- talks - lists.cs.princeton.edu

B Brown FPO
by Melissa M Lawson 24 Apr '08

24 Apr '08

Benedict Brown will present his FPO on Thursday May 1 at 10:30AM in Room 402. The members of his committee are; Szymon Rusinkiewicz, advisor; David Dobkin and Tom Funkhouser, readers; Adam Finkelstein and Ken Steiglitz, nonreaders. A copy of his thesis, "Registration and Matching of Large Geometric Datasets for Cultural Heritage Applications" is on display in Room 310. Everyone is invited to attend his talk.

1 0

1-2pm Wed May 7 talk on learning structured bayesian networks in room 302
by Jennifer Rexford 23 Apr '08

23 Apr '08

Title: *Learning Structured Bayesian Networks: Combining Abstraction Hierarchies and Tree-Structured Conditional Probability Tables Speaker: **Marie desJardins*, University of Maryland Date/time: 1-2pm Wed May 7 Location: room 302 in the CS building Abstract: In this talk, I will describe our research on incorporating background knowledge in the form of feature hierarchies during Bayesian network learning. Feature hierarchies enable the learning system to aggregate categorical variables in meaningful ways, thus enabling an appropriate "discretization" for a categorical variable. In addition, by choosing the appropriate level of abstraction for the parent of a node, we also support compact representations for the local probability models in the network. We combine this notion of selecting an appropriate abstraction with context-specific independence representations, which capture local ndependence relationships among the random variables in the Bayesian network. Capturing this local structure is important because it reduces the number of parameters required to represent the distribution. This can lead to more robust parameter estimation and structure selection, more efficient inference algorithms, and more interpretable models. I will describe our primary contribution, the Tree-Abstraction-Based Search (TABS) algorithm, which learns a data distribution by inducing the graph structure and parameters of a Bayesian network from training data. TABS combines tree structure and attribute-value hierarchies to compactly represent conditional probability tables. In order to construct the attribute-value hierarchies, we investigate two data-driven techniques: a global clustering method, which uses all of the training data to build the attribute-value hierarchies, and can be performed as a preprocessing step; and a local clustering method, which uses only the local network structure to learn attribute-value hierarchies. Empirical results in several benchmark domains show that (1) combining tree structure and attribute-value hierarchies improves the accuracy of generalization, while providing a significant reduction in the number of parameters in the learned networks, and (2) data-derived hierarchies perform as well or better than expert-provided hierarchies. BIOGRAPHY Dr. Marie desJardins is an associate professor in the Department of Computer Science and Electrical Engineering at the University of Maryland, Baltimore County. Her research is in artificial intelligence, focusing on the areas of machine learning, multi-agent systems, planning, interactive AI techniques, information management, reasoning with uncertainty, and decision theory. Dr. desJardins can be contacted at the Dept. of Computer Science and Electrical Engineering, University of Maryland Baltimore County, 1000 Hilltop Circle, Baltimore MD 21250, mariedj(a)cs.umbc.edu,(410) 455-3967.

1 0

Refreshments
by Ginny Hogan 23 Apr '08

23 Apr '08

Are now available in the tea room - Followed by Umesh Vazirani's talk At 4:30 -

1 0

Talk today - 4:30 - COS 105
by Ginny Hogan 23 Apr '08

23 Apr '08

Umesh Vazirani, UC Berkeley, will present his talk "The Truth about Quantum Computers", at 4:30pm today COS 105 - Tea will be at 3:30pm

1 0

Zhiyan Liu FPO
by Melissa M Lawson 21 Apr '08

21 Apr '08

Zhiyan Liu will present her FPO on Friday April 25 at 1:30PM in Room 302 (note room). The members of her committee are: Kai Li, advisor; Adam Finkelstein and Olga Troyanskaya, readers; Tom Funkhouser and Szymon Rusinkiewicz, nonreaders. A copy of her thesis, "Scalable Isosurface Visualization", is on display in Room 310. Everyone is invited to attend her talk.

1 0

C Huttenhower preFPO
by Melissa M Lawson 18 Apr '08

18 Apr '08

(Please note - talk is in Carl Icahn Lab, Room 200) Curtis Huttenhower will present his preFPO on Friday April 25 at 10AM in Carl Icahn Lab Room 200. The members of his committee are: Olga Troyanskaya, advisor; Mona Singh and Hilary Coller (MOL), readers; Kai Li and David Botstein (MOL/Genomics), nonreaders. Everyone is invited to attend his talk. ----- TITLE: Data mining in large biological data collections ABSTRACT: Modern biology has developed a wealth of high-throughput experimental techniques. Many of these, like whole-genome sequencing, produce measurements simultaneously for every gene in an organism's genome. This means that a single assay can produce tens or hundreds of thousands of structured data points; some assays detect pairwise interactions between genes, squaring the amount of data. Each such genome-scale dataset represents a highly descriptive snapshot of cellular biology. As the number of such datasets reaches the thousands for many organisms, however, new opportunities arise to understand systems-level biology by means of very large scale data integration and analysis. My thesis focuses on three areas of opportunity presented by this situation. First, analysis of single high-throughput datasets is by no means a solved problem: as new experimental assays become available, each resulting dataset reveals new aspects of molecular biology. I will discuss results examining the yeast phosphoproteome by electron transfer dissociation mass spectrometry, which allows global profiling of phosphorylation signaling, as well as a statistical model of the transcriptional response to changes in growth rate. This model is descriptive of yeast biology and can be applied to predict instantaneous growth rates from arbitrary expression measurements (e.g. from other experimental platforms or unicellular organisms). Second, it is critical to extend functional genomic analyses to higher organisms, particularly human beings. This presents both computational and biological challenges: the amount of available experimental data is orders of magnitude larger, the amount of pre-existing biological knowledge is smaller, and the biology of higher eukaryotes is itself substantially more complex. I will discuss HEFalMp, a machine learning system that builds off of our success in yeast to integrate hundreds of genome-scale datasets for functional genomic analysis in human beings. This has enabled new views of gene function, functional relationships, and pathway interactions in human biology. We are currently in the process of experimental followup for several predictions in the human autophagy pathway, and we are planning to extend HEFalMp to more specifically investigate human disease. Finally, the process of experimental data integration for functional genomics can be extended beyond single gene interactions or single organisms. By mining a sufficiently large body of experimental results, purely data-driven relationships can be derived between entire cellular pathways. This captures the complex interplay between pathways and processes at a systems level and can help to describe biological results in terms of their specific functional activity. Furthermore, all of these results can be extended across multiple organisms to study gene and pathway functional changes over the course of evolution. All of these results are made possible by a novel synthesis of large-scale machine learning with genome-scale biological data.

1 0

R. Dondero talk Friday 12:30
by Andrew W. Appel 15 Apr '08

15 Apr '08

Special PLC Seminar on a topic in Software Engineering: Robert Dondero will talk Friday April 18, room 302, at 12:30. [next door to the usual PLC room] Pizza will be provided. If you plan to attend AND you do not usually attend the PLC lunch, e-mail cbell(a)cs.princeton.edu so we order enough food. Predicting Software Change Coupling Robert Dondero Abstract: This project was an exploratory study of techniques for predicting future change coupling among a program's source code files. Two source code files are change coupled if programmers edit them together frequently, and separately infrequently. Specifically, this project investigated the predictive power of three approaches: mining of software change logs, software similarity detection, and software proximity detection. "Software mining" extracts patterns from source code databases, that is, version control systems containing source code and change histories. This project explored whether identification of past change coupling among source code files can predict future change coupling among those files. "Software similarity detection" finds files that contain similar, alias cloned, code. This project explored whether identification of similar code among source code files can predict future change coupling among those files. Finally, "software proximity detection" finds files that reference each other heavily. This project explored whether identification of proximity among source code files can predict future change coupling among those files. This project performed the study applied a software miner (created specifically for this project), three preexisting similarity detectors, and two proximity detectors (created specifically for this project) to four large open source code databases at multiple points in time. It determined that software mining generally generated the best predictions of the three approaches, followed by similarity detection, followed by proximity detection. Excessive source code change coupling can be a serious maintenance problem. So the prediction of future change coupling is an important challenge in software engineering. The results of this project shed light on the abilities of the three approaches, both in the absolute and relative senses, to predict change coupling. So the results of this project hold promise for decreasing program maintenance costs.

1 0

M Burns preFPO
by Melissa M Lawson 07 Apr '08

07 Apr '08

Michael Burns will present his preFPO on Friday April 11 at 10:30AM in Room 402. The members of his committee are: Adam Finkelstein, advisor; Szymon Rusinkiewicz and Doug DeCarlo (Rutgers), readers; Perry Cook and Tom Funkhouser, nonreaders. Everyone is invited to attend his talk. His abstract follows below. -------------------------- Title: Efficient and Comprehensible Visualization of Complex 3D Data As computing power increases, so does our power to scan or model increasingly complex 3D objects and scenes. Complexity may arise by the structure of the scene, such as having many nested objects, or by a large quantity of data. Traditional rendering pipelines fall short in exposing occluded objects or surfaces of interest in such complex scenes, and may result in important objects being visualized ineffectively or not at all. This thesis presents several techniques for creating comprehensible renderings of these complex scenes. Line drawings can be used to effectively visualize large scenes of volumetric data. Such data sets, which may result from fluid simulations or medical scans, often contain nested features of interest, such as bones being inside flesh in a CT scan. By using lines derived from implicitly defined isosurfaces, we can visualize several nested structures at once. Furthermore, by using lines instead of traditional volumetric rendering or isosurface extraction, we can generate images by processing far fewer data points for large, high- resolution volumetric data sets. For more traditional surfaces renderings, interior objects can be exposed by using cutaway structures. These view-dependent cutaways are defined by a cutaway surface that can be computed and represented either geometrically or by a screen-aligned height field. In either case, one can use this cutaway surface to affect both volumetric and polygonal renderings. We show how contextual cutaways can be used in volumetric renderings to expose interior materials of interest along with a specially defined region of interest, and present an example of rendering live Ultrasound data in the context of a medical CT scan, with various tissues being exposed around the plane of the Ultrasound probe. We also show how cutaways can be used in polygonal renderings to expose objects of interest in complex scenes, presenting examples in architectural modeling where furniture and interior structures can be viewed in the context of the exterior of the building.

1 0

Talk TOMORROW
by Ginny Hogan 26 Mar '08

26 Mar '08

Computer Science Talk Tomorrow - 4:15pm Computer Science Room 105. "Recent Directions in Nonparametric Bayesian Machine Learning Zoubin Ghahramani Carnegie Mellon University Machine learning is an interdisciplinary field which seeks to develop both the mathematical foundations and practical applications of systems that learn, reason and act. Machine learning draws from many fields, ranging from Computer Science, to Engineering, Psychology, Neuroscience, and Statistics. Because uncertainty, data, and inference play a fundamental role in the design of systems that learn, statistical methods have recently emerged as one of the key components of the field of machine learning. In particular, Bayesian methods, based on the work of Reverend Thomas Bayes in the 1700s, describe how probabilities can be used to represent the degrees of belief of a rational agent. Bayesian methods work best when they are applied to models that are flexible enough to capture to complexity of real-world data. Recent work on non-parametric Bayesian methods provides this flexibility. I will touch upon key developments in the field, including Gaussian processes, Dirichlet processes, and the Indian buffet process (IBP). Focusing on the IBP, I will describe how this can be used in a number of applications such as collaborative filtering, bioinformatics, cognitive modelling, independent components analysis, and causal discovery. Finally, I will outline the main challenges in the field: how to develop new models, new fast inference algorithms, and compelling applications.

1 0

Talk Today - 4:15 PM
by Ginny Hogan 26 Mar '08

26 Mar '08

Dan Grossman, University of Washington, will present his talk "Software Transacations: A Programming- Languages Perspective" at 4:14 today in Computer Science Room 105.

1 0