is based on a recent hardware trend, Graphical Processing Units (GPU). The GPUs have been built with several cores and are extremely capable. Our second algorithm efficiently utilizes the multiple processors to parallelize kernel summations. I will explain the ideas behind each of these accelerations and then show a couple of sample applications, where these have been utilized for efficient learning. The applications that I shall discuss will include object recognition, speaker verification / identification, Gaussian process regression, kernel density estimation and kriging.
Speaker Profile:
http://www.umiacs.umd.edu/%7Ebalajiv/home.htm I am Balaji Vasan Srinivasan, I am doing my doctoral Studies in the Department of Computer Science under Dr. Ramani Duraiswami. I am working as a Graduate Research Assistant at the Perceptual Interfaces and Reality Lab and UMIACS under Dr. Ramani Duraiswami. I finished my Masters in the Department of Electrical and Computer Engineering majoring in Communication and Signal Processing in November 2008. Prior to my Masters, I graduated from the Department of Electrical and Electronics Engineering (DEEE) in College of Engineering, Guindy, Anna University, Chennai with a Bachelors in July 2006. My primary research interests are in accelerating machine learning algorithms using Graphical processors and scientific computing based approximations. I am currently involved in research projects where I apply these ideas to speaker identification and weather modeling problems. My Masters thesis was on using efficient Gaussian process regression for estimating non-linear dynamical models. I have also had experience in working on image processing and power electronics.
List of Talks

Webmail

Username:
Password:
Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback