- talks - lists.cs.princeton.edu

Colloquium Speaker-Jason Lee, Monday, March 11 at 12:30pm
by Emily C. Lawrence 06 Mar '19

06 Mar '19

Colloquium Speaker Jason Lee from University of Southern California Monday, March 11 - 12:30pm Computer Science - Room 105 Host: Elad Hazan [ https://www.cs.princeton.edu/events/25822 | https://www.cs.princeton.edu/events/25822 ] On the Foundations of Deep Learning: SGD, Overparametrization , and Generalization Deep Learning has had phenomenal empirical successes in many domains including computer vision, natural language processing, and speech recognition. To consolidate and boost the empirical success, we need to develop a more systematic and deeper understanding of the elusive principles of deep learning. In this talk, I will provide analysis of several elements of deep learning including non-convex optimization, overparametrization , and generalization error. First, we show that gradient descent and many other algorithms are guaranteed to converge to a local minimizer of the loss. For several interesting problems including the matrix completion problem, this guarantees that we converge to a global minimum. Then we will show that gradient descent converges to a global minimizer for deep overparametrized networks. Finally, we analyze the generalization error by showing that a subtle combination of SGD, logistic loss, and architecture combine to promote large margin classifiers, which are guaranteed to have low generalization error. Together, these results show that on overparametrized deep networks SGD finds solution of both low train and test error. Bio: Jason Lee is an assistant professor in Data Sciences and Operations at the University of Southern California. Prior to that, he was a postdoctoral researcher at UC Berkeley working with Michael Jordan. Jason received his PhD at Stanford University advised by Trevor Hastie and Jonathan Taylor. His research interests are in statistics, machine learning, and optimization. Lately, he has worked on the foundations of deep learning, non-convex optimization algorithm, and adaptive statistical inference. He has received a Sloan Research Fellowship in 2019 and NIPS Best Student Paper Award for his work.

1 0

Marcela Melara will present her Pre FPO "Novel Systems for Intra-Process Least Privilege and Isolation" on Monday, March 11, 2019 at 2pm in CS 105
by Nicki Mahler 05 Mar '19

05 Mar '19

Marcela Melara will present her Pre FPO on Monday, March 11, 2019 at 2pm in CS 105. The members of her committee are as follows: Michael Freedman (adviser), Mic Bowman (Intel Labs), Amit Levy, Wyatt Lloyd, and Prateek Mittal All are welcome to attend. The talk title and abstract follow below. Title: Novel Systems for Intra-Process Least Privilege and Isolation Abstract: Third-party libraries reduce software development costs and effort. Designed for flexible reuse, libraries implement a small set of features, allowing developers to build applications by combining libraries that provide the desired functionality. However, third-party code also poses a great risk: because the source code is rarely inspected or even accessible by the application developer, bugs or vulnerabilities that can leak sensitive data may go unnoticed. Yet, existing data protection tools are insufficient because they do not enforce least privilege, restricting each library’s access to only those data it needs for its functionality. Prior proposals have addressed this issue with two main approaches: (1) running application components in separate processes for strong isolation, or (2) tracking individual data objects throughout the application to prevent unprivileged components from disclosing sensitive information. However, these approaches see limited adoption because they introduce significant integration complexity and performance overhead. This thesis proposes two systems for enforcing intra-process least privilege: both strongly isolate data within a single process address space while restricting access at the granularity of individual library functions. We first present Pyronia, a privilege separation system for language runtimes that targets IoT device applications. To protect sensitive OS resources and in-memory data objects, Pyronia combines three access control techniques: system call interposition, stack inspection, and memory protection domains. Developers then specify data access rules only for directly imported third-party functions in a central policy. We next present Griffin, a memory access control system for Intel SGX cloud applications. Intel SGX enables developers to run sensitive code inside an enclave, a hardware-protected memory region within an application’s address space. However, in practice, developers often include untrusted third-party libraries in the enclave, giving them unfettered access to all in-enclave data. Griffin leverages Intel memory protection keys (MPK) to partition an enclave and assign per-compartment access rules. As part of their SGX policy, developers declare sensitive data objects and access privileges for in-enclave functions. Griffin then automatically confines these data objects in MPK compartments. Pyronia and Griffin demonstrate the effectiveness of intra-process least privilege, reducing the performance impact on applications, and easing integration efforts for developers.

1 0

TODAY in B205 E-Quad - R. Srikant, University of Illinois at Urbana-Champaign
by Emily Lawrence 05 Mar '19

05 Mar '19

EE SEMINAR SERIES Speaker: R. Srikant, University of Illinois at Urbana-Champaign Title: Queues and Qs Day: Tuesday, March 5, 2019 Time: 4:30 pm Room: B205 Engineering Quadrangle Host: Prof. H. Vincent Poor Data Science relies on scalable resource allocation algorithms to operate data centers consisting of massive numbers of servers, as well as machine learning algorithms to extract information from data. In the first part of the talk, we will consider a resource allocation problem that arises in data centers and show that there exist low-complexity algorithms which use queue-length information to maximize throughput while ensuring that the packet delays do not scale with the size of the data center. Our result resolved a long-standing open conjecture at the intersection of probability, algorithms, and networks. In the second part of the talk, we will consider a reinforcement learning algorithm to estimate the Q-function of a given policy. We will present the first comprehensive finite-time performance bounds for the algorithm, which can potentially be used to characterize its sample complexity. The first part is joint work with Siva Theja Maguluri and the second part is joint work with Lei Ying. Bio: R. Srikant is the Fredric G. and Elizabeth H. Nearing Endowed Professor of Electrical and Computer Engineering and the Coordinated Science Lab at the University of Illinois at Urbana-Champaign. His research interests are in the areas of applied probability and stochastic networks, with applications to machine learning, cloud computing, and communication networks. He is the recipient of the 2019 IEEE Koji Kobayashi Computers and Communications Award and the 2015 IEEE INFOCOM Achievement Award. He has also received several Best Paper awards, including the 2017 Applied Probability Society Best Publication Award and the 2015 IEEE INFOCOM Best Paper Award. He was the Editor-in-Chief of the IEEE/ACM Transactions on Networking from 2013-2017. This seminar is supported with funds from the Korhammer Lecture Series

1 0

Colloquium Speaker-Tianqi Chen, TODAY at 12:30pm
by Emily C. Lawrence 05 Mar '19

05 Mar '19

Colloquium Speaker Tianqi Chen from the University of Washington Tuesday, March 5 - 12:30pm Computer Science - Room 105 Host: Ryan Adams [ http://www.cs.princeton.edu/events/25820 | http://www.cs.princeton.edu/events/25820 ] Learning-based Learning Systems Data, models, and computing are the three pillars that enable machine learning to solve real-world problems at scale. Making progress on these three domains requires not only disruptive algorithmic advances but also systems innovations that can continue to squeeze more efficiency out of modern hardware. Learning systems are in the center of every intelligent application nowadays. However, the ever-growing demand for applications and hardware specialization creates a huge engineering burden for these systems, most of which rely on heuristics or manual optimization. In this talk, I will present a new approach that uses machine learning to automate system optimizations. I will describe our approach in the context of deep learning deployment problem. I will first discuss how to design invariant representations that can lead to transferable statistical cost models, and apply these representations to optimize tensor programs used in deep learning applications. I will then describe the system improvements we made to enable diverse hardware backends . TVM, our end-to-end system, delivers performance across hardware back-ends that are competitive with state-of-the-art, hand-tuned deep learning frameworks. Finally, I will discuss how to generalize our approach to do full-stack optimization of the model, system, hardware jointly, and how to build systems to support life-long evolution of intelligent applications. Bio: Tianqi Chen is a Ph.D. candidate in the Paul G. Allen School of Computer Science & Engineering at the University of Washington, working with Carlos Guestrin on the intersection of machine learning and systems. He has created three major learning systems that are widely adopted: XGBoost , TVM, and MXNet (co-creator). He is a recipient of the Google Ph.D. Fellowship in Machine Learning.

1 0

Reminder: Colloquium Speaker-Tianqi Chen, tomorrow at 12:30pm
by Emily C. Lawrence 04 Mar '19

04 Mar '19

Colloquium Speaker Tianqi Chen from the University of Washington Tuesday, March 5 - 12:30pm Computer Science - Room 105 Host: Ryan Adams [ http://www.cs.princeton.edu/events/25820 | http://www.cs.princeton.edu/events/25820 ] Learning-based Learning Systems Data, models, and computing are the three pillars that enable machine learning to solve real-world problems at scale. Making progress on these three domains requires not only disruptive algorithmic advances but also systems innovations that can continue to squeeze more efficiency out of modern hardware. Learning systems are in the center of every intelligent application nowadays. However, the ever-growing demand for applications and hardware specialization creates a huge engineering burden for these systems, most of which rely on heuristics or manual optimization. In this talk, I will present a new approach that uses machine learning to automate system optimizations. I will describe our approach in the context of deep learning deployment problem. I will first discuss how to design invariant representations that can lead to transferable statistical cost models, and apply these representations to optimize tensor programs used in deep learning applications. I will then describe the system improvements we made to enable diverse hardware backends . TVM, our end-to-end system, delivers performance across hardware back-ends that are competitive with state-of-the-art, hand-tuned deep learning frameworks. Finally, I will discuss how to generalize our approach to do full-stack optimization of the model, system, hardware jointly, and how to build systems to support life-long evolution of intelligent applications. Bio: Tianqi Chen is a Ph.D. candidate in the Paul G. Allen School of Computer Science & Engineering at the University of Washington, working with Carlos Guestrin on the intersection of machine learning and systems. He has created three major learning systems that are widely adopted: XGBoost , TVM, and MXNet (co-creator). He is a recipient of the Google Ph.D. Fellowship in Machine Learning.

1 0

EE and CS Speaker-Stephen Tu, 3/7 at 12:30pm
by Emily C. Lawrence 01 Mar '19

01 Mar '19

EE and CS Speaker Stephen Tu from University of California, Berkeley Thursday, March 7 - 12:30pm Computer Science - Room 105 Hosts: Ryan Adams, CS and Yuxin Chen, EE [ https://www.cs.princeton.edu/events/25824 | https://www.cs.princeton.edu/events/25824 ] Safe and Reliable Reinforcement Learning for Continuous Control Many autonomous systems such as self-driving cars, unmanned aerial vehicles, and personalized robotic assistants are inherently complex. In order to deal with this complexity, practitioners are increasingly turning towards data-driven learning techniques such as reinforcement learning (RL) for designing sophisticated control policies. However, there are currently two fundamental issues that limit the widespread deployment RL: sample inefficiency and the lack of formal safety guarantees. In this talk, I will propose solutions for both these issues in the context of continuous control tasks. In particular, I will show that in the widely applicable setting where the dynamics are linear, model-based algorithms which exploit this structure are substantially more sample efficient than model-free algorithms, such as the widely used policy gradient method. Furthermore, I will describe a new model-based algorithm which comes with provable safety guarantees and is computationally efficient, relying only on convex programming. I will conclude the talk by discussing the next steps towards safe and reliable deployment of reinforcement learning. Bio: Stephen Tu is a PhD student in Electrical Engineering and Computer Sciences at the University of California, Berkeley advised by Benjamin Recht . His research interests are in machine learning, control theory, optimization, and statistics. Recently, he has focused on providing safety and performance guarantees for reinforcement learning algorithms in continuous settings. He is supported by a Google PhD fellowship in machine learning.

1 0

Olivier Savary Belanger will present his Pre FPO on Thursday, March 7, 2019 at 10am in CS 402.
by Nicki Mahler 28 Feb '19

28 Feb '19

Olivier Savary Belanger will present his Pre FPO on Thursday, March 7, 2019 at 10am in CS 402. The members of his committee are as follows: Andrew Appel (adviser), Greg Morrisett (Cornell), David Walker, Zak Kincaid, and Lennart Beringer All are welcome to attend. Please see below for the talk abstract. Interactive theorem provers allow for the development, in the same environment, of programs and of proofs about them. The programmatic portion of the development can then be extracted to code which is then compiled in an executable. However, unless both the extraction and compilation processes are formally verified, one has no guarantees that the proofs developed still apply to the resulting executable. In this talk, I will describe my work on CertiCoq, a verified extraction pipeline for the Coq theorem prover composing with the CompCert C verified compiler to achieve end-to-end correctness guarantees. First, I will present a proof framework to prove optimizations over the continuation-passing style (CPS) intermediate representation (IR) used in CertiCoq. This framework has been used by myself and others to prove the correctness of nontrivial optimizations. I will focus on a novel proof of correctness for a shrink reduction algorithm, a transformation combining in a single pass multiple optimizations which always result in smaller terms. Then, I will present a verified code generation translating the CPS IR into Clight, a front-end language of CompCert. I will show how it interfaces with a verified garbage collector and how its proof composes with the proof of correctness on CompCert. I will also describe the representation of values of inductive datatypes, inspired by OCaml, and the generated interface between the extracted program and regular C code.

1 0

The "D" word: solving for "diversity" on high-tech teams, TODAY at 4:30pm
by Emily C. Lawrence 28 Feb '19

28 Feb '19

The “D” word: solving for “diversity” on high-tech teams Janet Vertesi, Assistant Professor of Sociology at Princeton University Thursday, February 28th at 4:30 PM Bowen Hall 222 This talk gives an overview of the sociological factors that affect the construction of diverse, high-performing teams. Building on a summary of key issues that affect gender and racial disparities in high-performing occupations, the talk covers the current social science and vocabulary for addressing the problem, as well as strategies for moving forward, avoiding common traps, and protecting performance-based advancement. Bio: Dubbed “Margaret Mead among the Starfleet” by the Times Literary Supplement, Janet Vertesi is Assistant Professor of Sociology at Princeton University and an expert in the sociology of science, technology, and organizations. Vertesi’s past decade of research, funded by the National Science Foundation, examines how distributed robotic spacecraft teams work together effectively to produce scientific and technical results. Her book Seeing Like a Rover (University of Chicago Press, 2015) describes the collaborative work of the Mars Exploration Rover mission including the people, the images, and the robots who do science on Mars. Vertesi is also a long-time contributor to the Association of Computing Machinery conferences on human-computer interaction and computer-supported cooperative work. She is an advisory board member of the Data and Society institute in New York City and is a member of Princeton University’s Center for Information Technology Policy. To join the talk, please email seasdiversity(a)princeton.edu. This event co-sponsored by School of Engineering and Applied Science and Department of Computer Science.

1 0

The "D" word: solving for "diversity" on high-tech teams, tomorrow at 4:30pm
by Emily C. Lawrence 27 Feb '19

27 Feb '19

The “D” word: solving for “diversity” on high-tech teams Janet Vertesi, Assistant Professor of Sociology at Princeton University Thursday, February 28th at 4:30 PM Bowen Hall 222 This talk gives an overview of the sociological factors that affect the construction of diverse, high-performing teams. Building on a summary of key issues that affect gender and racial disparities in high-performing occupations, the talk covers the current social science and vocabulary for addressing the problem, as well as strategies for moving forward, avoiding common traps, and protecting performance-based advancement. Bio: Dubbed “Margaret Mead among the Starfleet” by the Times Literary Supplement, Janet Vertesi is Assistant Professor of Sociology at Princeton University and an expert in the sociology of science, technology, and organizations. Vertesi’s past decade of research, funded by the National Science Foundation, examines how distributed robotic spacecraft teams work together effectively to produce scientific and technical results. Her book Seeing Like a Rover (University of Chicago Press, 2015) describes the collaborative work of the Mars Exploration Rover mission including the people, the images, and the robots who do science on Mars. Vertesi is also a long-time contributor to the Association of Computing Machinery conferences on human-computer interaction and computer-supported cooperative work. She is an advisory board member of the Data and Society institute in New York City and is a member of Princeton University’s Center for Information Technology Policy. To join the talk, please email [ mailto:seasdiversity@princeton.edu | seasdiversity(a)princeton.edu ] . This event co-sponsored by School of Engineering and Applied Science and Department of Computer Science.

1 0

Colloquium Speaker-Tianqi Chen, 3/5 at 12:30pm
by Emily C. Lawrence 27 Feb '19

27 Feb '19

Colloquium Speaker Tianqi Chen from the University of Washington Tuesday, March 5 - 12:30pm Computer Science - Room 105 Host: Ryan Adams [ http://www.cs.princeton.edu/events/25820 | http://www.cs.princeton.edu/events/25820 ] Learning-based Learning Systems Data, models, and computing are the three pillars that enable machine learning to solve real-world problems at scale. Making progress on these three domains requires not only disruptive algorithmic advances but also systems innovations that can continue to squeeze more efficiency out of modern hardware. Learning systems are in the center of every intelligent application nowadays. However, the ever-growing demand for applications and hardware specialization creates a huge engineering burden for these systems, most of which rely on heuristics or manual optimization. In this talk, I will present a new approach that uses machine learning to automate system optimizations. I will describe our approach in the context of deep learning deployment problem. I will first discuss how to design invariant representations that can lead to transferable statistical cost models, and apply these representations to optimize tensor programs used in deep learning applications. I will then describe the system improvements we made to enable diverse hardware backends . TVM, our end-to-end system, delivers performance across hardware back-ends that are competitive with state-of-the-art, hand-tuned deep learning frameworks. Finally, I will discuss how to generalize our approach to do full-stack optimization of the model, system, hardware jointly, and how to build systems to support life-long evolution of intelligent applications. Bio: Tianqi Chen is a Ph.D. candidate in the Paul G. Allen School of Computer Science & Engineering at the University of Washington, working with Carlos Guestrin on the intersection of machine learning and systems. He has created three major learning systems that are widely adopted: XGBoost , TVM, and MXNet (co-creator). He is a recipient of the Google Ph.D. Fellowship in Machine Learning.

1 0