Events

Hide Sidebar Show Sidebar
Hide Sidebar Show Sidebar

Current Seminar Series

July 14, 2025

Thesis Defense: Topics in Geometric Machine Learning

Behrooz Tahmasebi
MIT CSAIL
1:00P
- 2:00P

Location

32-G882
MIT Stata Center (Building 32) (32 Vassar Street)
Add to Calendar 2025-07-14 13:00:00 2025-07-14 14:00:00 America/New_York Thesis Defense: Topics in Geometric Machine Learning Abstract: Recent advances and the widespread adoption of neural networks have revolutionized machine learning and artificial intelligence. These developments demand learning paradigms capable of processing data from diverse applications and sources. In structured domains such as molecules, graphs, sets, and 3D objects, as well as fields such as drug discovery, materials science, and astronomy, models must account for data structures. The emerging field of geometric machine learning has gained attention for enabling neural networks to handle geometric structures, unlocking novel solutions across scientific disciplines.Despite recent advances, theoretical gaps remain. This thesis aims to address these gaps by studying the benefits and limitations of leveraging geometric structures and symmetries in data. We explore sample complexity, generalization bounds, hypothesis testing for the presence of symmetries in data, time complexity of learning under symmetries, and regularization and optimization in symmetric settings. The goal is to build a robust theoretical framework that validates recent successes and sheds light on unexplored aspects, fostering future progress in geometric machine learning.Thesis Advisor: Stefanie Jegelka Thesis Committee: Tess Smidt, Tommi Jaakkola  TBD

July 15, 2025

Private Event

THESIS DEFENSE: Seeing Beyond Limits with Physics-Informed Priors

Part Of

2:00P
- 3:00P

Location

32-D463
Add to Calendar 2025-07-15 14:00:00 2025-07-15 15:00:00 America/New_York THESIS DEFENSE: Seeing Beyond Limits with Physics-Informed Priors THESIS DEFENSE: Seeing Beyond Limits with Physics-Informed PriorsSpeaker: Yang LiuSpeaker Affiliation: MIT EECS & CSAILHost: Frédo DurandHost Affiliation: MIT EECS & CSAILDate: Tuesday, July 15, 2025Time: 2:00 PM to 3:00 PMLocation: 32-D463 (Star) or Zoom Link: https://mit.zoom.us/j/98534109114Abstract:Conventional imaging systems face inherent dimensionality and visibility limits, primarily because image sensors are typically two-dimensional, and light tends to diffuse on rough surfaces or scatter within complex media. In this talk, I will reframe imaging systems through the lens of optical encoding and neural decoding, presenting my key contributions aimed at transcending the traditional limits of dimensionality and visibility. The idea is modelling the forward physical process and iteratively optimizing it with deep denoisers as visual priors, where eventually the priors are physics-informed. First, I introduce Privacy Dual Imaging, which reveals the privacy risk that ambient light sensors embedded in most smart devices could capture images of the scene in front of the screen. This idea of seeing the invisible from subtle intensity fluctuations is inspired by George Orwell’s novel 1984, wherein Big Brother is watching you through a two-way telescreen, and it closely relates to incoherent lensless imaging and non-line-of-sight imaging. Second, I present Snapshot Compressive Imaging, which encodes multiple temporal, spectral, or angular frames into a single measurement captured by a standard two-dimensional sensor. By learning high-dimensional visual priors from image or video data, we can efficiently reconstruct the original higher-dimensional data cube at scale. Lastly, I show that large AI models, particularly diffusion models, can serve as generic visual priors for both cases and beyond. I aim to push the boundaries of imaging and sensing within relevant domains of AI for science and healthcare (with an example).Committee Members: Frédo Durand (advisor, MIT), William T. Freeman (MIT & Google), Kaiming He (MIT & Google)Relevant URL: https://mit.zoom.us/j/98534109114For more information please contact: Roger White <[email protected]>  TBD

July 16, 2025

Thesis Defense: Foundational Abstractions for Quantum Programming

Charles Yuan
MIT CSAIL
2:00P
- 3:00P

Location

32-G575
Add to Calendar 2025-07-16 14:00:00 2025-07-16 15:00:00 America/New_York Thesis Defense: Foundational Abstractions for Quantum Programming Bringing the promise of quantum computation into reality requires not only building a quantum computer but also correctly programming it to run a quantum algorithm. To obtain asymptotic advantage over classical algorithms for applications including simulation, search, and optimization, quantum algorithms rely on the ability of data in quantum superposition to exhibit phenomena such as interference and entanglement. In turn, an implementation of the algorithm as a program must correctly orchestrate these phenomena in the states of qubits. Otherwise, it would yield incorrect outputs or lose quantum computational advantage.Given a quantum algorithm, what are the challenges and costs of realizing it as a program that can run on a physical quantum computer? In this thesis, I answer this question by showing how the basic abstractions of programming upon which many quantum algorithms rely – such as data structures and control flow – can fail to work correctly or efficiently on a quantum computer. I then demonstrate how we can leverage insights from research in programming languages to re-invent the software stack – including abstractions, libraries, and compilers – to meet the demands of quantum algorithms. This approach holds out a promise of expressive and efficient tools to program a quantum computer and thereby practically realize its computational advantage.Thesis Supervisor: Michael Carbin TBD

July 17, 2025

Thesis Defense: Vectorizing Complex Programs for Modern Architectures

Tom Chen
MIT-CSAIL
12:00P
- 2:00P

Location

32-G575
Add to Calendar 2025-07-17 12:00:00 2025-07-17 14:00:00 America/New_York Thesis Defense: Vectorizing Complex Programs for Modern Architectures Vector instructions are ubiquitous in modern processors, enabling efficient fine-grained parallelism across a wide range of architectures. Despite decades of research, auto-vectorization remains a challenging task, with current production vectorizers frequently failing to vectorize real-world programs due to limitations in handling complex control flow, ambiguous data dependencies, and complex instruction sets.This thesis revisits Superword-Level Parallelism (SLP) as a general framework for vectorization and presents three key contributions to address the core challenges to vectorization. First, I introduce SuperVectorization, a generalized SLP vectorizer that uses a new intermediate representation, Predicated SSA, to vectorize across basic blocks and loop nests. Second, I describe a fine-grained program versioning framework to handle ambiguous dependencies with runtime checks, allowing vectorization of programs with dependencies that cannot be analyzed by static techniques alone. Finally, I present VeGen, a system that can automatically use an increasingly common class of complex vector instructions that no longer fit into our traditional model of SIMD parallelism while using only their semantic descriptions as input.Together, these techniques extend the capability of compiler auto-vectorization and allow us to auto-vectorize a larger class of programs while using the vector hardware more effectively. My prototype achieves significant speedup across a variety of benchmarks, showing its potential as a foundation for the next generation of auto-vectorizers. Thesis Committee:  Professors, Saman Amarasinghe, Michael Carbin and Vikram Adve (UIUC)Zoom: https://mit.zoom.us/j/94654420771 TBD

July 28, 2025

Bespoke Threat Models: Achieving Realistic Privacy Guarantees for Deployed Protocols

Kyle Hogan
MIT CSAIL
1:00P
- 3:00P

Location

32-G882
Hybrid
Add to Calendar 2025-07-28 13:00:00 2025-07-28 15:00:00 America/New_York Bespoke Threat Models: Achieving Realistic Privacy Guarantees for Deployed Protocols This thesis focuses on the question of what degree of privacy is achievable in the real world for long running applications. We explore this question in two main settings: anonymous communication and private advertising. In doing so we consider constraints each application may have in practice and what adversarial model is realistic for the context in which the application will be deployed.In the space of private advertising, we propose a novel protocol, Adveil, that eliminates leakage of user data beyond that revealed by the input/output of the ads ecosystem as a whole. We also provide a minimal modeling of the functionality of digital advertising which we use to prove that, even for systems like Adveil with minimal leakage, the advertising metrics released at the end of the protocol are sufficient to leak information about end users to advertisers when combined with their audience targeting criteria. TBD

September 23, 2025

Explicit Lossless Vertex Expanders

Rachel Zhang
CSAIL, EECS

Part Of

4:15P
- 5:15P

Location

32-G449
Refreshments at 4:00 PM
Add to Calendar 2025-09-23 16:15:00 2025-09-23 17:15:00 America/New_York Explicit Lossless Vertex Expanders We give the first explicit construction of lossless vertex expanders. These are d-regular graphs where every small set S of vertices has (1-eps)d|S| distinct neighbors. Previously, the strongest known explicit vertex expanders were those given by Ramanujan graphs, whose spectral properties imply that every small set S of vertices has 0.5d|S| distinct neighbors.Based on joint work with Jun-Ting Hsieh, Ting-Chun Lin, Alex Lubotzky, Sidhanth Mohanty, Ryan O'Donnell, and Assaf Reiner. TBD