SUNMONTUEWEDTHUFRISAT
Events for April 21, 2010
-
CS Colloq: Barak Fishbain
Wed, Apr 21, 2010 @ 11:00 AM - 12:00 PM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Talk Title: Network flow algorithms for sensor networks and visual data analysisSpeaker: Barak FishbainHost: Prof. Cyrus ShahabiAbstract:As digital environments become increasingly complex, and the tools for managing information become increasingly advanced, it is essential to assist users in selecting their short term and long term attentional focus. In this talk a novel graph-cut based approaches for multi-dimensional data analysis are presented. These methods are highly robust and most efficient which allows for the analysis of significantly large data sets. Air quality control and video segmentation are presented as representative applications.
Air quality control is addressed by the use of sensors network, where each sensor is mounted on a moving vehicle, for the purpose of detecting various threats. An example scenario is that of multiple taxi cabs each carrying a detector. The detectors' positions are continuously reported from GPS data. The level of detected risk is then reported from each detector at each position. The problem is to delineate the presence of a potentially dangerous source and its approximate location by identifying a small area that has an elevated concentration of reported risk. This problem of using spatially deployed mobile
detector networks to identify and locate risks is modeled and formulated. Then it is shown to be solvable in polynomial time and with a combinatorial network flow algorithm. The efficiency of the algorithm enables its use in real time, and in areas containing a large number of deployed detectors.
In video segmentation a typical goal is to group together similar objects, or pixels in the case of image processing. At the same time another goal is to have each group distinctly dissimilar from the rest and possibly to have the group size fairly large. These goals are often combined as a ratio optimization problem. State-of-the-art methods address these ratio problems by employing nonlinear continuous approaches, such as spectral techniques.
These spectral techniques deliver solutions in real numbers which are not feasible to the discrete partitioning problem. Furthermore, these continuous approaches are relatively computationally expensive. In this talk a novel graph-cut based approaches for optimally solving a set of segmentation ratio problems are presented. These algorithms guarantee optimal solution to the respective problem and consistent output between different runs.
These methods are most efficient which allows for the segmentation of significantly large video data sets.
The work was done with Prof. Dorit S. Hochbaum, University of California at Berkeley.Bio:Barak Fishbain received his Ph.D in EE from Tel-Aviv University, Israel in 2008. His research interests are Computer Vision, Image Processing, Video Surveillance and Medical Imaging. Currently he is a postdoctoral fellow in the Dept. of Industrial Engineering and Operations Research in the University of California at Berkeley, USA
Location: Charles Lee Powell Hall (PHE) - 333
Audiences: Everyone Is Invited
Contact: CS Front Desk
-
CS Colloq: Chun-Nan Hsu - CANCELLED
Wed, Apr 21, 2010 @ 03:30 PM - 05:00 PM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
CANCELLEDTalk Title: Accelerating Machine Learning by Aggressive ExtrapolationSpeaker: Chun-Nan HsuHost: Prof. Dennis McLeodAbstract:This talk presents how to accelerate statistical machine learning algorithms for large scale applications by aggressive extrapolation. Extrapolation methods, such as Aitken's acceleration, have the advantage that they can achieve quadratic convergence with an overhead linear to the dimension of the training data. However, they can be numerically unstable and their convergence is only locally guaranteed. We show that this can be fixed by a double extrapolation method. There are two options for the extrapolation, global or component-wise. Previously, it was not clear which option is more effective. We show a general condition to determine which option will be more effective and show how to apply the condition to the training of Bayesian networks and conditional random fields (CRF). Then we show that extrapolation can accelerate on-line learning with a method called Periodic Step-size Adaptation (PSA). We show that PSA is an approximation of a theoretic "single-pass" on-line learning method, which can converge to an empirical optimum in a single pass through the training examples. With a single-pass on-line learning method, disk I/O can be minimized when a training set is too large to fit in memory. Experimental results for a wide variety of models, including CRF, linear SVM, and convolutional neural networks, show that single-pass performance of PSA is always very close to empirical optimum. Finally, an application to gene mention tagging for biological text mining will be presented, which achieved the top score in BioCreative 2 challenge.Bio:Dr. Chun-Nan Hsu is a computer scientist at Information Sciences Institute (ISI). Prior to joining ISI, he is Research Fellow and Leader of the Adaptive Internet Intelligent Agents (AIIA) Lab at the Institute of Information Science, Academia Sinica, Taipei, Taiwan. His research interests include machine learning, data mining, databases and bioinformatics. He earned his M.S. and Ph.D. degree in Computer Science from the University of Southern California, Los Angeles, CA, in 1992 and 1996, respectively. In 1996, before he passed his doctoral oral exam, he had been offered a position as Assistant Professor at the Department of Computer Science and Engineering, Arizona State University, Tempe, AZ. He taught there for two years before he returned to Taiwan in 1998. Since 2005, he has been the principal investigator of the Advanced Bioinformatics Core, National Research Program in Genomic Medicine, Taiwan, and leading one of the largest research efforts in computerized drug design and discovery in Taiwan. In 2006, the first drug candidate due to the use of the software his team developed was commercialized. In 2007, his teams achieved the best scores in the BioCreative 2 text mining challenge. Dr. Hsu has published 78 scientific articles since 1993. Some of the articles have been cited more than 300 times. Currently, Dr. Hsu has been working on applying artificial intelligence to computational biology and bioinformatics.
Location: Mark Taper Hall Of Humanities (THH) - 114
Audiences: Everyone Is Invited
Contact: CS Front Desk
