- talks - lists.cs.princeton.edu

Princeton Robotics Seminar: Elad Hazan, Friday Feb 9
by Emily C. Lawrence 06 Feb '24

06 Feb '24

Speaker: [ https://www.ehazan.com/ | Elad Hazan ] , Princeton Date: Friday, February 9, 2024 Time: 11am ET Location: Computer Science 105 Event page: [ https://robo.princeton.edu/seminar | https://robo.princeton.edu/seminar ] Title: The theory of online control and its application to robotics Abstract: In this talk we will discuss an emerging paradigm in differentiable reinforcement learning called “online nonstochastic control”. The new approach applies techniques from online convex optimization and convex relaxations to obtain new methods with provable guarantees for classical settings in optimal and robust control. Time permitting we will discuss recent extensions to nonlinear adaptive control and iterative planning, as well as model free reinforcement learning. This theory was, and continues to be, developed here in Princeton with numerous collaborators, including Naman Agarwal, Brian Bullins, Karan Singh, Max Simchowitz, Xinyi Chen, Ani Majumdar, Sham Kakade, Udaya Ghai, Edgar Minasyan, Paula Gradu, and many others. Bio: Elad Hazan is a professor of computer science at Princeton University. His research focuses on the design and analysis of algorithms for basic problems in machine learning and optimization. Amongst his contributions are the co-invention of the AdaGrad algorithm for deep learning, and the first sublinear-time algorithms for convex optimization. He is the recipient of the Bell Labs prize, the IBM Goldberg best paper award twice, in 2012 and 2008, a European Research Council grant, a Marie Curie fellowship and twice the Google Research Award. He served on the steering committee of the Association for Computational Learning and has been program chair for COLT 2015. In 2017 he co-founded In8 inc. focusing on efficient optimization and control, acquired by Google in 2018. He is the co-founder and director of Google AI Princeton.

1 0

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: Yao Xie, Feb 6 at 4:30pm
by Emily C. Lawrence 02 Feb '24

02 Feb '24

=== 2/6/24 ORFE Distinguished Lecture Series=== DATE: Tuesday, February 6, 2024 TIME: 4:30pm LOCATION: Sherrerd 101 SPEAKER: Yao Xie, Georgia Tech TITLE: Generative models for statistical inference Abstract: We consider the problem of learning a continuous probability density function from data, a fundamental problem in statistics known as density estimation. It also arises in distributionally robust optimization (DRO), where the goal is to find the worst-case distribution to represent scenario departure from observations. Such a problem is known to be hard in high dimensions and incurs a significant computational challenge. In this talk, I will present a new approach to tackle these challenges, leveraging recent advances in neural-networks-based generative models, a type of machine learning model that aims to generate new samples following the underlying distributions of the observations. We develop a new flow-based generative model with guarantees, inspired by the celebrated Jordan-Kinderleherer-Otto (JKO) scheme, to represent the data distribution using a composition of a sequence of optimal transport maps applying to a source distribution (such as Gaussian). Our method can greatly reduce computational costs when achieving competitive performance over existing generative models. The connection of our JKO-flow method with proximal gradient descent in the Wasserstein-2 space enables us to prove a density learning guarantee with an exponential convergence rate. Besides density estimation, we also demonstrate that the JKO-flow generative model can be used in various applications, including adversarial learning, robust hypothesis testing, and data-driven differential privacy. Bio: Yao Xie is the Coca-Cola Foundation Chair, Professor at Georgia Institute of Technology in the H. Milton Stewart School of Industrial and Systems Engineering, and Associate Director of the Machine Learning Center. She received her Ph.D. in Electrical Engineering (minor in Mathematics) from Stanford University in 2012 and was a Research Scientist at Duke University. Her research lies at the intersection of statistics, machine learning, and optimization in providing theoretical guarantees and developing computationally efficient and statistically powerful methods for problems motivated by real-world applications. She received the National Science Foundation (NSF) CAREER Award in 2017, the INFORMS Wagner Prize Finalist in 2021, and the INFORMS Gaver Early Career Award for Excellence in Operations Research in 2022. She is currently an Associate Editor for IEEE Transactions on Information Theory, Journal of the American Statistical Association-Theory and Methods, Operations Research, Sequential Analysis: Design Methods and Applications, INFORMS Journal on Data Science, and an Area Chair of NeurIPS, ICML, and ICLR.

1 0

Uthsav Chitra will present his FPO "Algorithms for understanding the spatial and network organization of biological systems" on Friday, March 1, 2024 at 10:00 AM in COS 302 and Zoom.
by CS Grad Department 01 Feb '24

01 Feb '24

Uthsav Chitra will present his FPO "Algorithms for understanding the spatial and network organization of biological systems" on Friday, March 1, 2024 at 10:00 AM in COS 302 and Zoom. Location: Zoom link: https://princeton.zoom.us/j/99220301104 The members of Uthsav’s committee are as follows: Examiners: Ben Raphael (Adviser), Bernard Chazelle, Yuri Pritykin Readers: Ellen Zhong, Fei Chen (Harvard) Everyone is invited to attend his talk. Abstract follows below: Biological systems are characterized by their spatial organization and network interactions at a hierarchy of scales. For example, the spatial arrangement of different cells in a tissue underlies fundamental multicellular processes such as tissue differentiation and disease response, while interactions between genes/proteins comprise the biological pathways that regulate cellular state and function. Recent developments in high-throughput sequencing have enabled the systematic analysis of spatial and network processes in many complex biological systems including the brain and tumor microenvironment. However, such analyses are challenged by high levels of sparsity and/or noise in high-throughput sequencing datasets—underscoring the need for principled and rigorous computational methods for biological data analysis. In this dissertation, we present a collection of mathematical frameworks and machine learning algorithms for modeling the spatial and network organization of biological systems. First, we derive a model of discrete and continuous spatial variation in gene expression. We present two algorithms, Belayer and GASTON, which learn the parameters of this model using complex analysis and interpretable deep learning, respectively. Second, we present a mathematical framework for the identification of altered subnetworks, or subnetworks of a biological interaction network containing genes/proteins that are differentially expressed, highly mutated, or otherwise aberrant compared to other genes/proteins. We prove that many existing algorithms are statistically biased, resolving the open question of why these algorithms often identify very large subnetworks that are difficult to interpret. We derive two altered subnetwork identification algorithms, NetMix and NetMix2, which we show are asymptotically unbiased and outperform existing approaches in practice. Finally, we present two frameworks for learning and modeling higher-order interactions. We first derive a statistical framework for learning higher-order genetic interactions from experimental fitness data, unifying decades of existing work in the genetics literature. Then, we derive a theoretical framework for modeling random walks on hypergraphs that provably utilizes higher-order interactions in data, in contrast to many existing hypergraph methods which only utilize pairwise interactions. Taken together, the approaches in this dissertation provide a theoretical and practical foundation for overcoming the computational challenges of modeling complex biological systems.

1 0

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