- talks - lists.cs.princeton.edu

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

Talk tomorrow - 3/26 -
by Ginny Hogan 25 Mar '08

25 Mar '08

Talk Wednesday 3/26 - 4:15pm - Computer Science Room 105 Software Transactions: A Programming-Languages Perspective Dan Grossman University of Washington With multicore processors bringing parallel computing to the masses, there is an urgent need to make concurrent programming easier. Software transactions hold great promise for simplifying shared-memory concurrency,and they have received enormous attention from the research community in the last couple years. This talk will provide an overview of work done at the University of Washington to help bring transactions to the next generation of programming languages. No prior knowledge of software transactions will be necessary. Our work complements research done on hardware and software algorithms for implementing transactions by considering essential issues regarding how transactions affect language semantics and language implementation. Our motivation takes the novel view that transactions can improve language support for concurrency much like garbage collection can improve language support for memory management. Our language design and language semantics work has considered the pitfalls of so-called "weak isolation" and how to avoid them, interaction with other language features like native calls and exceptions, and the implications for shared-memory consistency models. Our implementation work includes techniques for the special case of a uniprocessor, a whole-program static optimization that uses pointer information to remove unnecessary read- and write-barriers while providing "strong isolation", and ongoing work for allowing parallelism within transactions. Dan Grossman is an Assistant Professor in the Department of Computer Science & Engineering at the University of Washington. His research in the design and implementation of programming languages is aimed at improving software quality.

1 0

S Comandur preFPO
by Melissa M Lawson 19 Mar '08

19 Mar '08

Seshadhri Comandur will present his preFPO on Wednesday March 26 at 3:30PM in Room 402. The members of his committee are: Bernard Chazelle, advisor; Moses Charikar and Sanjeev Arora, readers; Boaz Barak and Rob Schapire, nonreaders. Everyone is invited to attend his talk. The abstract follows below. --------------------------------------- Title : Dealing with data : Online reconstruction and self-improving algorithms Abstract : With the ubiquity of large data sets, reading the whole input might itself by a luxury. Consider some large dataset that is supposed to (but does not) have some structural property, e.g., a set of numbers should be sorted, a set of points should in convex position, or a graph should be an expander. Let the input be represented by a function f (over a suitable domain and range). Our aim is to output a function g that satisfies the property, and does not differ too much from f. Furthermore, we want to generate g in an online sublinear fashion : given a domain point x, we can output g(x) in sublinear time. This is the problem of online reconstruction, which naturally has connections to property testing and other sublinear slgorithms. We study the reconstruction problem in various settings - monotonicity of functions over lattices, convexity of surfaces, and bounded degree expanders. The results obtained are also interesting in their own right, and provide sublinear algorithmic tools to study these properties. In an interesting aside, we use sublinear tools for online learning problems. We define stronger performance measures for learning problems, and provde very efficient algorithms that attain this performance. One of the key insights is how tools from the sublinear and streaming world allow us to construct efficient learning algorithms. Input in the real world is not always worst-case, and probably comes from some low entropy distribution over the input space. It seems natural that algorithms should run faster given this constraint. Is it possible to learn enough about the input distribution to optimize running time? This is the problem of designing self-improving algorithms, which can tune themselves to the input distribution and cut down their running time. We prove a variety of results for the classic problem of sorting numbers, and then extend these to computational geometry world.

1 0

11am Fri March 14 seminar on road/traffic monitoring in CS 105
by Jennifer Rexford 14 Mar '08

14 Mar '08

Hi folks, On Friday March 14, Venkat Padmanabhan -- who leads the Mobility, Networks, and Systems group at Microsoft Research India -- will be visiting. See http://research.microsoft.com/research/mns/default.aspx His talk is at 11am in the small auditorium (CS 105), with the title/abstract below. He'd also like to give interested folks an overview of the MSR lab in Bangalore, and to discuss career opportunities for internships, research positions, and visiting researchers. That session will start run from 2:15-3:30pm, in a room to-be-announced later. Please let me know if you'd like to meet with Venkat and/or join for lunch at noon. -- Jen Speaker: Venkat Padmanabhan Title: TrafficSense: Rich Road and Traffic Monitoring Using Mobile Smartphones Abstract: We consider the problem of monitoring road and traffic conditions in a city. Prior work in this area has required the deployment of dedicated sensors on vehicles and/or on the roadside, or the tracking of mobile phones by service providers. Furthermore, prior work has largely focused on the developed world, with its relatively simple traffic flow patterns. In fact, traffic flow in cities of the developing regions, which comprise much of the world, tends to be much more complex owing to varied road conditions (e.g., potholed roads), chaotic traffic (e.g., a lot of braking and honking), and a heterogeneous mix of vehicles (2-wheelers, 3-wheelers, cars, buses, etc.). To monitor road and traffic conditions in such a setting, we present TrafficSense, a system that piggybacks on smartphones that users carry around with them. Each participating smartphone performs rich sensing using its communication radios, GPS, accelerometer, and microphone, scans for and exchanges information from other participating phones in its neighbourhood, and reports processed information back to a central server for aggregation and reporting. We discuss key technical challenges, including energy optimization and the need for hands-free operation, present the design of TrafficSense, including how information from multiple sensors is combined to make inferences, and evaluate it in various road and traffic settings in Bangalore. (Joint work with Prashanth Mohan and Ram Ramjee.) Bio: Venkat Padmanabhan is a Senior Researcher and Research Manager at Microsoft Research India in Bangalore, where he founded and now leads the Mobility, Networks, and Systems group. Venkat was previously with Microsoft Research Redmond for 8.5 years. His research interests are in networked systems and his current projects focus on mobile and sensor systems, and network management. His professional service has included serving as program (co-)chair for ACM NOSSDAV 2004, ACM IMC 2005, and IEEE HotWeb 2008, and as an affiliate faculty member at the University of Washington, where he has taught and served in student thesis committees. Venkat holds a B.Tech. from IIT Delhi and an M.S. and Ph.D. from UC Berkeley, all in Computer Science.

1 1

Talk today 3/12 - 4:15PM - Computer Science Room 105 -
by Ginny Hogan 12 Mar '08

12 Mar '08

Reinventing Partially Observable Reinforcement Learning Eyal Amir University of Illinois, Urbana-Champaign Many complex domains offer limited information about their exact state and the way actions affect them. There, autonomous agents need to make decisions at the same time that they learn action models and track the state of the domain. This combined problem can be represented within the framework of reinforcement learning in POMDPs, and is known to be computationally difficult. In this presentation I will describe a new framework for such decision making, learning, and tracking. This framework applies results that we achieved about updating logical formulas (belief states) after deterministic actions. It includes algorithms that represent and update the set of possible action models and world states compactly and tractably. It makes a decision with this set, and updates the set after taking the chosen action. Most importantly, and somewhat surprisingly, the number of actions that our framework takes to achieve a goal is bounded polynomially by the length of an optimal plan in a fully observable, fully known domain, under lax conditions. Finally, our framework leads to a new stochastic-filtering approach that has better accuracy than previous techniques. * Joint work with Allen Chang, Hannaneh Hajishirzi, Stuart Russell, Dafna Shahaf, and Afsaneh Shirazi (IJCAI'03,IJCAI'05,AAAI'06,ICAPS'06,IJCAI'07,AAAI'07).

1 0