Events

Hide Sidebar Show Sidebar
Hide Sidebar Show Sidebar

Current Seminar Series

April 25, 2025

Foundations in Multimodal Mechanistic Interpretability (William Rudman, Brown University)

11:00A
- 12:15P

Location

32-D451
Add to Calendar 2025-04-25 11:00:00 2025-04-25 12:15:00 America/New_York Foundations in Multimodal Mechanistic Interpretability (William Rudman, Brown University) Abstract:  Mechanistic interpretability has been instrumental in understanding Large Language Models, yet remains underexplored in multimodal models. This is due to a lack of effective image-corruption methods needed for causal analysis. The first part of this talk introduces NOTICE, a novel corruption scheme designed for MLLMs, enabling causal mediation analysis for MLLMs.  Next, we examine the reasoning capabilities of MLLMs and find that MLLMs are shape-blind. Namely, vision-encoders in MLLMs embed geometrically dissimilar objects into the same regions of their representation space. We construct a side-counting dataset of abstract shapes, showing that current MLLMs achieve near-zero accuracy on a trivial task for humans.Finally, we present ongoing work on VisualCounterfact, a dataset designed to investigate the relationship between counterfactual visual inputs and world knowledge. VisualCounterfact consists of tuples that alter specific visual properties—color, size, and texture—of common objects. For instance, given (banana, color, yellow), we create a counterfact image (banana, color, purple) by modifying the object's pixels. Using VisualCounterfact, we locate a mechanism for reliably controlling whether a model will answer with the counterfactual property present in the image or retrieve the world-knowledge answer from its weights. TBD

April 26, 2025

New England Symposium on Graphics (NESG)

9:45A
- 4:45P

Location

34-101
Add to Calendar 2025-04-26 9:45:00 2025-04-26 16:45:00 America/New_York New England Symposium on Graphics (NESG) The New England Symposium on Graphics (NESG) is back! NESG is a one-day informal get-together for students, postdocs, and faculty in the area studying computer graphics and adjacent fields (computer-aided design, geometry, fabrication, vision, photography, etc.). The program consists of invited talks, a poster session, and plenty of time to catch up with collaborators. For more information, please check the event’s website https://nesg.graphics TBD

April 28, 2025

Industry Innovators Expo

2:00P
- 4:00P

Location

TBD
R&D Commons
Add to Calendar 2025-04-28 14:00:00 2025-04-28 16:00:00 America/New_York Industry Innovators Expo On Monday, April 28, from 2-4PM, CSAIL Alliances will host an Industry Innovators Expo where our member companies will have the opportunity to demonstrate their latest technology, talk about their technical progress and challenges, and give away swag. There will be light refreshments and opportunities for networking, so please join us in the R&D Commons on the 4th floor.   TBD

April 29, 2025

CSAIL Alliances Student Poster Session

5:00P
- 6:30P

Location

TBD
R&D Commons
Add to Calendar 2025-04-29 17:00:00 2025-04-29 18:30:00 America/New_York CSAIL Alliances Student Poster Session Come learn about the groundbreaking work happening in Stata on Tuesday, April 29, from 5-6:30PM at the CSAIL Student Poster Session. This is a chance for CSAIL students and postdocs to highlight their research, talk about its implications, and engage with industry members about its business applications. Meet your peers in the R&D Commons on the 4th floor—there will be refreshments! TBD

April 30, 2025

Managing Exploratory AI

Yongjoo Park
University of Illinois at Urbana-Champaign
1:00P
- 2:00P

Location

32-G882
Add to Calendar 2025-04-30 13:00:00 2025-04-30 14:00:00 America/New_York Managing Exploratory AI Abstract: Today’s data science systems, ranging from batch jobs to interactive interfaces, are surprisingly fragile. Data scientists typically use dozens of libraries, but a single bug in any of them can destroy hours or even days of computation, causing significant pain. This issue has been widely discussed in the data science community and academic literature. Yet, no principled mechanisms have been proposed to address the issue. It may be puzzling to database researchers. Existing databases implement checkpointing to periodically save changes for future recovery. Why haven’t data science systems adopted it? Are there any unique properties that challenge the adoption? In this talk, I will first identify a core challenge: the lack of mechanisms for detecting data changes, a key premise of checkpointing. Unlike databases with centralized buffer pools, data science systems intentionally omit centralized data spaces, allowing individual libraries to use shared memory, GPUs, and remote machines. Changes across these diverse locations must still be identified. To address this, we are making exciting progress around one central theme: a nonintrusive state manager that behaves like conventional buffer pools without requiring data to be placed in a central location. The key idea is to construct a mathematical map of library-managed data—including data dependencies—using graphs. These graphs enable new algorithms to detect changes, save them incrementally, and restore states correctly. We are actively developing an open-source system, Kishu, to benefit all data practitioners.Bio: Yongjoo Park is an Assistant Professor in the School of Computing and Data Science at the University of Illinois at Urbana-Champaign. His research focuses on systems for data-intensive AI. Yongjoo is also a Chief Scientist of Keebo, a start-up company he co-founded based on his Ph.D. research. Yongjoo obtained a Ph.D. in Computer Science and Engineering from the University of Michigan, Ann Arbor. He is a recipient of 2018 SIGMOD Jim Gray Dissertation Honorable Mention and ACM SIGMOD 2023 Best Artifact Award Honorable Mention.-- For the zoom passcode, contact the organizer at markakis@mit.edu TBD

How to Appease a Voter Majority

Prasanna Ramakrishnan
Stanford University

Part Of

4:00P
- 5:00P

Location

32-G575
Add to Calendar 2025-04-30 16:00:00 2025-04-30 17:00:00 America/New_York How to Appease a Voter Majority In 1785, Condorcet established a frustrating property of elections and majority rule: it is possible that, no matter which candidate you pick as the winner, a majority of voters will prefer someone else. You might have the brilliant idea of picking a small set of winners instead of just one, but how do you avoid the nightmare scenario where a majority of the voters prefer some other candidate over all the ones you picked? How many candidates suffice to appease a majority of the voters? In this talk, we will explore this question. Along the way, we will roll some dice — both because the analysis involves randomness and because of a connection to the curious phenomenon of intransitive dice, that has delighted recreational and professional mathematicians alike ever since Martin Gardner popularized it in 1970.Based on joint work with Moses Charikar, Alexandra Lassota, Adrian Vetta, and Kangning Wang. TBD

May 01, 2025

OPTIKS: An Optimized Key Transparency System

JULIA LEN
MIT / UNC Chapel Hill

Part Of

12:00P
- 1:00P

Location

32-D463
Stata (Star)
Add to Calendar 2025-05-01 12:00:00 2025-05-01 13:00:00 America/New_York OPTIKS: An Optimized Key Transparency System Abstract Key Transparency (KT) refers to a public key distribution system with transparency mechanisms proving its correct operation, i.e., proving that it reports consistent values for each user's public key. While prior work on KT systems have offered new designs to tackle this problem, relatively little attention has been paid on the issue of scalability. Indeed, it is not straightforward to actually build a scalable and practical KT system from existing constructions, which may be too complex, inefficient, or non-resilient against machine failures. In this paper, we present OPTIKS, a full featured and optimized KT system that focuses on scalability. Our system is simpler and more performant than prior work, supporting smaller storage overhead while still meeting strong notions of security and privacy. Our design also incorporates a crash-tolerant and scalable server architecture, which we demonstrate by presenting extensive benchmarks. Finally, we address several real-world problems in deploying KT systems that have received limited attention in prior work, including account decommissioning and user-to-device mapping.  TBD

Understanding the Trade-Offs Between Hallucinations and Mode Collapse in Language Generation

Grigoris Velegkas
Yale University

Part Of

4:00P
- 5:00P

Location

32-D507
Add to Calendar 2025-05-01 16:00:00 2025-05-01 17:00:00 America/New_York Understanding the Trade-Offs Between Hallucinations and Mode Collapse in Language Generation Specifying all desirable properties of a language model is challenging, but certain requirements seem essential. Given samples from an unknown language, the trained model should produce valid strings not seen in the training set, and be expressive enough to capture the language's full breadth. Otherwise, outputting invalid strings constitutes "hallucination," and failing to capture the full breadth leads to "mode collapse." Recent work by Kleinberg and Mullainathan [KM24], building on classical work on the closely related problem of language identification by Gold [Gol67] and Angluin [Ang79, 80], provides a concrete mathematical framework to study the problem of language generation. Kleinberg and Mullainathan showed that for all countable collections of languages, it is possible to create a language model that does not hallucinate but suffers from mode collapse. They asked whether this tension between validity and breadth is inherent for language generation.In this talk, we define various notions of breadth for language generation, and completely characterize when generation with validity and breadth is possible under each of these notions. Our results answer the question of [KM24] and show that this tension between validity and breadth is indeed inherent for language generation. Moreover, we formalize the notion of stable generation, a natural requirement derived from Gold’s work [Gol67], and discuss when this type of generation is achievable. Finally, we discuss the implications of our results in the universal rates setting of Bousquet, Hanneke, Moran, van Handel, and Yehudayoff [BGMvY21]. The talk is based on joint works with Alkis Kalavasis and Anay Mehrotra. TBD

May 02, 2025

(Thesis Defense) Building Intelligence that can Interact with the Physical World

Part Of

2:30P
- 3:30P

Location

32-G449
Add to Calendar 2025-05-02 14:30:00 2025-05-02 15:30:00 America/New_York (Thesis Defense) Building Intelligence that can Interact with the Physical World Speaker: Johnson Tsun-Hsuan WangAffiliation: MIT EECS (CSAIL)Title: [Thesis Defense] Building Intelligence that can Interact with the Physical WorldDate: Friday, May 2nd 2025Time: 2:30 pm EDTLocation: 32-G449 (Patil/Kiva)Zoom: https://mit.zoom.us/j/95448197150 Abstract: Recent advances in Artificial Intelligence (AI) have demonstrated remarkable success in parsing, reasoning, and generating digital content across modalities such as natural language, speech, images, videos, and 3D data. However, these breakthroughs have yet to extend meaningfully beyond the digital realm into the physical world. Developing AI for physical interaction poses challenges such as limited grounding, scarce physical data, and high reliability demands in safety-critical settings.This talk outlines a holistic approach to physical AI—through the lenses of data, brain, and body. We begin with data, the foundation of learning, and introduce data-driven and knowledge-driven robot simulation that generates data to improve policy learning and to systematically evaluate and probe existing models. Next, we turn to the brain, focusing on how to bridge the internet-scale knowledge of digital AI with the physical world to improve generalization and interpretability. Finally, we examine the body—the morphological component of intelligence—demonstrating how pre-trained generative models, when integrated with physics-based simulation, can automate the design of robot bodies. Together, this talk explores how digital AI can be extended into the physical world through a comprehensive investigation of data, brain, and body – laying the groundwork for building physical AI.Committee:Prof. Daniela Rus, MIT CSAIL (Advisor)Prof. Sertac Karaman, MIT LIDSProf. Wojciech Matusik, MIT CSAIL TBD

May 06, 2025

CSAIL Forum with Manish Raghavan: The role of information diversity in AI systems

Part Of

12:00P
- 1:00P

Location

TBD
Add to Calendar 2025-05-06 12:00:00 2025-05-06 13:00:00 America/New_York CSAIL Forum with Manish Raghavan: The role of information diversity in AI systems Registration required: https://mit.zoom.us/meeting/register/GP_RXB5BSTy_Ubf3wNJwxQBio: Manish Raghavan is the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management and Department of Electrical Engineering and Computer Science. Before that, he was a postdoctoral fellow at the Harvard Center for Research on Computation and Society (CRCS). His research centers on the societal impacts of algorithms and AI. TBD

May 13, 2025

May 14, 2025

Scalable Image AI via Self-designing Storage

Utku Sirin
Harvard University
1:00P
- 2:00P

Location

32-G575
Add to Calendar 2025-05-14 13:00:00 2025-05-14 14:00:00 America/New_York Scalable Image AI via Self-designing Storage Abstract: Image AI has the potential to improve every aspect of human life, providing more productive and safer services and tools. Image AI is, however, very expensive. It costs millions of dollars to train and deploy a single model with hundreds of thousands of pounds of carbon emission. We identify that the root cause of the problem is a long-overlooked and largely unexplored dimension: storage. The majority of the cost of reading and processing an image depends on how it is stored on disk and in main memory, as storage determines how much data is moved and processed. Most images today are stored as JPEG files. But, JPEG is designed for the human eye; it aims to maximally compress images with minimal loss in visual quality. In contrast, we observe that, during image AI, it is AI algorithms that “see” the images, rather than humans. Furthermore, JPEG is a single design. AI problems, however, are diverse; every problem is unique in terms of how data should be stored and processed. Using a fixed design, such as JPEG, for all problems results in excessive data movement and wasteful image AI pipelines.This talk presents Image Calculator, a self-designing storage system that creates and manages storage for image AI tasks. Unlike state-of-the-art that uses a fixed storage format, Image Calculator builds a design space of thousands of storage formats, each capable of storing and representing data differently, at different training and inference speed, accuracy, and space trade-offs. Given an AI task, Image Calculator searches and finds the optimal storage format that minimizes training and inference times and maximizes accuracy. Image Calculator consists of two main components: (i) search & training, and (ii) model-serving. Search & training efficiently searches within the design space by using locality among storage formats. Formats that have similar features also perform similarly. It clusters information-dense formats and quickly identifies high-quality candidates with scalable search time. Model-serving deploys storage formats at inference servers. It exploits the inherent frequency structure in image data. It breaks images into pieces, i.e., frequency components, and processes images frequency by frequency instead of image by image as conventionally done. This allows dramatically reducing data communication between clients & servers with fast and efficient inference.We evaluate Image Calculator across a diverse set of datasets, tasks, models, and hardware. We show that Image Calculator can generate storage formats that reduce end-to-end inference and training time by up to 14.2x and consume space by up to 8.2x with little or no loss in accuracy for image classification, object detection, and instance segmentation, compared to state-of-the-art image storage formats, such as JPEG and its recent variants. Image Calculator’s storage formats reduce individual time components, such as PCIe time, by up to 271x. Image Calculator is even more successful on small hardware devices providing sub-millisecond inference on CPUs, making inference scalable and cheap on commodity hardware. Tailoring storage to AI tasks results in heavily compressed and specialized data. Despite that, we show that Image Calculator is able to reconstruct images with high visual quality. Image Calculator’s incremental computation scheme allows moving just enough data for every image, further reducing data movement cost, and providing fast and scalable inference serving.This talk will mainly be based on the following two papers:[1] The Image Calculator: 10x Faster Image-AI Inference by Replacing JPEG with Self-designing Storage Format. Utku Sirin, Stratos Idreos. SIGMOD, 2024 [Link][2] Frequency-Store: Scaling Image AI by A Column-Store for Images. Utku Sirin, Victoria Kauffman, Aadit Saluja, Florian Klein, Jeremy Hsu, Stratos Idreos. CIDR, 2025 [Link]Bio: Utku Sirin is a postdoctoral researcher at the Data Systems lab at Harvard University, advised by Stratos Idreos. Utku’s work on the Image Calculator reimagines Image AI through self-designing AI storage, which always takes the best shape given the AI context and goals, bringing 10x speedup end-to-end. Utku was awarded the Microsoft Research PhD Fellowship in 2017 and the Swiss National Science Foundation Postdoctoral Fellowship in 2021 and 2023. Utku has also been a winner of the ACM SIGMOD Students Research Competition and a recipient of an IEEE ICDE best reviewer award. Before joining Harvard, Utku obtained his PhD from the Data-Intensive Applications and Systems lab at EPFL, advised by Anastasia Ailamaki on hardware-conscious data systems. -- For the zoom passcode, contact the organizer at markakis@mit.edu TBD