January 2024 - talks - lists.cs.princeton.edu

CS Distinguished Colloquium Speaker John Jumper: Monday, September 25 at 4:30
by Emily C. Lawrence 12 Feb '24

12 Feb '24

Speaker: John Jumper, DeepMind Date: Monday, September 25 Time: 4:30pm to 5:30pm Location: Friend Center 101 Host: Ellen Zhong Event page: [ https://www.cs.princeton.edu/events/26500 | https://www.cs.princeton.edu/events/26500 ] Title: Highly accurate protein structure prediction with deep learning Abstract: Our work on deep learning for biology, specifically the AlphaFold system, has demonstrated that neural networks are capable of highly accurate modeling of both protein structure and protein-protein interactions. In particular, the system shows a remarkable ability to extract chemical and evolutionary principles from experimental structural data. This computational tool has repeatedly shown the ability to not only predict accurate structures for novel sequences and novel folds but also to do unexpected tasks such as selecting stable protein designs or detecting protein disorder. In this lecture, I will discuss the context of this breakthrough in the machine learning principles, the diverse data and rigorous evaluation environment that enabled it to occur, and the many innovative ways in which the community is using these tools to do new types of science. I will also reflect on some surprising limitations -- insensitivity to mutations and the lack of context about the chemical environment of the proteins -- and how this may be traced back to the essential features of the training process. Finally, I will conclude with a discussion of some ideas on the future of machine learning in structure biology and how the experimental and computational communities can think about organizing their research and data to enable many more such breakthroughs in the future. Bio: John Jumper received his PhD in Chemistry from the University of Chicago, where he developed machine learning methods to simulate protein dynamics. Prior to that, he worked at D.E. Shaw Research on molecular dynamics simulations of protein dynamics and supercooled liquids. He also holds an MPhil in Physics from the University of Cambridge and a B.S. in Physics and Mathematics from Vanderbilt University. At DeepMind, John is leading the development of new methods to apply machine learning to protein biology.

1 13

MAE Seminar Series: Cynthia Sung, Friday 2/2
by Emily C. Lawrence 02 Feb '24

02 Feb '24

Speaker: [ https://sung.seas.upenn.edu/people/sung/ | Cynthia Sung ] , UPenn Date: Friday, February 2nd, 2024 Time: 12:30 PM ET Seminar: [ https://mae.princeton.edu/about-mae/events/computational-design-origami-and… | MAE Seminar Series ] Location: Bowen Hall Room 222 Title: Computational Design of Origami and Compliant Robots Abstract: Soft and compliant robots provide new opportunities for machines that are flexible, adaptable, safe, and robust. Origami-inspired engineering enables custom robots to be designed and fabricated within days, or even hours. These robots are capable of executing a variety of shape-changing and dynamical tasks by taking advantage of their folded shape and programmable mechanics. In this talk, I will show how the kinematics and compliance of a thin-shell mechanism can be designed algorithmically. I will also discuss how we can leverage these designs for better performance and simpler control, and how these approaches provide robots with capabilities such as dynamical hopping, swimming, and flight. Bio: Cynthia Sung is the Gabel Family Term Assistant Professor in the Department of Mechanical Engineering and Applied Mechanics (MEAM) and a member of the General Robotics, Automation, Sensing & Perception (GRASP) lab at the University of Pennsylvania. She received a Ph.D. in Electrical Engineering and Computer Science from MIT in 2016 and a B.S. in Mechanical Engineering from Rice University in 2011. Her research interests are computational methods for design automation of robotic systems, with a particular focus on origami-inspired and compliant robots. She is the recipient of a 2023 ONR Young Investigator award, a 2020 Johnson & Johnson Women in STEM2D Scholars Award, and a 2019 NSF CAREER award.

1 1

ORFE Distinguished Lecture Series: Benjamin Recht tomorrow at 4:30pm
by Emily C. Lawrence 29 Jan '24

29 Jan '24

=== 1/30/24 ORFE Distinguished Lecture Series=== DATE: Tuesday, January 30, 2024 TIME: 4:30pm LOCATION: Sherrerd 101 SPEAKER: Benjamin Recht, UC Berkeley TITLE: Statistics When n Equals 1 Abstract: From large trials of treatments and diseases, 21st-century medicine embraces inferences about people from studies of populations. But such population inferences tell us little about what to do with any particular person. In this talk, I will first describe some of the drawbacks of applying population statistics to decision-making about individuals. As an alternative, I will outline how we might design treatments and interventions to help those individuals directly. I will present a series of parallel projects that link ideas from systems, optimization, control, and experiment design to draw inferences and inform such individualized decision making. Bio: Benjamin Recht is a Professor in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley. His research has focused on applying mathematical optimization and statistics to problems in data analysis and machine learning. He is currently studying histories, methods, and theories of scientific validity and experimental design.

1 0

Xindi Wu will present her General Exam "Data-Centric Multimodal Machine Learning" on Wednesday, January 31, 2024 at 3:00 PM in CS 402 and via zoom.
by CS Grad Department 26 Jan '24

26 Jan '24

Xindi Wu will present her General Exam "Data-Centric Multimodal Machine Learning" on Wednesday, January 31, 2024 at 3:00 PM in CS 402 and via zoom. Zoom link: https://princeton.zoom.us/j/95305696158 Committee Members: Olga Russakovsky (advisor), Szymon Rusinkiewicz, Adji Bousso Dieng Abstract: Data is a cornerstone of innovation in multimodal machine learning. With the massive amount of vision-language data available, figuring out how to effectively use it becomes a major challenge. Dataset distillation methods aim to scale down large-scale datasets to significantly smaller sets of synthetic training examples while preserving sufficient information for model training. Current dataset distillation methods have shown success in condensing large-scale image datasets into smaller, yet effective, training examples for image classification. However, with the rise in capabilities of vision-language models (VLMs), and especially given the scale of datasets necessary to train these models, the time is ripe to expand dataset distillation methods beyond image classification. This task is challenging, given the nature of vision-language datasets, which typically lack discrete classes and involve complex image-text relationships. Our research marks a pioneering step in this direction. We take the first steps towards this goal by expanding the idea of trajectory matching to create a distillation method for vision-language datasets. Our method significantly reduces the required dataset size while maintaining the model's performance. We hope that our work lays the groundwork for future research aimed at understanding what is the minimum information required for a vision-language model to achieve comparable performance quickly, thereby building a better understanding of the compositionality of compact visual-linguistic knowledge. Reading List: https://docs.google.com/document/d/1A5ShHxlxNgs_E4r5iIs4Y1D1L-nUfp9Z68Vls6m… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Arushi Gupta will present her FPO "Understanding the Role of Data in Model Decisions" on Thursday, February 8, 2024 at 1:30 PM in CS 402.
by CS Grad Department 25 Jan '24

25 Jan '24

Arushi Gupta will present her FPO "Understanding the Role of Data in Model Decisions" on Thursday, February 8, 2024 at 1:30 PM in CS 402. Location: CS 402 The members of Arushi’s committee are as follows: Examiners: Sanjeev Arora (Adviser), Elad Hazan, Ryan Adams Readers: Karthik Narasimhan, Tom Griffiths A copy of her thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of thethesis. Everyone is invited to attend her talk. Abstract follows below: As neural networks are increasingly employed in high stakes applications such as criminal justice, medicine, etc, [1] it becomes increasingly important to understand why these models make the decisions they do. For example, it is important to develop tools to analyze whether models are perpetuating harmful demographic inequalities they have found in their training data in their future decision making [2]. However, neural networks typically require large training sets, have “black-box” decision making, and have costly retraining protocols, increasing the difficulty of this problem. This work considers three questions. Q1) What is the relationship between the elements of an input and the models’ decision? Q2) What is the relationship between the individual training points and the model’s decision. And finally Q3) to what extent do there exist (efficient) approximations that would allow practitioners to predict how model performance would change given different training data, or a different training protocol. Part I addresses Q1 for masking saliency methods. These methods implicitly assume that grey pixels in an image are “uninformative.” We find experimentally that this assumption may not always be true, and define “soundness,” which measures a desirable property of a saliency map. Part II addresses Q2 and Q3 in the context of influence functions, which aim to approximate the effect of removing a training points on the model’s decision. We use harmonic analysis to examine a particular type of influence method, namely datamodels, and find that there is a relationship between the coefficients of the datamodel, and the Fourier coefficients of the target function. Finally, Part III addresses Q3 in the context of test data. First, we assess whether held out test data is necessary to approximate the outer loop of meta learning, or whether recycling training data constitutes a sufficient approximation. We find that held out test data is important, as it learns representations that are low rank. Then, inspired by the PGDL competition [3] we investigate whether GAN generated data, despite well known limitations, can be used to approximate generalization performance when no test or validation set is available, and find that they can.

1 0

Yuan Wang will present her FPO "Systematic analysis of cellular and immune system responses for therapy development" on Monday, February 5, 2024 at 3:00 PM in Carl C. Icahn Laboratory - 200 and Zoom.
by CS Grad Department 24 Jan '24

24 Jan '24

Yuan Wang will present her FPO "Systematic analysis of cellular and immune system responses for therapy development" on Monday, February 5, 2024 at 3:00 PM in Carl C. Icahn Laboratory - 200 and Zoom. Location: Zoom link: https://princeton.zoom.us/j/3225545420?omn=94214228673 The members of Yuan’s committee are as follows: Examiners: Olga Troyanskaya (Adviser), Mona Singh, Ellen Zhong Readers: Yuri Pritykin, Wendell Lim (UCSF) Everyone is invited to attend her talk. Abstract follows below: The human body is a complex program where many important regulatory molecules are reused throughout the body, but their functions are specific to particular locations and biological contexts. This notion of bio-specificity is especially valuable in the context of immunology, which can be harnessed to develop effective diagnostics and therapeutics for human diseases. However, given the extensive number of potential candidate molecules, experimental analysis of the functions of all target genes is not feasible. Moreover, integrative interpretation of public omics datasets remains challenging due to substantial noise and heterogeneity present in the data. Systematically harmonizing and mapping transcriptional regulation and gene expression is thus of critical importance. In this thesis we developed and applied two computational frameworks, SPEEDI and TissueGPS, which harness large-scale heterogeneous omics datasets that systematically integrate and analyze epigenetic and transcriptional omics data to understand immune system dynamics. We applied these frameworks in the context of cellular and systemic immunity. First, we systematically study the systemic immune transcriptional responses of vaccinated subjects to SARS-CoV-2 infections. Second, by investigating the immune landscape of infected individuals, we predict epigenetic-informed gene signatures to assist rapid diagnosis for infectious disease assessments. Third, we designed cell-based therapeutics to arm the immune system with customized sensing circuits to detect and treat central nervous system disorders, such as malignant tumors, with reduced systemic off-target toxicity. Collectively, our methods offer novel frameworks for interrogating human diseases with precision and reproducibility across immune contexts.

1 0

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