- talks - lists.cs.princeton.edu

ORFE Distinguished Lecture Series: Costis Daskalakis on Monday, Jan 29 at 4:30pm
by Emily C. Lawrence 23 Jan '24

23 Jan '24

=== 1/29/24 ORFE Distinguished Lecture Series=== DATE: Monday, January 29, 2024 TIME: 4:30pm LOCATION: Sherrerd 101 SPEAKER: Costis Daskalakis, Massachusetts Institute of Technology TITLE: Training deep neural nets to be strategic Abstract: Many outstanding challenges in Deep Learning lie at its interface with Game Theory: from playing difficult games like Go to robustifying classifiers against adversarial attacks, training deep generative models, and training DNN-based models to interact with each other and with humans. In these applications, the utilities of the agents are non-concave in the parameters of the underlying DNNs; as a result, many standard notions of equilibrium fail to exist, and game-theoretic analysis becomes inadequate. So how does one train DNNs to be strategic? And what is even the goal of the training? We shed light on these challenges through a combination of learning-theoretic, complexity-theoretic, game-theoretic and topological techniques, presenting obstacles and opportunities for Deep Learning and Game Theory going forward. Bio: Constantinos (aka "Costis") Daskalakis is the Avanessians Professor of Electrical Engineering and Computer Science at MIT. He holds a Diploma in Electrical and Computer Engineering from the National Technical University of Athens, and a PhD in Electrical Engineering and Computer Science from UC Berkeley. He works on Computation Theory and its interface with Game Theory, Economics, Probability Theory, Machine Learning and Statistics. He has resolved long-standing open problems about the computational complexity of Nash equilibrium, and the mathematical structure and computational complexity of multi-item auctions. His current work focuses on multi-agent learning, high-dimensional statistics, learning from biased and dependent data, causal inference and econometrics. He has been honored with the ACM Doctoral Dissertation Award, the Kalai Prize from the Game Theory Society, the Sloan Fellowship in Computer Science, the SIAM Outstanding Paper Prize, the Microsoft Research Faculty Fellowship, the Simons Investigator Award, the Rolf Nevanlinna Prize from the International Mathematical Union, the ACM Grace Murray Hopper Award, the Bodossaki Foundation Distinguished Young Scientists Award, the ACM SIGECOM Test of Time Award, and the FOCS 2022 Test of Time Award. He is an ACM fellow, holds an honorary doctorate from the University of Patras, and was awarded the Golden Cross of the Order of the Redeemer by the Greek President. He has served in the scientific and advisory board of the Simons Institute for the Theory of Computing (2018-2020), and as the head of the theory of computation group at MIT (2018-2022). He is a co-founder of Archimedes AI research center where he maintains the role of chief scientist.

1 0

Jeffrey Helt will present his FPO "Holistic Consistency Models for Faster Applications and Systems" on Friday, February 16, 2024 at 1pm in CS 105.
by Gradinfo 22 Jan '24

22 Jan '24

Jeffrey Helt will present his FPO "Holistic Consistency Models for Faster Applications and Systems" on Friday, February 16, 2024 at 1pm in CS 105. The members of his committee are as follows: Examiners: Wyatt Lloyd (adviser), Michael Freedman, and Zachary Kincaid Readers: Ravi Netravali and Amit Levy Abstract follows below. All are welcome to attend. The correctness and performance of today’s applications are heavily influenced by the consistency models of their supporting services. Strong consistency models, like strict serializability and linearizability, restrict the space of possible service behaviors and in turn, simplify application programming. In exchange for their strong guarantees, however, strictly serializable and linearizable services incur worse performance than those with weaker consistency. But switching to such services can break applications. Consistency models thus offer a harsh trade-off between application correctness and service performance. Despite impacting both applications and services, most existing work on consistency models has taken a limited view. A long line of research sought to weaken the consistency of common services in an effort to improve their performance, largely without considering the impacts of weaker consistency on application correctness. A mostly separate line of work investigated the impacts of weaker consistency on applications. In contrast, this dissertation takes a holistic approach, considering applications and services simultaneously to develop better consistency models. First, we introduce Regular Sequential Serializability (RSS) and Regular Sequential Consistency (RSC). They are respectively as strong as strict serializability and linearizability for applications: we prove any application that is correct when using a strictly serializable (linearizable) service is also correct when using an RSS (RSC) service. And yet, they enable new, better-performing services, thus circumventing the correctness-performance trade-off. To demonstrate the latter, we design, implement, and evaluate variants of two systems, Spanner-RSS and Gryff-RSC. The variants achieve lower tail read latency than their existing counterparts. Second, we introduce Multi-dispatch Linearizability (MD-Linearizability), which relaxes linearizability’s assumption that a client is sequential and instead explicitly allows it to issue multiple operations concurrently. This can potentially reduce an application’s end-to-end latency. To this end, we present Ellis, the first multi-shard system to guarantee MD-Linearizability, and show it reduces end-to-end application latency up to 75%. Importantly, we also describe how to transform applications built atop linearizable services into ones that interact with comparable MD-Linearizable services. Following the transformation ensures the new applications behave identically to the originals, so programmers can reap MD-Linearizability’s performance benefits without worrying about breaking their applications.

1 0

Tim Alberdingk Thijm will present his FPO "Modular Control Plane Verification" on Friday, February 9, 2024 in CS 105 (and Zoom).
by Gradinfo 12 Jan '24

12 Jan '24

Tim Alberdingk Thijm will present his FPO "Modular Control Plane Verification" on Friday, February 9, 2024 in CS 105 (and Zoom). Zoom link: https://princeton.zoom.us/j/95302742818 The members of his committee are as follows: - Advisor: Aarti Gupta - Reader: David Walker - Reader: Ryan Beckett (Microsoft Research) - Examiner: Maria Apostolaki (ECE) - Examiner: Zak Kincaid Please see talk abstract below. All are welcome to attend. Abstract: Networks continue to experience costly outages due to misconfigured control planes. Avoiding misconfigurations is challenging: control planes often use notoriously inscrutable distributed routing protocols. One remedy is control plane verification, which analyzes control planes to uncover violations of desired properties. Unfortunately, most verification tools analyze the entire network at once. Hence, these monolithic tools often do not scale well as networks grow in size and properties increase in complexity, or compromise on the networks and properties they support. To address these limitations, we explore modular control plane verification, where the user divides their network into fragments to verify independently, and annotates each fragment with an interface. These interfaces summarize each fragment’s routing announcements (a.k.a. routes). We prove that if each fragment guarantees its interface, assuming the interfaces of its neighbors, then properties of the fragments’ routes hold of the monolithic network’s. We present Kirigami, a formal model of network fragments where users specify interfaces as cuts dividing the network into fragments, annotating fragment boundaries with nodes’ stable routes. We define a Satisfiability Modulo Theories (SMT)-based procedure for checking interfaces, implemented as an extension to the NV (monolithic) verification tool. Kirigami’s SMT checks are up to five orders of magnitude faster than NV’s for a range of industrial topologies with synthesized network policies. Unfortunately, Kirigami’s interfaces are not change-resilient if network updates change fragments’ stable routes. However, if we naïvely extend Kirigami’s interfaces to use over-approximate sets of routes instead of exact routes, the resulting verification procedure allows unsound circular reasoning. To prevent circularity, we introduce a new model, Timepiece, where interfaces are defined using temporal invariants inspired by temporal logic. We implement a new SMT-based checking procedure, which scales to verify networks with thousands of nodes in minutes (compared to hours for a monolithic baseline) and allows users to write change-resilient interfaces.

1 1

Naorin Hossain will present her FPO "Navigating Emerging Complexities of Modern Systems: Advancements in Automated Verification and Security Techniques" in CS 302 on January 11, 2024 at 9am.
by Gradinfo 11 Dec '23

11 Dec '23

Naorin Hossain will present her FPO "Navigating Emerging Complexities of Modern Systems: Advancements in Automated Verification and Security Techniques" in CS 302 on January 11, 2024 at 9am. The members of her committee are as follows: Margaret Martonosi (advisor, examiner) Aarti Gupta (reader) Pradip Bose (reader, IBM) Amit Levy (examiner) David August (examiner) Everyone is invited to attend the talk. Please see abstract below. Abstract: The increasing design complexity at the end of Moore’s Law and Dennard Scaling presents a new challenge for implementing modern systems correctly and securely. This dissertation presents focused efforts for advancing automated verification and security techniques to mitigate incorrect implementations and attackers, respectively. It explores two avenues: 1) correctness verification in virtual memory systems, where shared memory interactions occur across hardware and system-level events, and 2) enhancing security of heterogeneous systems-on-a-chip (SoCs) with anomalous activity detection and localization. In the first thrust, this dissertation develops tools for a systematic, end-to-end formalized approach for verifying and validating virtual memory system implementations using memory transistency model (MTM) specifications. First, I propose a novel formal language for reasoning about MTMs, enabling specification of the software-visible impacts of complex hardware- and system-level interactions that occur in parallel systems with virtual memory. Next, I present empirical MTM testing techniques for verification and validation on real system implementations using specialized stress measures to coax out MTM-specific bugs. The proposed techniques can alleviate complex system verification efforts by automating detection of bugs and vulnerabilities at the hardware-system interface of virtual memory systems. The second thrust of this dissertation develops processes to enhance security of heterogeneous SoC designs. Sophisticated attacks can take down parts of the SoC, resulting in diminished performance gains and failed processes. To design built-in defenses against such attacks, this dissertation proposes the use of network-on-chip (NoC) based hardware counters to monitor ongoing SoC activity in a holistic fashion. With these counters, I develop and demonstrate an anomalous activity detection and localization system to detect and pinpoint availability attacks on SoC components, leveraging machine learning techniques and the inherent interpretability offered by the NoC counters. These techniques enable automated attack and failure detection to be built into the SoC, a particularly valuable feature in the fast-changing heterogeneous SoC design space. Overall, this dissertation advocates for the use of formal and empirical modeling tools to effectively capture complex system behaviors such that vulnerabilities, bugs, and threats can be detected at design time and runtime with automated methods that provide elevated coverage and accuracy.

1 0

TODAY @ 5:30: Professor Cynthia Dwork, Harvard
by Emily C. Lawrence 07 Dec '23

07 Dec '23

REMINDER: Thursday, December 7 at 5:30pm: Princeton Public Lectures presents, Cynthia Dwork “It’s in Your Phone. It’s in Your Browser. It’s in Your Redistricting Data! … It’s Differential Privacy.” In-person event in McCosh Hall, Room 50 To participate virtually, you may either register to watch via [ https://princeton.zoom.us/webinar/register/WN_UDMs-l99RX-Czc6K3EbtdQ | Zoom Webinar ] , or Watch via Princeton’s Media Central (Live-Stream feed) by [ https://mediacentrallive.princeton.edu/ | clicking here ] . For more information about this event, [ https://lectures.princeton.edu/lectures/2023/cynthia-dwork | click here ] .

1 0

Thursday, Dec 7 @ 5:30: Professor Cynthia Dwork, Harvard
by Emily C. Lawrence 01 Dec '23

01 Dec '23

Thursday, December 7 at 5:30pm: Princeton Public Lectures presents, Cynthia Dwork “It’s in Your Phone. It’s in Your Browser. It’s in Your Redistricting Data! … It’s Differential Privacy.” In-person event in McCosh Hall, Room 50 To participate virtually, you may either register to watch via [ https://princeton.zoom.us/webinar/register/WN_UDMs-l99RX-Czc6K3EbtdQ | Zoom Webinar ] , or Watch via Princeton’s Media Central (Live-Stream feed) by [ https://mediacentrallive.princeton.edu/ | clicking here ] . For more information about this event, [ https://lectures.princeton.edu/lectures/2023/cynthia-dwork | click here ] .

1 0

DATE CORRECTION: ML Theory Seminar on neural network approximation by Prof. Song Mei from UC Berkeley
by Emily C. Lawrence 30 Nov '23

30 Nov '23

ML Theory Seminar Prof. Song Mei from UC Berkeley Statistics is giving a seminar talk on neural network approximation theory in the modern age. The talk will be on Friday (Dec 8th) 12:30 pm -1:30 pm at Equad B205. =============================================================== Title: Revisiting neural network approximation theory in the age of generative AI Abstract: Textbooks on deep learning theory primarily perceive neural networks as universal function approximators. While this classical viewpoint is fundamental, it inadequately explains the impressive capabilities of modern generative AI models such as language models and diffusion models. This talk puts forth a refined perspective: neural networks often serve as algorithm approximators, going beyond mere function approximation. I will explain how this refined perspective offers a deeper insight into the success of modern generative AI models. Bio: Song Mei is an assistant professor of statistics and EECS at UC Berkeley. He received his Ph.D. from Stanford in June 2020. Song’s research lies at the intersection of statistics and machine learning. His recent research focuses on mathematics of deep learning and foundation models, theory of reinforcement learning, and high dimensional statistics.

1 0

ML Theory Seminar on neural network approximation by Prof. Song Mei from UC Berkeley
by Emily C. Lawrence 29 Nov '23

29 Nov '23

ML Theory Seminar Prof. Song Mei from UC Berkeley Statistics is giving a seminar talk on neural network approximation theory in the modern age. The talk will be on Friday (Dec 1st) 12:30 pm -1:30 pm at Equad B205. =============================================================== Title: Revisiting neural network approximation theory in the age of generative AI Abstract: Textbooks on deep learning theory primarily perceive neural networks as universal function approximators. While this classical viewpoint is fundamental, it inadequately explains the impressive capabilities of modern generative AI models such as language models and diffusion models. This talk puts forth a refined perspective: neural networks often serve as algorithm approximators, going beyond mere function approximation. I will explain how this refined perspective offers a deeper insight into the success of modern generative AI models. Bio: Song Mei is an assistant professor of statistics and EECS at UC Berkeley. He received his Ph.D. from Stanford in June 2020. Song’s research lies at the intersection of statistics and machine learning. His recent research focuses on mathematics of deep learning and foundation models, theory of reinforcement learning, and high dimensional statistics.

1 0

TODAY: CSML SEMINAR and ECE KORHAMMER SEMINAR SERIES: Haifeng Xu, University of Chicago
by Emily C. Lawrence 29 Nov '23

29 Nov '23

CSML SEMINAR ECE KORHAMMER SEMINAR SERIES Speaker: Haifeng Xu, University of Chicago Title: Algorithms and Incentives in Machine Learning Day: Wednesday, November 29, 2023 Time: 4:30 PM Location: B205 Engineering Quadrangle Hosts: Chi Jin and Mengdi Wang Abstract: Machine learning designs approaches that transform data to predictions or estimations. The standard paradigm often assumes these data are objectively generated from distributions, without being affected by any human factors. However, this paradigm ceases to be true when our predictions or estimated parameters will in turn affect the data providers' welfare. In such situations, data providers have incentives to alter the data for their own benefits. Thus the design of any learning methods must account for potential data manipulations due to data providers' incentives. This talk will introduce a general ""incentive-aware"" framework for designing prediction methods. I will illustrate this design paradigm with two examples: (1) a very recent and timely application of eliciting authors' truthful private information for improving the peer review systems for today's massive scale machine learning conferences; (2) a very classic problem of PAC-learning classifiers but with strategic providers of data features. In both problems, I will illustrate how the presence of incentives can fundamentally change the problem's learning efficiency and how algorithms can help to overcome some statistic barriers. Bio: Haifeng Xu is an assistant professor in the Department of Computer Science and Data Science Institute at UChicago. He directs the Sigma Lab (Strategic IntelliGence for Machine Agents) which focuses on designing intelligent AI systems that can effectively learn and act in informationally complex multi-agent setups. His research has been recognized by a few awards including IJCAI early career spotlight, Google Faculty research award, the ACM SIGecom Dissertation Award and IFAAMAS Victor Lesser Distinguished Dissertation Award. This seminar is supported with funds from the Korhammer Lecture Series.

1 0

Dr. Fei-Fei Li in Conversation with Provost Jennifer Rexford Wednesday, Nov 29 at 6pm
by Emily C. Lawrence 29 Nov '23

29 Nov '23

Wednesday, November 29 at 6pm Princeton Public Lectures presents, Dr. Fei-Fei Li in Conversation with Jennifer Rexford in-person event in McCosh 50 Free copies of Dr. Li's book, The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI, will be handed out to the first 300 in-person attendees. To participate virtually, you may either register to watch via [ https://princeton.zoom.us/webinar/register/WN_L58PDluYR_K6huylyTPiPg | Zoom Webinar ] , or Watch via Princeton’s Media Central (Live-Stream feed) by [ https://mediacentrallive.princeton.edu/ | clicking here ] . For more information about this event, [ https://lectures.princeton.edu/lectures/2023/dr-fei-fei-li | click here ] .

1 1