March 2014 - talks - lists.cs.princeton.edu

Colloquium speaker: Richard Socher Tues, March 11, 4:30pm
by Nicole E. Wagenblast 11 Mar '14

11 Mar '14

Recursive Deep Learning for Modeling Compositional Meaning in Language Richard Socher ( Stanford University ) Tuesday, March 11, 4:30pm Computer science 105 Great progress has been made in natural language processing thanks to many different algorithms, each often specific to one application. Most learning algorithms force language into simplified representations such as bag-of-words or fixed-sized windows or require human-designed features. I will introduce three models based on recursive neural networks that can learn linguistically plausible representations of language. These methods jointly learn compositional features and grammatical sentence structure for parsing or phrase level sentiment predictions. They can also be used to represent the visual meaning of a sentence which can be used to find images based on query sentences or to describe images with a more complex description than single object names. Besides the state-of-the-art performance, the models capture interesting phenomena in language such as compositionality. For instance, people easily see that the "with" phrase in "eating spaghetti with a spoon" specifies a way of eating whereas in "eating spaghetti with some pesto" it specifies the dish. I show that my model solves these prepositional attachment problems well thanks to its distributed representations. In sentiment analysis, a new tensor-based recursive model learns different types of high level negation and how they can change the meaning of longer phrases with many positive words. They also learn that when contrastive conjunctions such as "but" are used the sentiment of the phrases following them usually dominates. Richard Socher is a PhD student at Stanford working with Chris Manning and Andrew Ng. His research interests are machine learning for NLP and vision. He is interested in developing new deep learning models that learn useful features, capture compositional structure in multiple modalities and perform well across different tasks. He was awarded the 2011 Yahoo! Key Scientific Challenges Award, the Distinguished Application Paper Award at ICML 2011, a Microsoft Research PhD Fellowship in 2012 and a 2013 "Magic Grant" from the Brown Institute for Media Innovation.

1 2

Colloquium Speaker Mohammad Alizadeh Monday March 10, 4:30pm
by Nicole E. Wagenblast 10 Mar '14

10 Mar '14

Mohammad Alizadeh ( Stanford University ) Monday, March 10, 4:30pm Computer Science 105 Packet Transport Mechanisms for Data Center Networks We are witnessing a fundamental shift in the computing landscape in recent years as much of the world's application workloads move to massive "cloud" data centers. These data centers enable the services we rely on daily, such as web search, social networking, and Internet commerce. They also require huge investments to deploy and operate. This has spurred significant interest in the industry and the research community in innovation for data centers, and in particular, for data center networks. A crucial feature of a data center network is its transport mechanism: the method by which data is transferred from one server to another. Ideally, this should occur at the highest possible rate and with the lowest possible latency. But there is an inherent tension between these requirements, especially with today 's state-of-the-art Transmission Control Protocol (TCP) in data centers. In this talk, I will present three data center transport designs with increasing levels of sophistication. First, I will describe Data Center TCP (DCTCP), a new congestion control algorithm for data centers that is now shipping with Windows Server 2012. DCTCP uses a simple modification to the TCP algorithm to achieve full throughput while reducing queueing latency in the switches by ~10x compared to TCP. I will present a control-theoretic analysis of DCTCP's control loop. I will then discuss HULL, an architecture that builds on DCTCP to deliver near-zero fabric latency: only propagation and switching latency, no queueing. Finally, I will describe pFabric, a minimalistic design that leverages application-level hints and very simple switch mechanisms to achieve near theoretically optimal scheduling of flows in the data center fabric. Mohammad Alizadeh is a Researcher at Insieme Networks (recently acquired by Cisco Systems). He received his Ph.D. in Electrical Engineering from Stanford University where he was advised by Balaji Prabhakar. Before that, he completed his undergraduate degree in Electrical Engineering at Sharif University of Technology. His research interests are broadly in networked systems, data center networking, and cloud computing. His dissertation work focused on designing high performance data center transport mechanisms. His research has garnered significant industry interest: the DCTCP algorithm has been implemented in Windows Server 2012; the QCN congestion control algorithm has been standardized as the IEEE 802.1Qau standard; and most recently, the CONGA adaptive load balancing mechanism has been implemented in Cisco's new flagship Application Centric Infrastructure products. Mohammad is a recipient of a Stanford Electrical Engineering Departmental Fellowship, the Caroline and Fabian Pease Stanford Graduate Fellowship, and the Numerical Technologies Inc. Prize and Fellowship.

1 2

C Park FPO
by Melissa M. Lawson 04 Mar '14

04 Mar '14

Christopher Park will present his FPO on Monday March 10 at 1PM in Room 302 (note room!) . The members of his committee are: Olga Troyanskaya, advisor; Kai Li and Mona Singh, readers; Moses Charikar and Rob Schapire, nonreaders. A copy of his thesis, "Computational Functional Genomics for Directing Biological Discoveries", is on display in Room 310. Everyone is invited to attend his talk.

1 0

Colloquium Speaker Barbara Engelhardt Monday, March 3, 4:30pm
by Nicole E. Wagenblast 03 Mar '14

03 Mar '14

Statistical models to study how a genetic variant impacts an organism Barbara Engelhardt ( Duke University ) Monday, March 3, 4:30pm Computer Science 105 Consider sequencing the genome of a newborn, and selecting targeted therapeutics early in life to reduce her lifetime risk of addiction, obesity, type II diabetes, or pancreatic cancer. While genome-wide association studies (GWAS) have unquestionably been successful in identifying reproducible genomic risk factors for complex human diseases, the promise of developing therapeutics to reduce the heritable portion of disease risk is far from fulfillment. The essential technological developments to fulfill this promise, however, are mainly in statistics and computation rather than in genomic experimental methods. I describe three genomic studies from my recent work. First, I identified a genetic variant that behaves differently depending on whether or not an individual takes cholesterol-reducing drugs. Second, I found that genetic variants that are associated with different cell traits are co-localized with a large variety of different regulatory mechanisms more often than expected. Third, I developed a model to uncover genetic variants that affect many traits simultaneously, where the trait measurements have substantial technical noise. Throughout, I emphasize statistical and computational challenges, and innovations necessary to fulfill this promise of genomic studies.

1 1