Add to Calendar
2024-03-19 10:00:00
2024-03-19 11:00:00
America/New_York
EECS Special Seminar: Emma Dauterman - Secure systems from insecure components
Abstract: Today’s computer systems provide weakest-link security: just one phishing attack, one software vulnerability, or one hardware flaw can have catastrophic consequences. And today’s systems have countless weak links – already in 2024, attackers have stolen millions of social security numbers and health records. In theory, cryptography makes it possible to survive compromise, but in practice, general-purpose tools are often prohibitively expensive. On top of that, it’s impossible to change many legacy systems. Weakest-link security and its dangerous ramifications seem inevitable.This talk will show otherwise. I will describe two systems that can withstand compromise at modest cost. One is a hardened authentication system that is usable today with most websites. The other is an encrypted backup system that protects against physical attacks and runs on legacy hardware. The key idea is to tailor the cryptographic tools to the system setting, providing precisely the necessary properties and pushing cryptographic work to where computation is cheap. This approach shows that even if we cannot protect every credential, patch every vulnerability, or secure every piece of hardware, we can still protect users and their data.Bio: Emma Dauterman is a Ph.D. candidate at UC Berkeley where she is advised by Raluca Ada Popa and Ion Stoica. Her research interests include computer security, systems, and applied cryptography. She has received the Microsoft Research Ada Lovelace fellowship, the NSF graduate research fellowship, and a UC Berkeley EECS excellence award.***https://mit.zoom.us/j/99765087011
Grier Room B
Events
March 18, 2024
No events scheduled
March 19, 2024
EECS Special Seminar: Sewon Min, "Rethinking Data Use in Large Language Models"
Sewon Min
Dept. CS & Engineering, University of Washington
Add to Calendar
2024-03-19 11:00:00
2024-03-19 12:00:00
America/New_York
EECS Special Seminar: Sewon Min, "Rethinking Data Use in Large Language Models"
Abstract:Large language models (LMs) such as ChatGPT have revolutionized natural language processing and artificial intelligence more broadly. In this talk, I will discuss my research on understanding and advancing these models, centered around how they use the very large text corpora they are trained on. First, I will describe our efforts to understand how these models learn to perform new tasks after training, demonstrating that their so-called in context learning capabilities are almost entirely determined by what they learn from the training data. Next, I will introduce a new class of LMs—nonparametric LMs—that repurpose this training data as a data store from which they retrieve information for improved accuracy and updatability. I will describe my work on establishing the foundations of such models, including one of the first broadly used neural retrieval models and an approach that simplifies a traditional, two-stage pipeline into one. I will also discuss how nonparametric models open up new avenues for responsible data use, e.g., by segregating permissive and copyrighted text and using them differently. Finally, I will envision the next generation of LMs we should build, focusing on efficient scaling, improved factuality, and decentralization.Bio:Sewon Min is a Ph.D. candidate in the Paul G. Allen School of Computer Science & Engineering at the University of Washington. Her research focuses on language models (LMs): studying the science of LMs, and designing new model classes and learning methods that make LMs more performant and flexible. She also studies LMs in information-seeking, legal, and privacy contexts. She is a co-organizer of multiple tutorials and workshops, including most recently at ACL 2023 on Retrieval-based Language Models and Applications and upcoming at ICLR 2024 on Mathematical and Empirical Understanding of Foundation Models. She won a paper award at ACL 2023, received a J.P. Morgan Fellowship, and was named an EECS rising star in 2022.
Grier A (34-401A) and zoom, https://mit.zoom.us/j/92345755112
From Large to Small Datasets: Size Generalization for Clustering Algorithm Selection
Ellen Vitercik
Stanford University
Add to Calendar
2024-03-19 16:15:00
2024-03-19 17:15:00
America/New_York
From Large to Small Datasets: Size Generalization for Clustering Algorithm Selection
Abstract: In clustering algorithm selection, we are given a massive dataset and must efficiently select which clustering algorithm to use. We study this problem in a semi-supervised setting, with an unknown ground-truth clustering that we can only access through expensive oracle queries. Ideally, the clustering algorithm's output will be structurally close to the ground truth. We approach this problem by introducing a notion of size generalization for clustering algorithm accuracy. We identify conditions under which we can (1) subsample the massive clustering instance, (2) evaluate a set of candidate algorithms on the smaller instance, and (3) guarantee that the algorithm with the best accuracy on the small instance will have the best accuracy on the original big instance. We verify these findings both theoretically and empirically.This is joint work with Vaggos Chatziafratis and Ishani Karmarkar.
32-G449
March 20, 2024
EECS Special Seminar: Silvia Sellan, "Stochastic Computer Graphics"
Silvia Sellan
Department of Computer Science, University of Toronto
Add to Calendar
2024-03-20 11:00:00
2024-03-20 12:00:00
America/New_York
EECS Special Seminar: Silvia Sellan, "Stochastic Computer Graphics"
AbstractComputer Graphics research has long been dominated by the interests of large film, television and social media companies, forcing other, more safety-critical applications (e.g., medicine, engineering, security) to repurpose Graphics algorithms originally designed for entertainment. In this talk, I will advocate for a perspective shift in our field that allows us to design algorithms directly for these safety-critical application realms. I will show that this begins by reinterpreting traditional Graphics tasks (e.g., 3D modeling and reconstruction) from a statistical lens and quantifying the uncertainty in our algorithmic outputs, as exemplified by the research I have conducted for the past five years. I will end by mentioning several ongoing and future research directions that carry this statistical lens to entirely new problems in Graphics and Vision and into specific applications.BioSilvia is a fifth year Computer Science PhD student at the University of Toronto, working in Computer Graphics and Geometry Processing. She is a Vanier Doctoral Scholar, an Adobe Research Fellow and the winner of the 2021 University of Toronto Arts & Science Dean’s Doctoral Excellence Scholarship. She has interned twice at Adobe Research and twice at the Fields Institute of Mathematics. She is also a founder and organizer of the Toronto Geometry Colloquium and a member of WiGRAPH.
Kiva 32-G448 and zoom, https://mit.zoom.us/j/94493487416
Building a Theory for Distributed Systems
Nancy Lynch
MIT/CSAIL
Add to Calendar
2024-03-20 12:00:00
2024-03-20 13:00:00
America/New_York
Building a Theory for Distributed Systems
Title: Building a Theory for Distributed SystemsThis presentation is part of a series of talks by holders of the Ellen Swallow Richards Professorship which was established by AMITA (the Association of MIT Alumnae) with the purpose of bringing leading women scholars to MIT. It honors the pioneering spirit and professional achievements of Mrs. Richards and strengthens the role of women on the MIT faculty. The Ellen Swallow Richards Chair is intended to recognize the national importance of research and education by women at MIT.Please register in advance even if you plan to attend in person at https://alumcommunity.mit.edu/events/77564/booking/fc08ed18-5b80-4e1c-b019-51b47647274b/ticket-selection. After registering, you will receive a confirmation email containing information about joining the webinar.Please indicate on the registration form if you plan to attend in person. This will help us determine whether the room is close to reaching capacity. Abstract:In this talk I will overview my work on developing a Theory for Distributed Systems---work that involved many students and other collaborators, at MIT and elsewhere. This effort started at Georgia Tech in the late 1970s, and has continued at MIT since 1981.Here I will emphasize the earlier contributions, and their impact on the directions of the field. These include abstract models for problems that are solved by distributed systems and for the algorithms used to solve them; new algorithms; lower bounds and other “impossibility results” expressing inherent limitations on the power of distributed systems; general mathematical foundations for modeling and analyzing distributed systems; and applications of these methods to understanding a variety of practical distributed systems.Bio:Nancy Lynch PhD, 1972, the first Ellen Swallow Richards Professor (1982-87), is currently a Professor (Post-Tenure) in the Department of Electrical Engineering and Computer Science at MIT. She heads the Theory of Distributed Systems research group in MIT's CSAIL laboratory. She started her career as a professor at Georgia Institute of Technology before she returned to MIT as the first Ellen Swallow Richards Professor.Lynch received her B.S. from Brooklyn College and her PhD from MIT, both in mathematics. She has written and co-written hundreds of research articles about distributed algorithms and impossibility results, and about formal modeling and verification of distributed systems. Her best-known contribution is the 1982 ``FLP'' impossibility result for distributed consensus in the presence of process failures, with Fischer and Paterson, followed by a paper with Dwork and Stockmeyer on algorithms for reaching consensus under restricted failure assumptions. Other contributions include the I/O automata system modeling frameworks, with Tuttle, Kaynar, Segala, and Vaandrager, as well as recent results on wireless network algorithms and biological distributed algorithms.Lynch is the author of the textbook ``Distributed Algorithms'' and co-author of ``Atomic Transactions'' and ``The Theory of Timed I/O Automata.'' She is an ACM Fellow, a Fellow of the American Academy of Arts and Sciences, and a member of both the National Academies of Sciences and Engineering. She has been awarded the Dijkstra Prize (twice), the van Wijngaarden prize, the Knuth Prize, the Piore Prize, and the Athena Prize, among others. She has supervised over 100 graduate students and postdocs, many of whom have themselves become leading researchers.
Durable Execution: The Backbone of Fault-Tolerant Software
Maxim Fateev
Temporal
Add to Calendar
2024-03-20 13:00:00
2024-03-20 14:00:00
America/New_York
Durable Execution: The Backbone of Fault-Tolerant Software
Title: Durable Execution: The Backbone of Fault-Tolerant SoftwareAbstract:The concept of Durable Execution is rapidly gaining traction in backend software development. Applications such as Snap stories, Coinbase transactions, Airbnb bookings, DoorDash deliveries, Taco Bell online orders, HashiCorp cloud provisioning operations, and tens of thousands of other mission-critical workloads are implemented with durable executions.Maxim, who originated the idea, played a pivotal role in its implementation at AWS and Uber. He is currently the CEO and co-founder of Temporal, a leading open-source Durable Execution Platform.In his talk, Maxim will delve into the concepts of Durable Execution, explore implementation options, and examine common use cases. He will also discuss future research areas, including resilient distributed operating systems, durable RPC, and more.Bio:Maxim has spent the last 20 years building massive distributed systems for Amazon, Microsoft, Google, and Uber. Among other things, he led the design and development of the AWS SQS backend, and AWS Simple Workflow Service. At Uber, Maxim led the effort on open source projects Cherami and Cadence Workflow.Since October of 2019, Maxim has been the CEO and Cofounder of Temporal Technologies. Its flagship open source project temporal.io is redefining the way large-scale reliable applications are developed and operated.Tianyu Li is inviting you to a scheduled Zoom meeting.Topic: Durable Execution: The Backbone of Fault-Tolerant SoftwareTime: Mar 20, 2024 01:00 PM Eastern Time (US and Canada)Join Zoom Meetinghttps://mit.zoom.us/j/95138205143?pwd=YjlpanN5VUp6cW92d0p6b0MwNld5QT09Password: 939318One tap mobile+16465588656,,95138205143# US (New York)+16699006833,,95138205143# US (San Jose)Meeting ID: 951 3820 5143US : +1 646 558 8656 or +1 669 900 6833International Numbers: https://mit.zoom.us/u/abKimXlUS8Join by SIP95138205143@zoomcrc.comJoin by Skype for Businesshttps://mit.zoom.us/skype/95138205143
32-G575
Key Overwriting Attacks
Miro Haller
Add to Calendar
2024-03-20 15:00:00
2024-03-20 16:00:00
America/New_York
Key Overwriting Attacks
n this talk, I will formally define key overwriting attacks and discuss some recent applications.After lying dormant for 20 years, a recent series of papers exploited key overwriting attacks to break the security of deployed end-to-end encrypted schemes. More and more, systems aim to protect users even against a malicious or compromised server. Together with complex key hierarchies, this lead to attacks where the adversary can overwrite (part of) the key material of users. By observing the client's operation on such (partially) corrupted key material, some attacks were able to go as far as recovering the key material.This talk is based on "MEGA: Malleable Encryption Goes Awry" and "Caveat Implementor! Key Recovery Attacks on MEGA" but I will also touch on other key recovery attacks.
G-882 Hewlett Room
Add to Calendar
2024-03-20 16:00:00
2024-03-20 17:00:00
America/New_York
Near Optimal Alphabet-Soundness Tradeoff PCPs
Abstract: We show a nearly optimal alphabet-soundness tradeoff for NP-hardness of 2-Prover-1-Round Games (2P1R). More specifically, we show that for all \eps > 0, for sufficiently large M, it is NP-hard to decide whether a 2P1R instance of alphabet size M has value nearly 1 or at most M^{-1+\eps}. 2P1R are equivalent to 2-Query PCPs, and are widely used in obtaining hardness of approximation results. As such, our result implies the following: 1) hardness of approximating Quadratic Programming within a factor of nearly log(n), 2) hardness of approximating d-bounded degree 2-CSP within a factor of nearly d/2, and 3) improved hardness of approximation results for various k-vertex connectivity problems. For the former two applications, our results nearly match the performance of the best known algorithms. Joint work with Dor Minzer.
32-G575
Algorithmic Discrepancy Theory and Randomized Controlled Trials
Daniel Spielman
Yale University
Add to Calendar
2024-03-20 17:00:00
2024-03-20 18:00:00
America/New_York
Algorithmic Discrepancy Theory and Randomized Controlled Trials
Abstract:Theorems in discrepancy theory tell us that it is possible to partition a set of vectors into two sets that look surprisingly similar to each other. In particular, these sets can be much more similar to each other than those produced by a random partition. For many measures of similarity, computer scientists have developed algorithms that produce these partitions efficiently.A natural application for these algorithms is the design of randomized controlled trials (RCTs). RandomizedControlled Trials are used to test the effectiveness of interventions, like medical treatments and educationalinnovations. Randomization is used to ensure that the test and control groups are probably similar. When we know nothing about the experimental subjects, a random partition into test and control groups is the best choice. When experimenters have measured quantities about the experimental subjects that they expect could influence a subject's response to a treatment, the experimenters try to ensure that these quantities are evenly distributed between the test and control groups. That is, they want a random partition of low discrepancy.In this talk, I will survey some fundamental results in discrepancy theory, present a model for the analysis of RCTs, and summarize results from my joint work with Chris Harshaw, Fredrik Sävje, and Peng Zhang that uses algorithmic discrepancy theory to improve the design of randomized controlled trials.Bio:Daniel Alan Spielman is the Sterling Professor of Computer Science, and Professor of Statistics and Data Science, and of Mathematics at Yale. He received his B.A. in Mathematics and Computer Science from Yale in 1992, and his Ph.D in Applied Mathematics from M.I.T. in1995. After spending a year as an NSF Mathematical Sciences Postdoctoral Fellow in the Computer Science Department at U.C. Berkeley, he became a professor in the Applied Mathematics Department at M.I.T. He moved to Yale in 2005.He has received many awards, including the 1995 ACM Doctoral Dissertation Award, the 2002 IEEE Information Theory Paper Award, the 2008 and 2015 Godel Prizes, the 2009 Fulkerson Prize, the 2010 Nevanlinna Prize, the 2014 Polya Prize, the 2021 NAS Held Prize, the 2023 Breakthrough Prize in Mathematics, a Simons Investigator Award, and a MacArthur Fellowship. He is a Fellow of the Association for Computing Machinery and a member of the National Academy of Sciences, the American Academy of Arts and Sciences, and the Connecticut Academy of Science and Engineering.His main research interests include the design andanalysis of algorithms, network science, machine learning, digital communications and scientific computing.
32-123
March 21, 2024
EECS Special Seminar: Omar Khattab - Building More Reliable and Scalable AI Systems with Language Model Programming
Omar Khattab
Stanford NLP
Add to Calendar
2024-03-21 10:00:00
2024-03-21 11:00:00
America/New_York
EECS Special Seminar: Omar Khattab - Building More Reliable and Scalable AI Systems with Language Model Programming
It is now easy to build impressive demos with language models (LMs) but turning these into reliable systems currently requires brittle combinations of prompting, chaining, and finetuning LMs. I will present LM programming, a systematic way to address this by defining and improving three layers of the LM stack. I start with how to adapt LMs to search for information most effectively (ColBERT) and how to scale such search to billions of tokens (PLAID). I then discuss the right architectures and supervision strategies (e.g., ColBERT-QA, Baleen, Hindsight) for allowing LMs to retrieve and cite verifiable sources in their responses. This leads to DSPy, a programming model that replaces ad-hoc techniques for using LMs with composable modules for building and automatically supervising controllable programs built with LMs. Even simple systems expressed in DSPy routinely outperform large standalone LMs and standard hand-crafted prompting pipelines, in some cases while using only small models. I highlight how ColBERT and DSPy have sparked applications at dozens of leading tech companies and academic labs. I then conclude by discussing how DSPy enables a new degree of research modularity around LMs, one that stands to allow open research to again lead the development of AI systems.Bio: Omar is a fifth-year CS Ph.D. candidate at Stanford NLP, whose work spans Natural Language Processing (NLP), information retrieval (IR), and ML systems. Omar's research creates models, systems, supervision strategies, and programming abstractions for building reliable, transparent, and scalable NLP systems. Omar is the author of the ColBERT retrieval model, which has helped shape the modern landscape of neural information retrieval. Omar's lines of work on ColBERT and DSPy form the basis of influential open-source projects (e.g., exceeding 500,000 downloads per month) and have sparked applications at Google, Amazon, IBM, VMware, Databricks, Baidu, AliExpress, and numerous startups. Omar's Ph.D. has been supported by the Eltoukhy Family Graduate Fellowship and the Apple Scholars in AI/ML PhD Fellowship.
Grier Room B
Extracting Secret Keys from a Device's Power LED Using COTS Video Cameras
Ben Nassi
Cornell Tech
Add to Calendar
2024-03-21 12:00:00
2024-03-21 13:00:00
America/New_York
Extracting Secret Keys from a Device's Power LED Using COTS Video Cameras
Over the past 25 years, research has highlighted the fact that high-end hardware can be used by attackers to recover secret keys from devices. Numerous studies have demonstrated innovative secret key extraction techniques that rely on dedicated professional equipment to capture data-dependent physical leakage from target devices. These methods employ equipment like scopes to obtain power traces, software-defined radio and probes to capture electromagnetic radiation (EMR) traces, as well as ultrasonic microphones to capture acoustic traces. While these methods have deepened our understanding regarding the cryptanalytic risks associated with various types of leakage (EMR, acoustic, power) and high-end sensors to secret keys, much less is known about the cryptanalytic risks posed by optical leakage and accessible ubiquitous equipment such as video cameras.In this talk, we will reveal the findings from the two research papers, optical cryptanalysis (CCS’23) and video-based cryptanalysis (SP’24), and discuss how attackers can extract cryptographic keys using video footage of a device’s power LEDs captured by standard video cameras. In the first part of the talk, we will review the history of the side-channel cryptanalytic attacks from the first timing attack that was published in 1996, through the cryptanalytic power-based attacks and cryptanalytic EMR attacks that were published since 1998 until the acoustic attack that was published at 2014 and conclude interesting insights regarding the lessons we learned from these works. Next, we will discuss information leakage from power LEDs (based on the findings presented at CCS 23), and understand why the intensity of the light emitted by a device’s power LED can be used as an alternative to power traces obtained from the device to recover secret keys (2048-bit RSA, 256-bit ECDSA and 378-bit SIKE keys) from commonly used cryptographic libraries (Libgcrypt, GnuPG, PQCryptoSIDH) using a photodiode. In the second part of the talk, we will discuss how standard video cameras (e.g., of an iPhone 13 PRO Max, and security camera) can be used as alternatives for the photodiodes (based on the findings presented at SP’24) to extract secret keys (256-bit ECDSA and 378-bit SIKE keys). We will discuss a video camera’s rolling shutter and understand how it can be used to increase the sampling rate of a video camera from the frame-per-second rate (60 measurements per second) to the rolling shutter rate (60,000 measurements per second). We will see videos of secret key recoveries that were taken by a smartphone and by an Internet-connected security camera to recover a 256-bit ECDSA key (using the Minerva side-channel attack) and a 378-SIKE key (using the HertzBleed side-channel attack). At the end of the talk, we will discuss countermeasures, and provide insights regarding the real potential of extracting cryptographic keys by video cameras in our days and in the near future, taking into account the expected improvements in the specifications of video cameras expected by Moore’s Law.
Advances in Speaker Diarization at Google
Quan Wang
Google
Add to Calendar
2024-03-21 14:00:00
2024-03-21 15:00:00
America/New_York
Advances in Speaker Diarization at Google
Abstract: In this talk, Dr. Quan Wang will introduce the development and evolution of speaker diarization technologies at Google in the past decade, and how they landed as impactful products such as Cloud Speech-to-Text and the Pixel Recorder app. The talk will cover four critical milestones of the speaker diarization technologies at Google: (1) leveraging deep speaker embeddings; (2) leveraging supervised clustering; (3) leveraging sequence transducers; and (4) leveraging large language models. The talk will also discuss how speaker diarization will evolve in the new era of multimodal large language models.Bio: Dr. Quan Wang is a Senior Staff Software Engineer at Google, leading the Hotword Research team, Hotword Quality team, and Speaker, Voice & Language team. Quan is an IEEE Senior Member, and was a former Machine Learning Scientist at Amazon Alexa team. Quan received his B.E. degree from Tsinghua University, and received his Ph.D. degree from Rensselaer Polytechnic Institute. Quan is the author of the Chinese textbook "Voice Identity Techniques: From core algorithms to engineering practice", winning the Distinguished Author of Year 2020 Award. Quan is also an online instructor, and his Speaker Recognition course is rated as a bestselling course on Udemy.
32-G449 (Kiva/Patil Conference Room) Stata Center
EI Seminar - Rob Platt - Leveraging Symmetries to Make Robot Learning More Data Efficient
Northeastern University
Add to Calendar
2024-03-21 16:00:00
2024-03-21 17:00:00
America/New_York
EI Seminar - Rob Platt - Leveraging Symmetries to Make Robot Learning More Data Efficient
Title: Leveraging Symmetries to Make Robotic Learning More Data EfficientAbstract: Many robotics problems have transition dynamics that are symmetric in SE(2) and SE(3) with respect to rotation, translation, scaling, reflection, etc. In these situations, any optimal policy will also be symmetric over these transformations. In this talk, I leverage this insight to improve the data efficiency of policy learning by encoding domain symmetries directly into the neural network model using group invariant and equivariant layers. The result is that we can learn non-trivial visuomotor control policies with much less data than is typically the case. For imitation learning, this significantly reduces the number of demonstrations required. For reinforcement learning, it reduces the amount of experience needed to learn a good policy. In fact, we can sometimes learn good policies from scratch training directly on physical robotic hardware in real time.Bio: Rob Platt is an Associate Professor in the Khoury College of Computer Sciences at Northeastern University and a Faculty Fellow at BDAI. He is interested in developing robots that can perform complex manipulation tasks alongside humans in the uncertain everyday world. Much of his work is at the intersection of robotic policy learning, planning, and perception. Prior to coming to Northeastern, he was a Research Scientist at MIT and a technical lead at NASA Johnson Space Center.
32-G449 (Patil/Kiva)
When is Agnostic Reinforcement Learning Statistically Tractable?
Zeyu Jia
LIDS MIT
Add to Calendar
2024-03-21 16:00:00
2024-03-21 16:30:00
America/New_York
When is Agnostic Reinforcement Learning Statistically Tractable?
Abstract: We study the problem of agnostic PAC reinforcement learning (RL): given a policy class Π, how many rounds of interaction with an unknown MDP (with a potentially large state and action space) are required to learn an ε-suboptimal policy with respect to Π? Towards that end, we introduce a new complexity measure, called the spanning capacity, that depends solely on the set Π and is independent of the MDP dynamics. With a generative model, we show that for any policy class Π, bounded spanning capacity characterizes PAC learnability. However, for online RL, the situation is more subtle. We show there exists a policy class Π with a bounded spanning capacity that requires a superpolynomial number of samples to learn. This reveals a surprising separation for agnostic learnability between generative access and online access models (as well as between deterministic/stochastic MDPs under online access). On the positive side, we identify an additional sunflower structure, which in conjunction with bounded spanning capacity enables statistically efficient online RL via a new algorithm called POPLER, which takes inspiration from classical importance sampling methods as well as techniques for reachable-state identification and policy evaluation in reward-free exploration.
32-370
March 22, 2024
Learning from Nisan's natural proofs
Ari Karchmer (Boston University)
Add to Calendar
2024-03-22 10:30:00
2024-03-22 12:00:00
America/New_York
Learning from Nisan's natural proofs
In this talk, we will discuss the concept of Natural Proofs of circuit lower bounds (Razborov and Rudich, 1994), and their general relationship to cryptography and computational learning theory. Then, we will dive into some recent progress in the area by highlighting a new technique for extracting efficient distribution-independent learning algorithms or fully-agnostic learning algorithms from a restricted class of Natural Proofs due to Nisan (1993). The new results presented in this talk are based in part on two papers:- Ari Karchmer. Distributional pac-learning from nisan’s natural proofs. In 15th Innovations in Theoretical Computer ScienceConference (ITCS 2024). Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, 2024.- Ari Karchmer. Agnostic membership query learning with nontrivial savings: New results, techniques. In 35th InternationalConference on Algorithmic Learning Theory (ALT 2024). PMLR, 2024.
32-G882 Hewlett Room
March 26, 2024
Byte Bites with Stephen Moskal
Stephen Moskal
Add to Calendar
2024-03-26 12:00:00
2024-03-26 13:00:00
America/New_York
Byte Bites with Stephen Moskal
AI Agents in Cyber Space - LLMs Killed the Script KiddieCybersecurity is a deeply technical, complex, and ever-evolving field where there is a struggle to defend against the constant barrage of adversarial attacks from around the world. With the aid of an LLM like ChatGPT, adversaries now need little actual cybersecurity knowledge to plan and conduct cyberattack campaigns. In this talk, Stephen will discuss how LLMs are changing the landscape of cybersecurity from both an offensive and defensive perspective. Stephen will show how their autonomous AI cyberagent can conduct reconnaissance, exploitation, and exfiltration actions on a real network environment with no human intervention. The goal of this talk is to bring awareness to the current and future cybersecurity capabilities of LLMs and harbor discussion of how to defend against such threats.Reference paper:https://arxiv.org/pdf/2310.06936.pdf
Quest | CBMM Seminar Series: Physical and Social Human-Robot Interaction with the iCub Humanoid
Giorgio Metta
Istituto Italiano di Tecnologia (IIT)
Add to Calendar
2024-03-26 16:00:00
2024-03-26 17:30:00
America/New_York
Quest | CBMM Seminar Series: Physical and Social Human-Robot Interaction with the iCub Humanoid
Abstract: The iCub is a humanoid robot designed to support research in embodied AI. At 104 cm tall, the iCub has the size of a five-year-old child. It can crawl on all fours, walk, and sit up to manipulate objects. Its hands have been designed to support sophisticate manipulation skills. The iCub is distributed as Open Source following the GPL licenses (http://www.iCub.org). More than 50 robots have been built so far which are available in laboratories across Europe, US, Korea, Singapore, and Japan. It is one of the few platforms in the world with a sensitive full-body skin to deal with the physical interaction with the environment including possibly people. In this talk I report about the work of two of research units of the Italian Institute of Technology whose focus is the design of methods to enable natural interaction with the iCub robot.Bio: Giorgio Metta is the Scientific Director of the Istituto Italiano di Tecnologia (IIT). He holds a MSc cum laude (1994) and PhD (2000) in electronic engineering both from the University of Genoa. From 2001 to 2002, Giorgio was postdoctoral associate at the MIT AI-Lab. He was previously with the University of Genoa and from 2012 to 2019 Professor of Cognitive Robotics at the University of Plymouth (UK). He was member of the board of directors of euRobotics aisbl, the European reference organization for robotics research. Giorgio Metta served as Vice Scientific Director of IIT from 2016 to 2019. He coordinated IIT's participation into two of the Ministry of Economic Development Competence Centers for Industry 4.0 (ARTES4.0, START4.0). He was one of the three Italian representatives at the 2018 G7 forum on Artificial Intelligence and, more recently, one of the authors of the Italian Strategic Agenda on AI. Giorgio coordinated the development of the iCub robot for more than a decade making it de facto the reference platform for research in embodied AI. Presently, there are more than 40 robots reaching laboratories as far as Japan, China, Singapore, Germany, Spain, UK and the United States. Giorgio Metta research activities are in the fields of biologically motivated and humanoid robotics and, in particular, in developing humanoid robots that can adapt and learn from experience. Giorgio Metta is author of more than 300 scientific publications. He has been working as principal investigator and research scientist in about a dozen international research as well as industrial projects.
April 01, 2024
Compositional Approaches to Modelling Language and Concepts
University of Bristol
Add to Calendar
2024-04-01 16:00:00
2024-04-01 17:00:00
America/New_York
Compositional Approaches to Modelling Language and Concepts
Zoom Link: https://mit.zoom.us/j/95408129253Abstract: Recent neural approaches to modelling language and concepts have proven quite effective, with a proliferation of large models trained on correspondingly massive datasets. However, these models still fail on some tasks that humans, and symbolic approaches, can easily solve. Large neural models are also, to a certain extent, black boxes - particularly those that are proprietary. There is therefore a need to integrate compositional and neural approaches, firstly to potentially improve the performance of large neural models, and secondly to analyze and explain the representations that these systems are using. In this talk I will present two compositional approaches for modelling language and concepts. I will describe applications to reasoning and aspects of language such as ambiguity or metaphor. I will go on to give extensions from a textual to a multimodal setting, and how to generalize to unseen concepts using composition. Finally, I will present some future directions in modelling analogy and for understanding the types of reasoning or symbol manipulation that large neural models may be performing.Bio:Martha is a Lecturer in the School of Engineering Mathematics and Technology at the University of Bristol, UK. Prior to Bristol, she held a Veni fellowship at the University of Amsterdam, was a postdoc in the Quantum Group at the University of Oxford, and is currently visiting the Santa Fe Institute working on approaches to modelling analogy. Her research interests are in compositional approaches to modelling language and concepts, and in how these can be integrated with neural or distributed systems.
Seminar Room D463 (Star)
How does one bit-flip corrupt an entire deep neural network, and what to do about it
Yanjing Li
University of Chicago
Add to Calendar
2024-04-01 16:00:00
2024-04-01 17:00:00
America/New_York
How does one bit-flip corrupt an entire deep neural network, and what to do about it
Abstract:Deep neural networks are increasingly susceptible to hardware failures. The impact of hardware failures on these workloads is severe – even a single bit-flip can corrupt an entire network during both training and inference. The urgency of tackling this challenge, known as the Silent Data Corruption challenge in a broader context, has been widely raised by both the industry and academia.In this talk, I will first present the first in-depth resilience study targeting DNN workloads and hardware failures that occur in the logic portion of deep learning accelerator systems, including a comprehensive characterization of hardware failure effects, along with the fundamental understanding of how hardware failures propagate in hardware devices and interact with the workloads. Next, based on the insights obtained from our study, I will present ultra lightweight yet highly effective techniques to mitigate hardware failures in deep learning accelerator systems.Bio:Yanjing Li is an Assistant Professor in the Department of Computer Science at the University of Chicago. Prior to joining the university, she was a senior research scientist at Intel Labs. Professor Li received her Ph.D. in Electrical Engineering from Stanford University, an M.S. in Mathematical Sciences (with honors) and a B.S. in Electrical and Computer Engineering (with a double major in Computer Science) from Carnegie Mellon University.Professor Li has received various awards, including the NSF CAREER Award, DAC under-40 innovators award, Google research scholar award, NSF/SRC energy-efficient computing: from devices to architectures (E2CDA) program award, Intel Labs Gordy academy award (highest honor in Intel Labs) and several other Intel recognition awards, outstanding dissertation award (European Design and Automation Association), and multiple best paper awards (ACM Great Lakes Symposium on VLSI, IEEE VLSI Test Symposium, and IEEE International Test Conference).
32-G575
April 02, 2024
EECS Special Seminar: Rachit Nigam - Modular Abstractions for Hardware Design
Rachit Nigam
Cornell University
Add to Calendar
2024-04-02 11:00:00
2024-04-02 12:00:00
America/New_York
EECS Special Seminar: Rachit Nigam - Modular Abstractions for Hardware Design
Abstract. The exponential performance improvements of general-purpose processors have long driven the modern computing revolution. But with the end of Dennard scaling and the rise of dark silicon, processor-based systems can no longer sustain these trends. Instead, hardware accelerators, which sacrifice computational generality for efficiency, have become ubiquitous and enabled dramatic improvements in domains from machine learning (Google TPU) to networking (Intel Tofino). In the acceleration era, we must rethink the strict separation between software and hardware design and enable domain experts to design and deploy accelerators.In this talk, I will present two new systems for designing hardware accelerators. First, Filament, a hardware description language that uses a novel type system to enable modular reasoning of hardware designs and eliminates a large class of bugs at compile time. Second, Calyx, a modular compiler infrastructure that transforms high-level languages, like C++, Halide, and PyTorch, and optimizes them to produce efficient hardware designs. Together, these systems represent a first step towards an ecosystem for hardware design, where users seamlessly intermix high- and low-level abstractions, package up reusable components, and implement efficient accelerators. I will conclude by discussing next steps as well as the challenges with the complementary goal of designing programming abstractions for using accelerators.Biography. Rachit Nigam (he/him) is a PhD student at Cornell University and a visiting scholar at MIT working on new programming systems for designing and using hardware accelerators. His research is supported by a Jane Street Fellowship and has been adopted by broad open-source communities such as the LLVM CIRCT project and by industrial teams at Google and Jane Street. Rachit is the founder of PLTea, a virtual, worldwide organization for people interested in programming languages.
Grier Room A
Quest | CBMM Seminar Series: The Debate Over “Understanding” in AI’s Large Language Models
Melanie Mitchell
Santa Fe Institute
Add to Calendar
2024-04-02 16:00:00
2024-04-02 17:30:00
America/New_York
Quest | CBMM Seminar Series: The Debate Over “Understanding” in AI’s Large Language Models
Abstract: I will survey a current, heated debate in the AI research community on whether large pre-trained language models can be said to "understand" language—and the physical and social situations language encodes—in any important sense. I will describe arguments that have been made for and against such understanding, and, more generally, will discuss what methods can be used to fairly evaluate understanding and intelligence in AI systems. I will conclude with key questions for the broader sciences of intelligence that have arisen in light of these discussions. Short Bio: Melanie Mitchell is Professor at the Santa Fe Institute. Her current research focuses on conceptual abstraction and analogy-making in artificial intelligence systems. Melanie is the author or editor of six books and numerous scholarly papers in the fields of artificial intelligence, cognitive science, and complex systems. Her 2009 book Complexity: A Guided Tour (Oxford University Press) won the 2010 Phi Beta Kappa Science Book Award, and her 2019 book Artificial Intelligence: A Guide for Thinking Humans (Farrar, Straus, and Giroux) was shortlisted for the 2023 Cosmos Prize for Scientific Writing.
Buildig 46
Hot Topics in Computing: Beyond information retrieval: what does Search mean these days?
Prabhakar Raghavan
Add to Calendar
2024-04-02 17:00:00
2024-04-02 18:00:00
America/New_York
Hot Topics in Computing: Beyond information retrieval: what does Search mean these days?
TITLE:Beyond information retrieval: what does Search mean these days?ABSTRACT:Classic information retrieval systems aimed to retrieve the documents best matching a user’s query. The advent of the web dramatically changed this landscape and led to the creation of what is considered the modern day search engine. In this talk, we will explore this evolution and the nuanced questions that arise in the operation of a search engine at planetary scale: information quality vs. misinformation; limitations in query understanding, the corpus of content, and information fragmentation; and the economics and value exchange of web search. We are in the early stages of these staples being transformed by the advent of large language models. The future of search lies in the continued refinement of the interplay between evolving technologies, economic models, and user behaviors. We will delve into the new set of opportunities and challenges posed by this shift.BIO:Prabhakar Raghavan is a Senior Vice President at Google. He is responsible for Google’s Knowledge & Information products, including Google Search, News, Assistant, Bard, Geo, Ads, Commerce and Payments. Prabhakar is one of the foremost authorities on Search and is the co-author of two widely-used graduate texts on algorithms and on search: Randomized Algorithms and Introduction to Information Retrieval. He has over 20 years of research spanning algorithms, web search and databases, has published over 100 papers in various fields, and holds 20 issued patents, including several on link analysis for web search.He joined Google in 2012. Prior to his current role, he was Vice President of Google Apps, Google Cloud. Under his leadership, the Apps business expanded from a set of consumer apps to an enterprise solution that is a major contributor to Google’s Cloud business. He grew both Gmail and Drive past 1 billion monthly active users (MAUs) and introduced a number of machine intelligence features in G Suite, including Smart Reply, Smart Compose, and Drive Quick Access — each leading to measurable improvements in user experience. In 2018, he became responsible for the Ads & Commerce teams, including search, display and video advertising, analytics, shopping, payments, and travel. He’s helped drive double-digit growth, while remaining centered on longstanding principles of user trust and fair value exchange among users, publishers, and advertisers.Before joining Google, Prabhakar founded and led Yahoo! Labs where he was responsible for search and ad ranking, as well as ad marketplace design. He also served as CTO at Verity, and held various positions over the course of 14 years at IBM Research, working on algorithms. Prabhakar holds a Ph.D. from U.C. Berkeley in Electrical Engineering and Computer Science and a Bachelor of Technology from the Indian Institute of Technology, Madras. He is a member of the National Academy of Engineering; a Fellow of the ACM and IEEE; a former editor in chief for the Journal of the ACM; and was a Consulting Professor of Computer Science at Stanford University. In 2009, he was awarded a Laurea honoris causa from the University of Bologna.
32-G449
April 04, 2024
EECS Special Seminar: Paul Krogmeier - Learning Symbolic Concepts and Domain-specific Languages
Paul Krogmeier
University of Illinois Urbana-Champaign
Add to Calendar
2024-04-04 11:00:00
2024-04-04 12:00:00
America/New_York
EECS Special Seminar: Paul Krogmeier - Learning Symbolic Concepts and Domain-specific Languages
Abstract: Symbolic languages are fundamental to computing:they help us understand and orchestrate unfamiliar concepts andcomputations in complex domains. Symbolic learning aims to synthesizeconcepts expressed in these languages, e.g., formulas or programs,given a few examples, with many applications in programming, testing,and verification of computer systems. Effective algorithms forsymbolic learning rely on domain-specific heuristics, which makes themhard to build and limits application in new domains.In this talk I will discuss my work on foundations of symboliclearning, which connects language semantics to uniform learningalgorithms via an algorithmic meta-theorem. By writing specializedlanguage interpreters, we are able to effectively describe learningalgorithms and simultaneously prove new theorems about thedecidability of learning in several well-studied symbolic languages incomputer science. With this connection, I will explain how afundamental technique based on version space algebra, as realized inprogram synthesizers from industry, e.g., Microsoft Excel's FlashFill,is in fact an instance of a deeper concept related to tree automata. Iwill discuss how this connection between interpreters and algorithmsuncovers a path to efficient specification and design of symboliclearning algorithms for new domains. I will also discuss my work onlearning logical formulas and applications to visual discriminationand automated discovery of axiomatizations.Finally, I will discuss my work on learning domain-specific languages(DSLs) for few-shot learning, which explores the problem ofconstructing DSLs that balance expressive power, succinctness, andtractability for effective symbolic learning in specific domains. Iwill conclude with some ideas for practically realizing an effectivetranslation from interpreters to learning algorithms and someinteresting applications of symbolic learning to music, math, andmachine learning.Short bio: Paul Krogmeier is a PhD candidate at theUniversity of Illinois Urbana-Champaign. Paul's research is focused onalgorithms for symbolic learning and the problem of learning symboliclanguages and abstractions that capture specific domains. His work onsymbolic learning was recognized with distinguished paper awards atPOPL 2022 and OOPSLA 2023. He has also published in the areas ofprogram synthesis, program verification, and differential privacy.
Grier Room A
End-to-End Encrypted Group Chats with MLS: Design, Implementation and Verification
Théophile Wallez
Inria Paris
Add to Calendar
2024-04-04 12:00:00
2024-04-04 13:00:00
America/New_York
End-to-End Encrypted Group Chats with MLS: Design, Implementation and Verification
Messaging Layer Security (MLS), currently undergoing standardization at the IETF, is an asynchronous group messaging protocol that aims to be efficient for large dynamic groups, while providing strong guarantees like forward secrecy (FS) and post-compromise security (PCS). While prior work on MLS has extensively studied its group key establishment component (called TreeKEM), many flaws in early designs of MLS have stemmed from its group integrity and authentication mechanisms that are not as well-understood.In this work, we identify and formalize TreeSync: a sub-protocol of MLS that specifies the shared group state, defines group management operations, and ensures consistency, integrity, and authentication for the group state across all members. We present a precise, executable, machine-checked formal specification of TreeSync, and show how it can be composed with other components to implement the full MLS protocol. Our specification is written in F* and serves as a reference implementation of MLS; it passes the RFC test vectors and is interoperable with other MLS implementations. Using the DY* symbolic protocol analysis framework, we formalize and prove the integrity and authentication guarantees of TreeSync, under minimal security assumptions on the rest of MLS.Our analysis identifies a new attack and we propose several changes that have been incorporated in the latest MLS draft. Ours is the first testable, machine-checked, formal specification for MLS, and should be of interest to both developers and researchers interested in this upcoming standard.
G-882 Hewlett Room