- talks - lists.cs.princeton.edu

Computational Math IDeAS seminar: Thursday, March 26 at 2pm
by Emily C. Lawrence 24 Mar '26

24 Mar '26

COMPUTATIONAL MATH – IDeAS SEMINAR Recurring weekly series · Thursdays, 2:00 – 3:00 PM This week's talk: Speaker: Leticia Mattos Da Silva (MIT) Title: Leveraging Latent Mathematical Structure in Geometry Processing Date: Thursday, March 26, 2026 Time: 2:00 PM – 3:00 PM Room: 224 Fine Hall Abstract: Myriad problems in geometry processing are readily formulated through continuous mathematics, including variational energies, stochastic dynamics, geometric constraints, and partial differential equations. Often, the challenge in solving these formulations in practice is not the absence of a principled theoretical formulation, but that standard computational approaches are limited by design constraints or technical barriers, resulting in instability or poor scalability. Common numerical techniques resort to fragile nonlinear solvers, repeated large optimizations, or forgo analytical or physical properties of the continuous problem. In this talk, I will present methods that tackle these limitations by recognizing and exploiting latent problem structure, including convexity, variational reformulations, and symmetry. After a prelude illustrating the value of this approach in two problems of interest in geometry---simulating elastodynamics and solving PDEs on surfaces---I will describe a new approach to resampling that leverages a time-symmetric variant of the Schrödinger bridge problem. In each of these works, we discuss how algorithmic design, when carefully guided by underlying mathematical structure, can yield numerical tools that offer practitioners greater control and flexibility. About the speaker: Leticia is a fifth-year Ph.D. student at MIT, where she works in the Geometric Data Processing Group under the supervision of Justin Solomon. Her research involves developing numerical algorithms to solve nonlinear PDEs and other dynamical problems that appear in geometry processing. She is the recipient of two MathWorks Fellowships, as well as a Google Fellowship. Previously, she received a B.S. in Mathematics from the University of California, Los Angeles, and in summers past, she has conducted research as part of internships at the Flatiron Institute, at Adobe, Inc. and at the Fields Institute. Up next: Thursday, April 2 – Florian Schaefer (NYU).

1 0

ECE Seminar: Dylan Foster, Microsoft Research TODAY at 4:30pm
by Emily C. Lawrence 23 Mar '26

23 Mar '26

ECE SEMINAR Speaker: Dylan Foster, Microsoft Research Title: Understanding the foundation model pipeline: From imitation to exploration Day: Monday, March 23, 2026 Time: 4:30 PM Location: B205 Engineering Quadrangle Host: Guillermo Sapiro Abstract: The prevailing recipe of ever-larger models trained on passively collected data is showing diminishing returns and facing growing constraints on data. To sustain progress, the next phase of AI will hinge on experience: agents and systems that actively curate their own data through interaction. This talk asks: which principles behind today's foundation model pipeline will carry us there, and where do we need new ones? I. Rethinking the mechanisms behind next-token prediction. Pre-training with next-token prediction is the basis for language modeling, but will it hit a wall as we scale to longer horizons? We will argue that the answer is no, but for reasons that overturn conventional wisdom and point to new interventions. II. Quantifying what makes a good foundation model. Post-training uses interaction to surface capabilities from the pre-trained model, but when does this amplify existing behaviors versus discover new ones? We introduce the coverage profile, a simple property of the model that characterizes post-training success and failure, and that standard pre-training secretly optimizes for. III. Revisiting reinforcement learning for the foundation model era. Classical RL offers powerful mechanisms for going beyond what the model already knows, but leveraging them in the foundation model era requires new ideas. We will show that outcome-based methods are limited by coverage, and that overcoming this requires rethinking how models explore. Bio: Dylan Foster is a principal researcher at Microsoft Research, New England and New York City. Previously, he was a postdoctoral fellow at MIT IDSS with Sasha Rakhlin, and received his PhD from the Department of Computer Science at Cornell University, advised by Karthik Sridharan. His research develops principles and algorithms for learning from interaction, including reinforcement learning, foundation model training, and interactive decision making. His work has received several awards, including best paper at COLT (2019), best student paper at COLT (2018, 2019), and the Cornell CS doctoral dissertation award.

1 0

Sata Sengupta will present his FPO "Network-Boosted Applications" on Friday, April 3, 2026 at 1pm in CS 402.
by Gradinfo 19 Mar '26

19 Mar '26

Sata Sengupta will present his FPO "Network-Boosted Applications" on Friday, April 3, 2026 at 1pm in CS 402. The members of his committee are as follows: Examiners: Jennifer Rexford (adviser), David Walker, Maria Apostolaki (ECE) Readers: Ravi Netravali, Hyojoon Kim (University of Virginia, CS) Zoom link: [ https://princeton.zoom.us/j/2332835952?pwd=tYVB2ta8EI4aC3eX4biZCbu4VyYK0J.1 | https://princeton.zoom.us/j/2332835952?pwd=tYVB2ta8EI4aC3eX4biZCbu4VyYK0J.1 ] Abstract follows below: Modern Internet applications such as video conferencing, cloud gaming, and AR/VR demand high media quality, low latency, and massive scale. Yet networks still operate largely as best-effort conduits with limited visibility into application behavior or quality of experience (QoE), leaving operators unable to react in time to short-lived congestion or stealthy attacks that disproportionately affect application performance or security. Addressing these problems requires packet-level measurement and real-time control in the network itself, where traffic can be observed and acted upon before users feel the impact. High-speed programmable switches and Smart Network Interface Cards (SmartNICs) offer a path to "boosting" applications at up to terabit rates, but only with careful designs that operate within stringent limits on memory and computation, while handling sophisticated, stateful protocols such as Transmission Control Protocol (TCP) for on-demand traffic (for example, web browsing and video streaming) and Real-Time Communication (RTC) for real-time interactive traffic (for example, video conferencing and online gaming). This dissertation develops efficient, hardware-amenable techniques for real-time network monitoring and control across both traffic classes. For on-demand traffic, we present DART (Data-plane Actionable Round-trip Times), which measures TCP round-trip time (RTT) at line rate by carefully handling retransmissions, reordering, and selective acknowledgments---TCP behaviors that can otherwise distort RTT samples or exhaust limited switch memory. Building on this capability, we develop HiDe (Hijack Defense) to promptly detect and mitigate stealthy long-distance routing attacks using passively monitored RTT variation as the detection signal. For interactive traffic, we analyze representative video-conferencing systems, including Zoom (proprietary) and MediaSoup (open-source), and derive detailed packet-level performance metrics. For Zoom, we show that despite encryption and a proprietary RTC protocol, packet headers retain enough information to recover per-media-stream quality indicators such as bitrate, frame rate, and jitter. These insights motivate Scallop, a scalable hardware-software co-designed architecture for video-conferencing infrastructure inspired by Software-Defined Networking (SDN). In Scallop, a high-speed data plane executes high-frequency, latency-sensitive operations, while a lean software control plane handles infrequent tasks, improving performance and scalability, and reducing cost relative to conventional software solutions. Overall, this dissertation advances scalable, application-aware network monitoring and control techniques that improve the performance and security of the Internet’s two most prevalent application classes. Our research artifacts, including hardware prototypes and traffic-analysis tools, are already being used by both academia and industry for further study and development.

1 0

TODAY: ECE Seminar Dhruv Rohatgi, MIT on Wednesday, March 18, 2026
by Emily C. Lawrence 18 Mar '26

18 Mar '26

ECE SEMINAR Speaker: Dhruv Rohatgi, Massachusetts Institute of Technology Title: Designing Principled ML Algorithms via Modularity Day: Wednesday, March 18, 2026 Time: 4:30 PM Location: B205 Engineering Quadrangle Host: Pramod Viswanath Abstract: Designing the right algorithm for a problem requires modelling the problem well. But modern machine learning increasingly tackles problems that appear far too complex to mathematically model: empirical successes are driven by elusive properties of natural data, as well as significant engineering expertise. Can we still develop theory to help illuminate the space of possible algorithms for these problems? In this talk, I will argue the benefits of a modular approach to theory-building, where instead of defining "natural data" explicitly, we define what machine learning can already do with it. More precisely, we take existing building blocks of machine learning -- classifiers, regressors, even generative models -- as given, and we ask: how should we use these building blocks to provably solve more complex problems? I will apply this lens at several stages of the language model pipeline -- fine-tuning, reinforcement learning, and generation -- to prototype principled new algorithms and illuminate fundamental limits. Throughout, a key theme will be understanding when and how the "curse of horizon" -- a folklore principle throughout machine learning, that long decision-making horizons tend to amplify learning errors -- can be algorithmically mitigated. Bio: Dhruv Rohatgi is a final-year PhD student at the Massachusetts Institute of Technology, advised by Ankur Moitra. His research focuses on understanding how fundamental building blocks of machine learning can be fit together to solve more complex problems, particularly those that involve reliably making sequences of decisions. His PhD has been supported by an Akamai Presidential Fellowship and an NDSEG Fellowship.

1 0

Computational Math – IDeAS Seminar Thursdays at 2pm
by Emily C. Lawrence 17 Mar '26

17 Mar '26

COMPUTATIONAL MATH – IDeAS SEMINAR Recurring weekly series · Thursdays, 2:00 – 3:00 PM, 224 Fine Hall This week's talk: Speaker: Diana Halikias (NYU) Title: Operator learning without the adjoint. Date: Thursday, March 19, 2026 Time: 2:00 PM – 3:00 PM Room: 224 Fine Hall Abstract: There is a mystery at the heart of operator learning: how can one recover a non-self-adjoint operator from data without probing the adjoint? Current practical approaches suggest that one can accurately recover an operator while only using data generated by the forward action of the operator without access to the adjoint. However, naively, it seems essential to sample the action of the adjoint. We partially explain this mystery by proving that without querying the adjoint, one can approximate a family of non-self-adjoint infinite-dimensional compact operators via projection onto a Fourier basis. We then apply the result to recovering Green's functions of elliptic partial differential operators and derive an adjoint-free sample complexity bound. While existing theory justifies low sample complexity in operator learning, ours is the first adjoint-free analysis that attempts to close the gap between theory and practice. We also explore a closely related question in numerical linear algebra: when is access to both forward and transpose matrix-vector products essential? We discuss the role of transpose access in sketching algorithms for low-rank approximation, least-squares problems, and norm estimation. About the speaker: Diana Halikias is a Courant Instructor in the mathematics department at NYU. Previously, she was a postdoctoral fellow for the semester-long program in Complexity and Linear Algebra at the Simons Institute in Berkeley. She completed a PhD in mathematics in 2025 at Cornell, where she was advised by Alex Townsend. Up next: Thursday, March 26 – Leticia Mattos Da Silva (MIT).

1 0

CS Colloquium Speaker: Stefano Soatto, Tues. Feb 24 at 3:20pm
by Emily C. Lawrence 17 Mar '26

17 Mar '26

CS Colloquium Speaker: Stefano Soatto from UCLA & Amazon Date & Time: Tuesday, February 24, 2026 - 3:20pm Location: CS 105 Host: Jia Deng Webinar registration: [ https://princeton.zoom.us/webinar/register/WN_geQ0rhfcQR-fwkdQGAo7Ug | https://princeton.zoom.us/webinar/register/WN_geQ0rhfcQR-fwkdQGAo7Ug ] Title: AI Agents as Universal Task Solvers: Time, Information, and Intelligence in LLMs viewed as Maximalistic Models of Computation Abstract: Scaling laws predict that AI agents will steadily improve and eventually exceed human performance across a wide range of tasks. Yet at the limit of these scaling laws lies a form of inference that involves no intelligence at all: with increasing compute and memory, a model can brute-force any verifiable task without learning anything from past experience. Universally optimal inference, pioneered by Solomonoff and Levin, requires no insight — only exhaustive search. This raises a basic question: if scaling alone does not foster intelligence, what does? And if performance on downstream tasks is insufficient to measure intelligence, what is? In this talk, I will point to the critical role of time in both analyzing and fostering the emergent reasoning behavior of AI agents. Building on insights that Solomonoff sketched in 1985 but that remained theoretical curiosities for decades, I will show that the value of learning is measured not by a reduction in uncertainty — the core of inductive learning and generalization — but by a reduction in the time needed to solve new tasks. A key result is that data can make a universal solver exponentially faster, with the speed-up tightly characterized by a single quantity: the algorithmic mutual information between past experience and the solution to unforeseen tasks. Connecting these ideas to modern AI requires rethinking what computation means for systems powered by large language models. Unlike minimalistic models of computation such as Turing Machines, LLMs are stochastic dynamical systems whose computational elements — context, weights, activations, chain-of-thought — do not resemble a "program" in the ordinary sense. I will show that LLMs are instead maximalistic models of computation: universal, like Turing Machines, but operating through entirely different and in many ways antithetical mechanisms. Programming such systems can be achieved through two-level control strategies — open-loop planning and closed-loop feedback — in abstract space, a framework we have recently released Strands Agents open-source library (www.strandsagents.com). Once time is properly accounted for, scaling laws reveal an inversion : beyond a critical point, increasing resources improve benchmark accuracy while diminishing conceptual depth— a savant regime in which models improve while learning less. I will discuss what this means for how we build, evaluate, and scale AI agents. Bio: Stefano Soatto is a Vice President at AWS Agentic AI, and a Professor of Computer Science at UCLA. He received his PhD in Control and Dynamical Systems from the California Institute of Technology, his D.Ing from the University of Padova, Italy and was a postdoctoral scholar at Harvard University. He is a Fellow of the ACM and of the IEEE.

1 4

CoE Seminar - Prof, Justin Notestein, 3/19 at 10:00 am in Maeder Hall
by Emily C. Lawrence 12 Mar '26

12 Mar '26

CoE Seminar Thursday, March 19 at 10:00am in Maeder Hall Auditorium Justin Notestein, Northwestern University Title: Comparisons between MOF-supported and oxide-supported catalysts Abstract: The Notestein group studies catalytic transformations with the goal of improving the production of large-scale chemicals and fuels. Most of our research examines supported catalysts, and it is in this light that we carry out much of our work on metal organic frameworks. Setting aside the question of industrial-scale relevance, we primarily ask the question, how does supporting an active site on a MOF node – or as a MOF node – compare to the ‘analogous’ structure on other supports or free in solution? The talk will chart an evolution of our work in the field, and we will pause to discuss a few specific examples, including the use of Zr-based MOFs as catalysts for the MPV reaction and supported complexes and clusters for challenging reactions like dehydrogenation or the isomerization of aromatics. Through these examples, and by analogy to some of our work with zeolite-supported catalysts, we will try to identify areas where loading into these porous solids presents opportunities for improved catalytic function or understanding.

1 0

Robotics seminar: Friday, Mar 6 at 11am
by Emily C. Lawrence 05 Mar '26

05 Mar '26

Hi everyone, Our next robotics seminar speaker is Prof. Abhishek Gupta, Assistant Professor at the University of Washington. Speaker: Abhishek Gupta Date: 03/06/2026 Time: 11:00 AM EST Location: CS 105 We are looking forward to seeing you there! Please stay tuned for more details. Title: [ https://robotics.princeton.edu/seminars | Simulation-driven Robot Learning for the Lazy Researcher ] Abstract: This talk will delve into the workflow of building robot learning methods powered by large-scale physics simulation. While the promise of simulation is low-effort data generation that scales cheaply with human effort, the reality is often not as simple, with considerable human effort required to actually make simulation a useful tool for real-world robotic manipulation. In this talk, I will study how to develop methods that explicitly reduce manual human effort in enabling simulation as a data generation tool - from environment and behavior generation, to enabling downstream transfer to the real-world. In doing so, I will outline an effort-efficient path to leveraging simulation as a data source for robot learning - hopefully to make robotics easier for lazy researchers like me! Best Regards, Robotics Seminar Committee

1 0

Computational Math – IDeAS Seminar Thursdays at 2pm
by Emily C. Lawrence 04 Mar '26

04 Mar '26

COMPUTATIONAL MATH – IDeAS SEMINAR Recurring weekly series · Thursdays, 2:00 – 3:00 PM · Jadwin 111 This week's talk: Speaker: Per-Gunnar Martinsson (The University of Texas) Title: Randomized algorithms for linear algebraic computations. Date: Thursday, March 5, 2026 Time: 2:00 PM – 3:00 PM Room: Jadwin 111 Abstract: The talk will describe how randomized algorithms can effectively, accurately, and reliably solve linear algebraic problems that are omnipresent in scientific computing and in data analysis. We will focus on techniques for low rank approximation, since these methods are particularly simple and powerful, and are well understood mathematically. The talk will also briefly survey how randomized techniques can be applied to approximate global operators that arise in scientific computing such as solution operators to elliptic PDEs, boundary-to-boundary operators such as the Dirichlet-to-Neumann map, and time evolution operators of parabolic PDEs. About the speaker: Gunnar Martinsson is a Professor of Mathematics at the University of Texas at Austin. He is also Deputy Director of the Oden Institute for Computational Engineering and Sciences, and he holds the W.A. "Tex" Moncrief, Jr. Chair in Simulation-Based Engineering and Sciences. Prior to joining UT Austin, Martinsson served as a Professor of Mathematics at the University of Oxford, and he has previously held faculty positions at the University of Colorado at Boulder and at Yale University. He earned his Ph.D. in computational and applied mathematics in 2002 from UT Austin. He received an M.Sc. degree in Engineering Physics in 1996 and a "Licentiate" degree in Mathematics in 1998, both from the Chalmers University of Technology in Sweden. Martinsson was the recipient of the SIAM 2017 Germund Dahlquist Prize, was named a Fellow of SIAM in 2021, and held a Simons Fellowship in 2025/26. Martinsson's research concerns the development of faster algorithms for ubiquitous computational tasks in scientific computing and data sciences. Recent work has focused on randomized methods in linear algebra, fast solvers for elliptic PDEs, O(N) complexity direct solvers, structured matrix computations, and high order accurate methods for scattering and fluid problems. Up next (after spring break): Thursday, March 19 – Diana Halikias.

1 0

ECE Seminar: Weijia Shi, Monday, March 2
by Emily C. Lawrence 27 Feb '26

27 Feb '26

ECE SEMINAR Speaker: Weijia Shi, University of Washington Day: Monday, March 2, 2026 Time: 4:30 PM Location: B205 Engineering Quadrangle Host: Tom Silver Title: Breaking the Language Model Monolith Abstract: Language models (LMs) are typically monolithic: a single model storing all knowledge and serving every use case. This design presents significant challenges; they often generate factually incorrect statements, require costly retraining to add or remove information, and face serious privacy and copyright issues. In this talk, I will discuss how to break this monolith by introducing modular architectures and training algorithms that separate capabilities across composable components. I'll cover two forms of modularity: (1) External modularity, which augments LMs with external tools like retrievers to improve factuality and reasoning; and (2) internal modularity, which builds inherently modular LMs from decentrally trained components to enable flexible composition and an unprecedented level of control. Bio: Weijia Shi is a Ph.D. candidate at the University of Washington, where she is advised by Luke Zettlemoyer and Noah Smith. Her research focuses on developing augmented and modular architectures and training algorithms that enable language models to be more controllable, collaboratively developed, and factual. She received an Outstanding Paper Award at ACL 2024 and was recognized as Rising Stars in Machine Learning in 2023 and in Data Science in 2024.

1 0