Events

Hide Filters Show Filters
Hide Filters Show Filters

Current Seminar Series

CSAIL Forum
Dertouzos Distinguished Lecture
Hot Topics in Computing
AI@MIT Reading Group
Algorithms and Complexity (A&C) 2025 - 2026
Bioinformatics Seminar 2025
Biomedical Imaging and Analysis 2025 - 2026
Boston IEEE/ACM 2025 -2026
Brains, Minds and Machines 2025 - 2026
CIS Seminar 2025-2026
CSAIL Security Seminar 2025 - 2026
EECS Special Seminar
Embodied Intelligence 2025-2026
HCI Seminar 2025-2026
ML+Crypto Seminar
ML Tea
Theory of Computation (ToC) 2025 - 2026
Thesis Defense
Previous Seminar Series

February 4, 2026

From Agentic LLMs to an Agentic Data System for Autonomous Data Science

Ju Fan
Renmin University of China
1:00P
- 2:00P

Location

32-G882
Add to Calendar 2026-02-04 13:00:00 2026-02-04 14:00:00 America/New_York From Agentic LLMs to an Agentic Data System for Autonomous Data Science Abstract: Autonomous data science, a longstanding goal of the data community, aims to automate the entire data science pipeline for extracting insights from structured data. Existing approaches to applying large language models (LLMs) to autonomous data science are either domain-specific or rely on predefined pipelines, which limits their autonomy and adaptivity in end-to-end data science tasks. Recent advances in agentic LLMs create new opportunities to support autonomous planning, execution, and iteration over multi-step tasks. However, moving from agentic LLMs to true end-to-end autonomy for data science introduces new challenges and calls for a shift toward an agentic data system. In this talk, I will first present a vision of agentic data systems and highlight key research challenges, including autonomous pipeline orchestration, environment-aware iterative reasoning, and data-intensive execution environments. I will then discuss our recent work that begins to address these challenges and outline future directions.Bio: Ju Fan is a Professor at Renmin University of China. He received his Ph.D. from Tsinghua University and was previously a research fellow at the National University of Singapore. His research interests lie in intelligent data systems (AI4DB), with a current focus on building agentic data science systems that enable end-to-end autonomous data science. He has published over 70 papers in top conferences and journals, including SIGMOD, VLDB, ICDE, and TKDE. He served as Publication Chair for VLDB 2023 and 2024, and has been a Program Committee member for leading database conferences, including SIGMOD, VLDB, ICDE, and KDD. He also led the development of DeepAnalyze, an early end-to-end agentic LLM system for autonomous data science. His work has received the ACM SIGMOD 2024 Research Highlight Award and the ICDE 2025 Best Paper Runner-Up Award, and he is also a recipient of the ACM China Rising Star Award.----Please reach out to markakis@mit.edu for the Zoom password. TBD

Agreement testers and PCPs from coset complexes

Noah Singer
Carnegie Mellon

Part Of

4:00P
- 5:00P

Location

32-G882
Hewlett
Add to Calendar 2026-02-04 16:00:00 2026-02-04 17:00:00 America/New_York Agreement testers and PCPs from coset complexes "Agreement testers” are objects used in the design of (some) probabilistically checkable proofs, which, in turn, play a fundamental role in modern complexity theory and cryptography. Recent breakthrough works [Bafna–Lifshitz–Minzer, Dikstein–Dinur–Lubotzky 2024] analyzed a certain sophisticated construction and showed that it has strong agreement testing properties. In our work, we establish the same result for the so-called "Kaufman–Oppenheim (KO) complex”, an alternative construction which is more elementary, explicit, and symmetric. Ultimately, our proof boils down to a bound on the 'complexity', in a precise sense, of the group of upper triangular matrices with 1's on the diagonal over a finite field.In the talk, I will informally define the agreement testing problem and its relationship with "higher-dimensional analogues" of expander graphs, before presenting, from first principles, our bound and some ideas from its proof. Based on joint work with Ryan O'Donnell. TBD

February 5, 2026

FAST CODE SEMINAR: Highway to Performance -- Portable C++ SIMD

Jan Wassenberg
Google DeepMind in Zürich
12:00P
- 1:00P

Location

TBD
Virtual see announcement for registration
Add to Calendar 2026-02-05 12:00:00 2026-02-05 13:00:00 America/New_York FAST CODE SEMINAR: Highway to Performance -- Portable C++ SIMD Speaker: Jan WassenbergSpeaker Affiliation: Google DeepMindHost: Roberto PalmieriHost Affiliation: Lehigh University---Date: Thursday, Feb 5, 2026Time: 12:00 PM - 1:00 PM ET---Virtual seminar: Zoom registration required at https://mit.zoom.us/meeting/register/oRdRZlMtRQ-gAoFOVQC6fwAbstract:Modern CPUs offer significant speedups through Single Instruction, Multiple Data (SIMD), but accessing this performance often implies a difficult choice: rely on unpredictable compiler auto-vectorization, or hand-write intrinsics for each CPU.Based on a series of workshops for Google engineers, this talk introduces a better alternative: Highway. This open-source C++ library allows you to write SIMD code once and efficiently target Arm, LoongArch, POWER, RISC-V, WASM, x86, and IBM Z.We will cover the fundamentals of vectorization and discuss real-world use cases, including LLM inference (gemma.cpp) and image processing. The session provides a practical guide on porting standard C++ loops to portable and type-safe vector code. Going deeper, we demonstrate how to handle loop remainders and variable-length vectors, and how to implement runtime dispatch for selecting the best available instructions. This enables you to unlock hardware performance while keeping your codebase maintainable.---Bio:Jan Wassenberg is a Senior Staff Software Engineer at Google DeepMind in Zürich, where he develops infrastructure for high-performance software. With a PhD in efficient algorithms and over 20 years of focus on SIMD, Jan bridges the gap between raw hardware capabilities and software usability.He is the main author of Highway, the open-source C++ library that helps numerous projects efficiently target six CPU families with much less toil. Its users include Chromium, NumPy, ScaNN, TensorFlow, and V8.Jan is currently tech lead of gemma.cpp (open-source LLM inference). His previous work includes optimizing the JPEG XL image codec (also co-chairing its standardization), devising the fastest known vectorized quicksort (vqsort), and designing Randen, Google’s default secure random number generator.---Contact:Bruce Hoppebehoppe@mit.edu--- Event URL: https://fastcode.org/events/fastcode-seminar/jan-wassenberg/ TBD

February 6, 2026

[Thesis Defense] Morris Yau: On Structure, Parallelism, and Approximation in Modern Neural Sequence Modeling

Morris Yau
9:30A
- 10:30A

Location

32-D463
Add to Calendar 2026-02-06 9:30:00 2026-02-06 10:30:00 America/New_York [Thesis Defense] Morris Yau: On Structure, Parallelism, and Approximation in Modern Neural Sequence Modeling Abstract:  Is there an algorithm that learns the best fit parameters of a Transformer to any dataset? If I trained a neural sequence model and promised you it is equivalent to a program, how would you even be convinced? Modern RNNs are functions that admit parallelizable recurrence; what is the design space of parallelizable recurrences? Are there unexplored function families that lie between RNNs and Transformers? We explore these questions from first principles starting with state, polynomials, and parallelism. TBD

[Thesis Defense] Performance Portable Scientific Computing through Multi-Level Compiler Optimizations

Avik Pal
CSAIL

Part Of

11:00A
- 1:00P

Location

Room 34-401A
Grier Room A
Add to Calendar 2026-02-06 11:00:00 2026-02-06 13:00:00 America/New_York [Thesis Defense] Performance Portable Scientific Computing through Multi-Level Compiler Optimizations Scientific computing faces a trade-off between mathematical expressiveness and performance across heterogeneous hardware. This work presents a compiler-driven approach that automatically preserves and recovers mathematical structure in existing scientific codes to generate optimized code without manual re-engineering. By combining graph optimizations, abstraction raising, communication optimizations, multi-level automatic differentiation, and learned cost models within a unified compiler infrastructure, the system enables optimizations that are infeasible in isolation. Our evaluations on large-scale scientific applications, including differentiable climate and hypersonic flow simulations, demonstrate consistent performance gains across thousands of GPUs and TPUs. The methods and artifacts presented in this thesis demonstrate a path toward scientific computing where the complexity of heterogeneous execution is handled automatically, leaving scientists free to focus on the mathematics. TBD

February 10, 2026

February 13, 2026

[Thesis Defense] Learning Intelligent Contact for Dynamic Robots

Gabriel Margolis
MIT CSAIL

Part Of

11:00A
- 1:00P

Location

32-D463
Add to Calendar 2026-02-13 11:00:00 2026-02-13 13:00:00 America/New_York [Thesis Defense] Learning Intelligent Contact for Dynamic Robots When a robot's foot strikes the ground or its hand presses against an object, interaction forces propagate through the kinematic chain and appear as joint torques and motions, providing a proprioceptive sense of touch. Foundational work in learning-based legged locomotion implicitly processes proprioceptive information to achieve forceful tasks, as presented across the first three chapters of this thesis, which demonstrate systems for highly dynamic running, contact-parameterized walking, and leg-based object manipulation. Subsequently, the core of this thesis develops novel learning-based formulations for controlling contact-force interactions in legged robots. First, we address force regulation through virtual force fields that simulate resistance during training, demonstrating that whole-body force control is achievable through reinforcement learning without dedicated force sensors. Second, we train policies rewarded for estimating physical properties, allowing informative probing behaviors to emerge; when friction estimation is rewarded, the robot learns to scuff its foot while walking, generating informative shear forces. Third, we advance the safety and generalization of humanoid control by reframing compliant whole-body control as an augmented motion-imitation problem, enabling humanoids to yield to external forces while learning from human movements. Together, these contributions show how data-driven control can treat contact as a source of information and opportunity for control rather than a disturbance to be rejected. TBD

February 17, 2026

Jia-Bin Huang visiting seminar & one-on-one meetings: Controllable Visual Imagination

11:00A
- 12:00P

Location

45-792
Large seminar room on 7th floor of Schwarzman College of Computing. Exit elevator - turn right - turn right again.
Add to Calendar 2026-02-17 11:00:00 2026-02-17 12:00:00 America/New_York Jia-Bin Huang visiting seminar & one-on-one meetings: Controllable Visual Imagination Prof. Jia-Bin Huang is visiting MIT for a day! He will give a talk, and you may also sign up for one-on-one meetings in this Google Sheet:https://docs.google.com/spreadsheets/d/12tvBn91bdDkAW-oaIrThR5DlcIzY_WEkGt1vYBuOlEk/edit?usp=sharingTalk abstract:Generative models have empowered human creators to visualize their imaginations without artistic skills and labor. A prominent example is large-scale text-to-image/video generation models. However, these models are often difficult to control and do not respect 3D perspective geometry and the temporal consistency of videos. In this talk, I will showcase several of our recent efforts to improve controllability for visual imagination. Specifically, I will discuss how we enable semantic and spatial control for 2D image generation, facilitate layered decompositions for video editing, and synthesize object and camera motions from monocular videos. Short bio: Jia-Bin Huang is a Capital One-endowed Associate Professor of Computer Science at the University of Maryland, College Park. Before coming to UMD, Huang was a research scientist at Meta Reality Labs and an Assistant Professor of Electrical and Computer Engineering at Virginia Tech. Huang received his Ph.D. from the University of Illinois, Urbana-Champaign (UIUC) in 2016. His research interests include 3D computer vision, generative models, and computational photography. Huang is the recipient of the Thomas & Margaret Huang Award, NSF CRII award, faculty award from Samsung, Google, 3M, Qualcomm, Meta, and a Google Research Scholar Award.  TBD

Visual Computing Seminar: Learning a distance measure from the information-estimation geometry of data

Guy Ohayon
Flatiron Institute
12:00P
- 1:00P

Location

32-D463
Add to Calendar 2026-02-17 12:00:00 2026-02-17 13:00:00 America/New_York Visual Computing Seminar: Learning a distance measure from the information-estimation geometry of data Abstract:The perceptual distance between images is widely believed to be related to the distribution of natural images. But how can a probability distribution give rise to a distance measure—let alone one that aligns with human perception? What properties should such a distance satisfy, and how can it be learned from an image database in an unsupervised manner? In this talk, I will address these questions by presenting the Information–Estimation Metric (IEM), a novel form of distance function derived from a given probability density over a domain of signals. The IEM is rooted in a fundamental relationship between information theory and estimation theory, which links the log-probability of a signal with the errors of an optimal denoiser, applied to noisy observations of the signal. For Gaussian-distributed signals, the IEM coincides with the Mahalanobis distance. But for more complex distributions, it adapts, both locally and globally, to the geometry of the distribution. I will discuss and illustrate the theoretical properties of the IEM—including its global and local behavior. Finally, I will demonstrate that the IEM effectively predicts human perceptual judgments when trained (unsupervised) on natural images.Bio:Guy is a postdoctoral researcher working with Eero Simoncelli at the Flatiron Institute. His research focuses on developing computational models of human perception that are grounded in principles from information theory. He received his PhD in Computer Science from the Technion—Israel Institute of Technology, where he worked with Michael Elad and Tomer Michaeli on the design and theoretical analysis of image restoration and compression methods that rely on generative models. TBD

February 19, 2026

Human-Machine Partnerships in Computer-Integrated Interventional Medicine: Yesterday, Today, and Tomorrow

Russell H. Taylor
Johns Hopkins University, Baltimore, MD

Part Of

11:00A
- 12:00P

Location

32-370
Add to Calendar 2026-02-19 11:00:00 2026-02-19 12:00:00 America/New_York Human-Machine Partnerships in Computer-Integrated Interventional Medicine: Yesterday, Today, and Tomorrow This talk will discuss insights gathered over 35 years of research on medical robotics and computer-integrated interventional medicine (CIIM), both at IBM and at Johns Hopkins University. The goal of this research has been the creation of a three-way partnership between physicians, technology, and information to improve treatment processes. CIIM systems combine innovative algorithms, robotic devices, imaging systems, sensors, and human-machine interfaces to work cooperatively with surgeons in the planning and execution of surgery and other interventionalprocedures. For individual patients, CIIM systems can enable less invasive, safer, and more cost-effective treatments. Since these systems have the ability to act as “flight data recorders” in the operating room, they can enable the use of statistical methods to improve treatment processes for future patients and to promote physician training. We will illustrate these themes with examples from our past and current work, with special attention to the human-machine partnership aspects, and will offer some thoughts about future research opportunities and system evolution. TBD

February 24, 2026

March 3, 2026

March 12, 2026

Interdiction problems and 2-person sequential games: beyond NP-completeness

Jim Orlin
MIT Sloan School of Management

Part Of

4:15P
- 5:15P

Location

TBD
Add to Calendar 2026-03-12 16:15:00 2026-03-12 17:15:00 America/New_York Interdiction problems and 2-person sequential games: beyond NP-completeness AbstractIn the Knapsack Problem (KP), a decision maker wants to select items of value V or more to put into a knapsack subject to a weight limit.  In the Interdiction Knapsack Problem (IKP), an adversary can block K items from being selected.  The adversary’s goal is to prevent the decision maker from achieving a value of V. The IKP is a sequential game with an initial move by the adversary followed by a move from the decision maker. The Knapsack Problem is NP-complete. The IKP is one level harder than NP-complete. We have proved a meta-theorem which establishes that the Interdiction Traveling Salesman Problem, the Interdiction Set Packing Problem, the Interdiction Set Cover Problem, and hundreds of other interdiction problems are all one level harder than NP-complete.  We have also established meta-theorems showing that hundreds of min max regret problems are one level harder than NP-complete.  We discuss our meta-theorems and relate them to the complexity of 2-person sequential games.This talk does not assume a knowledge of complexity other than an understanding of NP-completeness.This work is joint with Christoph Grüne, Berit Johannes, and Lasse Wulf.Speaker BioJames Orlin is the E. Pennell Brooks (1917) Professor of Operations Research at the MIT Sloan School.  He is best known for his research on obtaining faster algorithms for problems in network and combinatorial optimization and for his text with Ravi Ahuja and Tom Magnanti entitled Network Flows: Theory, Algorithms, and Applications. He has won various awards for his co-authored publications: including the 1993 Lanchester Prize for the best publication in O.R., the 2016 ACM SIGecom Test of Time Award, for a paper published between 10 and 25 years ago that has had “significant impact on research or applications that exemplify the interplay of economics and computation,” and the 2020 Khachyian prize for lifetime achievements in the area of optimization.  He is also a MacVicar fellow, an honor that is awarded at MIT to faculty for their “outstanding contributions to undergraduate education, and for their exceptional teaching, mentoring, and educational innovation.”  TBD

March 17, 2026

March 31, 2026

April 7, 2026

April 14, 2026

April 21, 2026

April 28, 2026

May 5, 2026

May 12, 2026