September 2015 - talks - lists.cs.princeton.edu

Colloquium Speaker- Percy Liang Thurs Oct 1, 4:30pm
by Nicole E. Wagenblast 01 Oct '15

01 Oct '15

Colloquium Speaker Percy Liang, Stanford University Thursday, October 1st- 4:30pm Computer Science 105 Learning Hidden Computational Processes We are interested in prediction problems in which evaluating the learned function requires multiple intermediate steps of computation. One motivation is building a system that can answer complex questions: here the function would need to map "How many countries have held the Summer Olympics more than once?" to "3" by applying a sequence of aggregation and filtering operations on a database. In this talk, we examine two key machine learning problems that arise in this setting. First, how do we model the computational process? We argue that the classic paradigm of decoupling modeling from inference is inadequate, and we propose techniques that directly model the inference procedure. Second, learning is very difficult: in our example, the supervision "3" constrains the hidden computational process in a very indirect way. We propose methods that relax the output supervision in a parameterized way, and learn both the relaxation and model parameters jointly subject to an explicit computational constraint. Finally, we show some empirical progress on a new challenging question answering task. Percy Liang is an Assistant Professor of Computer Science at Stanford University (B.S. from MIT, 2004; Ph.D. from UC Berkeley, 2011). His research interests include (i) modeling natural language semantics, (ii) developing machine learning methods that infer rich latent structures from limited supervision, (iii) and studying the tradeoff between statistical and computational efficiency. He is a 2015 Sloan Research Fellow, 2014 Microsoft Research Faculty Fellow, a 2010 Siebel Scholar, and won the best student paper at the International Conference on Machine Learning in 2008.

1 4

Colloquium Speaker- Ming Zhang Tues Sept 29, 4:30pm
by Nicole E. Wagenblast 29 Sep '15

29 Sep '15

Colloquium Speaker Ming Zhang, Microsoft Research Tuesday, September 29, 4:30pm Computer Science 105 Software-Defined Datacenter Networks Today ’s cloud-scale services (e.g., search, social networking and cloud computing) operate on large datacenter networks (DCNs). Keeping these networks running smoothly is difficult, due to the sheer number of physical devices, the complexity of network software stack, and the dynamic nature of the environment. At any given moment, multiple devices experience component failures, are brought down for planned maintenance, are upgraded with new firmware, or are reconfigured to adapt to prevailing traffic demand. In this talk, I will go through some of the automated systems that are developed for managing the routing, configuration, and firmware and power of network devices in Microsoft Azure. I will first describe how NetPilot achieves fast, safe mitigation of common network failures by automatically deactivating or restarting offending links or devices. I will then describe how SWAN boosts the utilization of inter-datacenter networks by centrally controlling when and how much traffic each service sends and frequently reconfiguring the network’s data plane to match current traffic demand. Finally, I will present Network State Service (NSS) that manages the entire network state for Azure datacenter and wide-area networks. It serves as the foundation of multiple network management applications (including NetPilot and SWAN), allowing these applications to operate independently, while maintaining network-wide safety. Ming Zhang is a Senior Researcher in the Mobility and Networking group at Microsoft Research, where he focuses on datacenter and wide-area networking. He creates cutting-edge technologies that power Microsoft’s massive cloud networks. His research is published in leading systems and networking conferences including SIGCOMM, NSDI, OSDI, and MobiSys. His work on MobiPerf won the Open Internet App Award and People's Choice App Award in FCC Open Internet Challenge. He received his Ph.D. in Computer Science from Princeton University in 2005 and his B.S. in Computer Science from Nanjing University in 1999.

1 0

Colloquium Speaker- Ming Zhang Tues Sept 29, 4:30pm
by Nicole E. Wagenblast 28 Sep '15

28 Sep '15

Colloquium Speaker Ming Zhang, Microsoft Research Tuesday, September 29, 4:30pm Computer Science 105 Software-Defined Datacenter Networks Today ’s cloud-scale services (e.g., search, social networking and cloud computing) operate on large datacenter networks (DCNs). Keeping these networks running smoothly is difficult, due to the sheer number of physical devices, the complexity of network software stack, and the dynamic nature of the environment. At any given moment, multiple devices experience component failures, are brought down for planned maintenance, are upgraded with new firmware, or are reconfigured to adapt to prevailing traffic demand. In this talk, I will go through some of the automated systems that are developed for managing the routing, configuration, and firmware and power of network devices in Microsoft Azure. I will first describe how NetPilot achieves fast, safe mitigation of common network failures by automatically deactivating or restarting offending links or devices. I will then describe how SWAN boosts the utilization of inter-datacenter networks by centrally controlling when and how much traffic each service sends and frequently reconfiguring the network’s data plane to match current traffic demand. Finally, I will present Network State Service (NSS) that manages the entire network state for Azure datacenter and wide-area networks. It serves as the foundation of multiple network management applications (including NetPilot and SWAN), allowing these applications to operate independently, while maintaining network-wide safety. Ming Zhang is a Senior Researcher in the Mobility and Networking group at Microsoft Research, where he focuses on datacenter and wide-area networking. He creates cutting-edge technologies that power Microsoft’s massive cloud networks. His research is published in leading systems and networking conferences including SIGCOMM, NSDI, OSDI, and MobiSys. His work on MobiPerf won the Open Internet App Award and People's Choice App Award in FCC Open Internet Challenge. He received his Ph.D. in Computer Science from Princeton University in 2005 and his B.S. in Computer Science from Nanjing University in 1999.

1 0

PACM/CS Colloquium Speaker- Ben Recht, Monday Sept 28, 4:30pm
by Nicole E. Wagenblast 28 Sep '15

28 Sep '15

Ben Recht , University of California, Berkeley The Resilience of the Perceptron Monday, September 28th- 4:30pm 214 Fine Hall The most widely used optimization method in machine learning practice is the Perceptron Algorithm, also known as the Stochastic Gradient Method (SGM). This method has been used since the fifties to build statistical estimators, iteratively improving models by correcting errors observed on single data points. SGM is not only scalable, robust, and simple to implement, but achieves the state-of-the-art performance in many different domains. In contemporary systems, SGM powers enterprise analytics systems and is the workhorse tool used to train complex pattern-recognition systems in speech and vision. In this talk, I will explore why SGM has had such staying power, focusing on notions of stability and robustness. I will first discuss how SGM is robust to perturbations of the model and the updates. From a computing systems perspective, this robustness enables parallel implementations with minimal communication, with no locking or synchronization, and with strong spatial locality. I will then show how SGM is robust to perturbations of the data itself, and prove that any model trained with stochastic gradient method in a reasonable amount of time attains small generalization error. I will subsequently provide a new interpretation of common practices in neural networks, and provide a formal rationale for many popular techniques in training large, deep models.

1 2

4:30pm Thu Oct 15 talk on online learning and optimization in wireless networks (E-Quad B-205)
by Jennifer Rexford 24 Sep '15

24 Sep '15

> From: "Katharine G. Furda" <kfurda(a)Princeton.EDU> > Subject: EE Seminar - Thursday, October 15, 4:30pm, EQuad B205 - Bhaskar Krishnamachari > Date: September 24, 2015 at 11:00:02 AM EDT > To: ee-seminar(a)Princeton.EDU > Reply-To: "Katharine G. Furda" <kfurda(a)Princeton.EDU> > > > DEPARTMENT OF > ELECTRICAL ENGINEERING SEMINAR SERIES > > > > Speaker: Bhaskar Krishnamachari, University of Southern California > Title: Bandits and Newsvendors: Joint Online Learning and Optimization in Wireless Networks > Date: Thursday, October 15, 2015 > Time: 4:30pm > Room: E-Quad, B205 > Host: Mung Chiang > > > Abstract: Algorithms for online learning and decision-making under uncertainty have become popular in recent years to improve the performance of wireless networks in unknown dynamic environments. I will give a brief overview of certain classic problem formulations such as multi-armed bandits (MAB) and newsvendor problems, talk about their applications to wireless networking, and present some recent results from my group's research in this area. These include results for decentralized MAB, combinatorial MAB, contextual MAB, multi-period newsvendors, and optimized robotic network formation in unknown environments. This talk will cover joint work with students Dr. Yi Gai, Pranav Sarkar, Parisa Mansourifard, and Shangxing Wang, and faculty collaborators Rahul Jain, Tara Javidi, and Nora Ayanian. > > > Bio: Bhaskar Krishnamachari is an Associate Professor and Ming Hsieh Faculty Fellow in Electrical Engineering, and Director of the Autonomous Networks Research Group (http://anrg.usc.edu/ <http://anrg.usc.edu/>) at the University of Southern California's Viterbi School of Engineering. He works on the design and analysis of algorithms and protocols for next generation wireless networks. His co-authored papers have received best paper awards at IPSN (2004, 2010), MSWiM (2006) and MobiCom (2010), a best paper runner-up at SECON (2012), and a top-three paper at MSWiM (2014). He has received the NSF CAREER award (2004), the ASEE Terman Award (2010), and has been included on Technology Review Magazine's TR-35 list (2011), and Popular Science's Brilliant 10 list (2015). He has authored a book titled "Networking Wireless Sensors" published by Cambridge University Press. He is an Editor for the ACM Transactions on Sensor Networks, was the TPC Co-Chair for IPSN 2015, and is a TPC Co-Chair for WiOpt 2016.

1 0

Colloquium Speaker- Ming Zhang Tues Sept 29, 4:30pm
by Nicole E. Wagenblast 23 Sep '15

23 Sep '15

Colloquium Speaker Ming Zhang, Microsoft Research Tuesday, September 29, 4:30pm Computer Science 105 Software-Defined Datacenter Networks Today ’s cloud-scale services (e.g., search, social networking and cloud computing) operate on large datacenter networks (DCNs). Keeping these networks running smoothly is difficult, due to the sheer number of physical devices, the complexity of network software stack, and the dynamic nature of the environment. At any given moment, multiple devices experience component failures, are brought down for planned maintenance, are upgraded with new firmware, or are reconfigured to adapt to prevailing traffic demand. In this talk, I will go through some of the automated systems that are developed for managing the routing, configuration, and firmware and power of network devices in Microsoft Azure. I will first describe how NetPilot achieves fast, safe mitigation of common network failures by automatically deactivating or restarting offending links or devices. I will then describe how SWAN boosts the utilization of inter-datacenter networks by centrally controlling when and how much traffic each service sends and frequently reconfiguring the network’s data plane to match current traffic demand. Finally, I will present Network State Service (NSS) that manages the entire network state for Azure datacenter and wide-area networks. It serves as the foundation of multiple network management applications (including NetPilot and SWAN), allowing these applications to operate independently, while maintaining network-wide safety. Ming Zhang is a Senior Researcher in the Mobility and Networking group at Microsoft Research, where he focuses on datacenter and wide-area networking. He creates cutting-edge technologies that power Microsoft’s massive cloud networks. His research is published in leading systems and networking conferences including SIGCOMM, NSDI, OSDI, and MobiSys. His work on MobiPerf won the Open Internet App Award and People's Choice App Award in FCC Open Internet Challenge. He received his Ph.D. in Computer Science from Princeton University in 2005 and his B.S. in Computer Science from Nanjing University in 1999.

1 0

Thur Sep 17, 4:30pm, E-Quad B205-Anna Choromanska-Optimization for large-scale machine learning: large data and large model
by Jennifer Rexford 10 Sep '15

10 Sep '15

> > DEPARTMENT OF > ELECTRICAL ENGINEERING SEMINAR SERIES > > > > Speaker: Anna Choromanska, > Courant Institute of Mathematical Sciences, New York University > Title: Optimization for large-scale machine learning: large data and large model > Date: Thursday, September 17, 2015 > Time: 4:30pm > Room: E-Quad, B205 > Host: Emmanuel Abbe > > Abstract: The talk will focus on selected challenges in modern large-scale machine learning in two settings: i) large data setting and ii) large model (deep learning) setting. The first part of the talk will focus on the case when the learning algorithm needs to be scaled to large data. The multi-class classification problem will be addressed, where the number of classes (k) is extremely large, with the goal of obtaining train and test time complexity logarithmic in the number of classes. A reduction of this problem to a set of binary classification problems organized in a tree structure will be discussed. A top-down online tree construction approach for constructing logarithmic depth trees will be demonstrated, which is based on a new objective function. Under favorable conditions, the new approach leads to logarithmic depth trees that have leaves with low label entropy. Discussed approach comes with theoretical guarantees following from convex analysis, though the underlying problem is inherently non-convex. The second part of the talk focuses on the theoretical analysis of more challenging non-convex learning setting, deep learning with multilayer networks. Despite the success of convex methods, deep learning methods, where the objective is inherently highly non-convex, have enjoyed a resurgence of interest in the last few years and they achieve state-of-the-art performance. In the second part of the talk we move to the world of non-convex optimization where recent findings suggest that we might eventually be able to describe these approaches theoretically. The connection between the highly non-convex loss function of a simple model of the fully-connected feed-forward neural network and the Hamiltonian of the spherical spin-glass model will be established. It will be shown that under certain assumptions i) for large-size networks, most local minima are equivalent and yield similar performance on a test set, (ii) the probability of finding a “bad” local minimum, i.e. with high value of loss, is non-zero for small-size networks and decreases quickly with network size, (iii) struggling to find the global minimum on the training set (as opposed to one of the many good local ones) is not useful in practice and may lead to overfitting. Discussion of open problems concludes the talk. > > Bio: Anna Choromanska is a Post-Doctoral Associate in the Computer Science Department at Courant Institute of Mathematical Sciences, New York University. She is working in the Computational and Biological Learning Lab, which is a part of Computational Intelligence, Learning, Vision, and Robotics Lab, of Prof. Yann LeCun. She graduated with her PhD from Columbia University, Department of Electrical Engineering, where she was the The Fu Foundation School of Engineering and Applied Science Presidential Fellowship holder. She was advised by Prof. Tony Jebara. She completed her MSc with distinctions in the Department of Electronics and Information Technology, Warsaw University of Technology with double specialization, Electronics and Computer Engineering and Electronics and Informatics in Medicine. She was working with various industrial institutions, including AT&T Research Laboratories, IBM T.J. Watson Research Center and Microsoft Research New York. Her research interests are in machine learning, optimization and statistics with applications in biomedicine and neurobiology. She also holds a music degree from Mieczyslaw Karlowicz Music School in Warsaw, Department of Piano Play. She is an avid salsa dancer performing with the Ache Performance Group.

1 0

Alex Iriza will present his Master's thesis talk on Thursday, September 10, 2015 at 1:30pm in CS 401
by Nicki Gotsis 09 Sep '15

09 Sep '15

Alex Iriza will present his Master's thesis talk on Thursday, September 10, 2015 at 1:30pm in CS 401. The members of his committee are Jérémie Lumbroso (adviser) and Bob Sedgewick (co-adviser). Everyone is invited to attend his talk. His abstract follows below. Title: Enumeration and random generation of unlabeled classes of graphs: a practical study of cycle pointing and the dissymmetry theorem Abstract: Our work studies the enumeration and random generation of unlabeled combinatorial classes of unrooted graphs. While the technique of vertex pointing provides a straightforward procedure for analyzing a labeled class of unrooted graphs by first studying its rooted counterpart, the existence of nontrivial symmetries in the unlabeled case causes this technique to break down. Instead, techniques such as the dissymmetry theorem (of Otter) and cycle pointing (of Bodirsky et al.) have emerged in the unlabeled case, with the former providing an enumeration of the class and the latter providing both an enumeration and an unbiased sampler. In this work, we extend the power of the dissymmetry theorem by showing that it in fact provides a Boltzmann sampler for the class in question. We then present an exposition of the cycle pointing technique, with a focus on the enumeration and random generation of the underlying unpointed class. Finally, we apply cycle pointing to enumerate and implement samplers for the classes of distance-hereditary graphs and three-leaf power graphs.

1 0

Joey Dodds will present his FPO "Computation Improves Interactive Symbolic Execution" on Thursday, 9/10/2015 at 10am in CS 302
by Nicki Gotsis 01 Sep '15

01 Sep '15

Joey Dodds will present his FPO, "Computation Improves Interactive Symbolic Execution," on Thursday, 9/10/2015 at 10:00 AM, CS 302. The members of his committee are: Andrew Appel (adviser); Readers: David Walker, Jesper Bengtson(IT University of Copenhagen); Nonreaders: Sharad Malik, Aarti Gupta A copy of his thesis is available in Room 310. Everyone is invited to attend his talk. The talk title and abstract follow below: ________________________________________________________ "Computation Improves Interactive Symbolic Execution" As it becomes more prevalant throughout our lives, correct software is more important than it has ever been before. Verifiable C is an expressive Hoare logic (higher-order impredicative concurrent separation logic) for proving functional correctness of C programs. The program logic is foundational—it is proved sound in Coq w.r.t. the operational semantics of CompCert Clight. Users apply the program logic to C programs using semiautomated tactics in Coq, but these tactics are very slow. This thesis shows how to make an efficient (yet still foundational) symbolic executor based on this separation logic by using computational reflection in several different ways. Our execution engine is able to interact gracefully with the user by reflecting application-specific proof goals back to the user for interactive proof—necessary in functional correctness proofs where there is almost always domain-specific reasoning to be done. We use our “mostly sound” type system, computationally efficient finite-map data structures, and the MirrorCore framework for computationally reflected logics. Measurements show a 40× performance improvement.

1 1