Add to Calendar
2024-08-12 13:00:00
2024-08-12 14:00:00
America/New_York
CRYPTO practice talk - Seyoon Ragavan
Seminar Room G575
Events
August 10, 2024
No events scheduled
August 12, 2024
Thesis Defense: Hanshen Xiao, "Automated and Provable Privatization for Black-Box Processing"
Hanshen Xiao
MIT
Add to Calendar
2024-08-12 14:00:00
2024-08-12 16:00:00
America/New_York
Thesis Defense: Hanshen Xiao, "Automated and Provable Privatization for Black-Box Processing"
Abstract: Can we automatically and provably quantify and control the information leakage from a black-box processing? From a statistical inference standpoint, in this talk, I will start from a unified framework to summarize existing privacy definitions based on input-independent indistinguishability and unravel the fundamental challenges in crafting privacy proof for general data processing. Yet, the landscape shifts when we gain access to the (still possibly black-box) secret generation. By carefully leveraging its entropy, we unlock the black-box analysis. This breakthrough enables us to automatically “learn" the underlying inference hardness for an adversary to recover arbitrarily-selected sensitive features fully through end-to-end simulations without any algorithmic restrictions. Meanwhile, a set of new information-theoretical tools will be introduced to efficiently minimize additional noise perturbation assisted with sharpened adversarially adaptive composition. I will also unveil the win-win situation between the privacy and stability for simultaneous algorithm improvements. Concrete applications will be given in diverse domains, including privacy-preserving machine learning on image classification and large language models, side-channel leakage mitigation and formalizing long-standing heuristic data obfuscations. Thesis Committee: Profs. Srini Devadas, Edward Suh and Henry Corrigan-GibbsZoom link: https://mit.zoom.us/j/96012380112
G882
August 13, 2024
Add to Calendar
2024-08-13 11:00:00
2024-08-13 12:00:00
America/New_York
Thesis Defense: Building Strategic AI Agents for Human-centric Multi-agent Systems
Title: "Building Strategic AI Agents for Human-centric Multi-agent Systems"Date: Tuesday, August 13, 2024Time: 11:00 AM - 12:00 PM (Eastern Time)Location: - In-person: D463 (Star), Stata Center, 32 Vassar Street, Cambridge, MA, 02139 - Virtual: https://mit.zoom.us/j/94031036791Thesis Supervisor: Jacob Andreas Thesis Committee: Gabriele Farina, Constantinos Daskalakis, Roger LevyAbstract:This thesis addresses the challenge of developing strategic AI agents capable of effective decision-making and communication in human-centric multi-agent systems. While significant progress has been made in AI for strategic decision-making, creating agents that can seamlessly interact with humans in multi-agentic settings remains a challenge. This research explores the limitations of current approaches, such as self-play reinforcement learning (RL) and imitation learning (IL), and proposes novel methods to overcome these constraints. Modeling human-like communication and decision making is a crucial first step toward building effective strategic agents. The initial part of the thesis addresses this through two approaches. We start by developing a regret minimization algorithm for modeling actions of strong and human-like agents called piKL, that incorporates a cost term proportional to the KL divergence between a search policy and a human IL policy. This approach improves reward while keeping behavior close to a human IL policy, producing agents that predict human actions accurately while improving performance in the benchmark game of no-press Diplomacy. Then, we develop a procedure for modeling populations of agents that communicate with humans using natural language. Our sample-efficient multitask training scheme for latent language policies improves the reward obtained by these policies while preserving the semantics of language in a complex real-time strategy game. Building on these foundations, the second part of the thesis focuses on building strategic agents for human-centric multi-agent domains. The research introduces the DiL-piKL planning algorithm and its extension, RL-DiL-piKL, which regularize self-play RL and search towards a human IL policy. These algorithms enable the training of Diplodocus, an agent achieving expert human-level performance in no-press Diplomacy. A significant milestone is reached with Cicero, the first AI agent to achieve human-level performance in full-press Diplomacy, integrating a language model (LM) with planning and RL algorithms based on piKL. The final part of the thesis revisits language generation tasks, applying piKL to model pragmatic communication and improving LM truthfulness. It presents Regularized Conventions, a model of pragmatic language understanding that outperforms existing best response and rational speech act models across several datasets. Furthermore, a novel approach to LM decoding is introduced, casting it as a regularized imperfect-information sequential signaling game. This results in the Equilibrium-Ranking algorithm, which consistently improves performance over existing language model decoding procedures.
D463 (Star)
August 14, 2024
TALK: Thesis Defense: Video Understanding in Artificial and Biological Neural Networks
Camilo Fosco
MIT-CSAIL
Add to Calendar
2024-08-14 14:00:00
2024-08-14 15:00:00
America/New_York
TALK: Thesis Defense: Video Understanding in Artificial and Biological Neural Networks
Abstract:The human brain processes visual information in ways that remain largely unknown. Research has been conducted to understand complex human patterns, such as abstractions, in videos, but the problem remains challenging. Additionally, there is still a significant open question regarding our ability to extract meaningful information related to video content from brain signals. This thesis addresses two primary questions: First, can we leverage human-like priors—i.e., priors derived from our biological networks—to build better machine learning models for video classification and forecasting, and use them to predict membership inference over abstract concepts in a human-like manner? Second, can we comprehend the signals traveling through our biological networks well enough to reconstruct embedded visual concepts? Specifically, can we reconstruct exact videos that were viewed from fMRI signals?We explore solutions and applications to both problems. By constructing a semantic relational graph, a model can be trained to predict the abstraction connecting different videos. Additionally, by collecting a curated fMRI-video dataset, realistic reconstructions of watched videos can be achieved using diffusion models. This work aims to explore both dimensions and propose machine learning models that approximate working solutions to both challenges.Thesis Committee: Profs. Bill Freeman and Phillip Isola
32-D463
August 16, 2024
Yevgeniy Dodis: Perpetual Encryption
Yevgeniy Dodis (New York University)
Add to Calendar
2024-08-16 10:30:00
2024-08-16 12:00:00
America/New_York
Yevgeniy Dodis: Perpetual Encryption
We consider the problem of building a private blockchain (BC) on top of a public one. This has the advantage that users of the private BC do not need to build expensive consensus protocol, while still maintaining privacy. When building this object, we realize the need for a novel encryption scheme we call perpetual encryption (PE), which is a new primitive of independent interest. Very roughly, PE schemes support dynamically changing keys sk_1, sk_2, ... in a way that ciphertext c_n for epoch n:(1) can be efficiently decrypted using any future key sk_n, sk_{n+1},...(2) looks random even given all past keys sk_1,..,sk_{n-1}. We also show how to build an efficient PE from any public-key encryption, hashing and other standard symmetric primitives.
32-G882 Hewlett Room
Thesis Defense: "Leveraging Mechanics for Multi-step Robotic Manipulation Planning" (Rachel Holladay)
CSAIL / EECS
Add to Calendar
2024-08-16 13:00:00
2024-08-16 14:30:00
America/New_York
Thesis Defense: "Leveraging Mechanics for Multi-step Robotic Manipulation Planning" (Rachel Holladay)
Abstract: This thesis focuses on enabling robots to robustly perform complex, multi-step manipulation tasks, like chopping vegetables or wielding a wrench. Completing such tasks requires a robot to plan and execute long sequences of actions, where each action involves many connected, discrete and continuous choices that are critically impacted by constraints relating to force, motion and contact. To tackle this, this thesis contributes models and algorithms that exploit the physics and geometry of the world in order to address the dual challenges of long-horizon decision-making and acting under uncertainty. We apply this in the context of three domains: in-hand manipulation, forceful manipulation and briefly-dynamic manipulation.First, to reorient a grasped object, we develop a sampling-based motion planner to generate sequences of pushes that slide the object in-hand. We derive an abstraction for pushing to enable the planner to reason about frictional constraints. Second, we focus on forceful manipulation tasks, such as opening a childproof medicine bottle or twisting a nut on a bolt, where the robot's planning choices are impacted by the need to exert force. We define constraints that explicitly consider torque and frictional limits and integrate these into an existing task and motion planning framework. We leverage cost-sensitive planning to enable the robot to generate plans that are robust to uncertainty in the physical parameters. Finally, we frame planning with dynamic actions, like shoveling or toppling, as requiring the robot to reason about both action uncertainty and potential dead ends. We learn a simple action model and formulate a sample-based manipulation planner that guards against dead ends in the face of uncertainty. Throughout this thesis, we validate the practical applicability of our model-based approaches by evaluating them on real robots.Thesis Committee: Tomás Lozano-Pérez, Alberto Rodriguez, Leslie Kaelbling, Dmitry Berenson
34-401A (Grier) (zoom link available upon request)
August 19, 2024
Exascale Climate Emulators: Enhancing Earth System Model Outputs and Reducing Petabytes of Storage
Sameh Abdulah
King Abdullah University of Science and Technology
Add to Calendar
2024-08-19 14:00:00
2024-08-19 15:00:00
America/New_York
Exascale Climate Emulators: Enhancing Earth System Model Outputs and Reducing Petabytes of Storage
Title:Exascale Climate Emulators: Enhancing Earth System Model Outputs and Reducing Petabytes of Storage Abstract:We present an exascale climate emulator to tackle the rising computational and storage demands of high-resolution Earth System Model simulations. Using spherical harmonic transform, we model spatio-temporal climate variations stochastically, providing tunable resolution and significantly enhancing emulation fidelity. We extend linear solver software to multi-precision arithmetic GPUs, adapting to different correlation strengths. The PaRSEC runtime system optimizes parallel matrix operations by balancing computation, communication, and memory. Our BLAS3-rich code is optimized for diverse GPU systems, achieving 0.523 EFlops/s on 4,096 ORNL Frontier nodes (projected 1.04 EFlops/s on 8,192 nodes), 0.243 EFlops/s on 1,024 Cineca Leonardo nodes, and 0.375 EFlops/s on 3,072 ORNL Summit nodes. Bio:Sameh Abdulah obtained his MS and Ph.D. degrees from Ohio State University, Columbus, USA, in 2014 and 2016, respectively. Presently, he serves as a research scientist at the Extreme Computing Research Center (ECRC), King Abdullah University of Science and Technology, Saudi Arabia. His research focuses on various areas, including high-performance computing applications, big data, bitmap indexing, handling large spatial datasets, parallel spatial statistics applications, algorithm-based fault tolerance, and machine learning and data mining algorithms. Sameh was a part of the KAUST team nominated for the ACM Gordon Bell Prize in 2022 for their work on large-scale climate/weather modeling and prediction.
Seminar Room G449 (Patil/Kiva)
August 21, 2024
Add to Calendar
2024-08-21 15:00:00
2024-08-21 16:00:00
America/New_York
Thesis Defense: Techniques for Foundational End-to-End Verification of Systems Stacks
Zoom Link:https://mit.zoom.us/j/94297415474Abstract:Today's software is full of bugs and vulnerabilities. Formal verification provides a promising remedy through mathematical specifications and machine-checked proofs that the implementations conform to the specifications. However, there could still be bugs in the specifications or in the verification tools, which could lead to missed bugs in the software being verified. Therefore, this dissertation advocates for foundational end-to-end verification, a proof-based software development method that can mitigate both of these concerns:It is end-to-end in the sense that the correctness proofs of individual components are used to discharge the assumptions of adjacent components throughout the whole stack, resulting in end-to-end theorems that only mention the top-most and bottom-most specifications, so that bugs in intermediate specifications cannot invalidate the soundness of the end-to-end statement anymore.The method is foundational in the sense that the soundness of the proofs only relies on the foundations of mathematics and on the correctness of a small proof-checking kernel, but not on the correctness of other, domain-specific verification tools, because these tools are either proven correct once-and-for-all, or they output proofs that are checked by the kernel.Ensuring that all the reasoning can be checked by the same small foundational kernel requires considerable effort, and the first part of this dissertation presents techniques to reduce this effort:Omnisemantics, a new style of semantics that can be used instead of traditional small-step or big-step operational semantics, offer a smooth way of combining undefined behavior and nondeterminism, and enable forward-simulation compiler correctness proofs with nondeterministic languages, whereas previous approaches need to fall back to the much less convenient backward simulations if support for nondeterminism is needed.Live Verification is proposed, a technique to turn an interactive proof assistant into a programming assistant that displays the symbolic state of the program as the user writes it and allows the user to tweak the symbolic state as long as the tweaks are provably sound. An additional convenience-improving feature is that instead of stating lengthy loop invariants, the user only needs to give the diff between the symbolic state before the loop and the desired loop invariant, resulting in shorter and more maintainable annotations. Finally, in order to make Live Verification practical, a number of additional proof techniques is presented.The second part of the dissertation shows how these techniques were useful in three collaborative case studies: An embedded system running on a verified processor with an end-to-end proof where the software-hardware interface specification cancels out, a cryptographic server with an end-to-end proof going from high-level elliptic-curve math all the way down to machine code, and a trap handler to catch unsupported-instruction exceptions whose correctness proof combines program-logic proofs about C-level functions, a compiler correctness proof, and proofs about hand-written assembly.
32-G882
August 22, 2024
Tackling Algorithmic Problems on Massive Graphs
Amartya Shankha Biswas
MIT
Add to Calendar
2024-08-22 11:00:00
2024-08-22 12:00:00
America/New_York
Tackling Algorithmic Problems on Massive Graphs
Abstract: Confronted with increasingly large data sets, the traditional models of computation that read the entire input are often impractical, due to constraints on time, memory, randomness and other computational resources. These limitations have led to the emergence and popularity of new models such as streaming algorithms, sublinear algorithms, oracle access models, and various models of parallel and distributed algorithms for large scale data analysis.This thesis studies such alternative algorithmic approaches to process massive graphs under constraints of time, memory, and randomness. Part of our focus is on sublinear time algorithms that aim to deliver accurate approximations of graph properties without the need to process the entire graph. These settings typically necessitate a careful definition of input access and output delivery. We also investigate models of parallel and distributed computation for very large graph datasets. We will focus on algorithms for the problems of subgraph counting and graph sparsification in sublinear, distributed, parallel, and local models of computation.Committee: Ronitt Rubinfeld, Piotr Indyk, VIrginia Williams
32-D463
August 30, 2024
Add to Calendar
2024-08-30 10:00:00
2024-08-30 13:00:00
America/New_York
CSAIL new grad orientation
CSAIL orientation program for incoming new grad students
32-G449
September 11, 2024
Reading Alan Turing
Avi Wigderson, IAS
IAS, Princeton U.
Add to Calendar
2024-09-11 17:00:00
2024-09-11 18:00:00
America/New_York
Reading Alan Turing
Abstract: I will discuss some well-known and less-known papers of Turing, exemplify the scope of deep, prescient ideas he put forth, and mention follow-up work on these by the Theoretical CS community.No special background will be assumed.
32-123
September 20, 2024
Cancelled
Seyoon Ragavan: Indistinguishability Obfuscation from Bilinear Maps and LPN Variants
This event has been cancelled
December 01, 2024
Cancelled
Test event
This event has been cancelled
February 19, 2025
Dertouzos Distinguished Lecture: Deborah Estrin
Deborah Estrin
Cornell Tech
Add to Calendar
2025-02-19 17:00:00
2025-02-19 18:00:00
America/New_York
Dertouzos Distinguished Lecture: Deborah Estrin
32-123