Add to Calendar
2021-05-10 15:00:00
2021-05-10 16:00:00
America/New_York
Approximation Algorithms for Fair Clustering
Approximation Algorithms for Fair ClusteringGrowing use of automated machine learning in high-stake decision making tasks has led to an extensive line of research on the societal aspects of algorithms. In particular, the design of fair algorithms for the task of clustering, which is a core problem in both machine learning and algorithms, has received lots of attention in recent years. The fair clustering problem was first introduced in the seminal work of (Chierichetti, Kumar, Lattanzi and Vassilvitskii, 2017) where they proposed the “proportionality/balance inside the clusters” as the notion of fairness. Since then, fairness in clustering has been advanced in this setting and has been further studied in other contexts such as equitable representation and individual fairness.In this talk, I will overview my recent works on the design of efficient approximation algorithms for fair clustering in the three aforementioned contexts. My talk is based on three papers co-authored with Arturs Backurs, Piotr Indyk, Sepideh Mahabadi, Yury Makarychev, Krzysztof Onak, Baruch Schieber, Tal Wagner.Bio: Ali Vakilian is a postdoctoral researcher at the IDEAL institute hosted by TTIC. His research interests include learning-based algorithms, algorithmic fairness and algorithms for massive data. Ali received his PhD from MIT under the supervision of Erik Demaine and Piotr Indyk.
https://mit.zoom.us/j/97637459251
Events
May 08, 2021
No events scheduled
May 10, 2021
May 11, 2021
THESIS DEFENSE: Modeling Intelligence via Graph Neural Networks
Keyulu Xu
MIT CSAIL
Add to Calendar
2021-05-11 9:30:00
2021-05-11 11:00:00
America/New_York
THESIS DEFENSE: Modeling Intelligence via Graph Neural Networks
Abstract: Artificial intelligence is beyond human intelligence. This thesis is about modeling intelligence that can learn to represent and reason about the world. We study both questions from the lens of graph neural networks. First, we can abstract many objects in the world as graphs and learn their representations with graph neural networks. Second, learning to reason implies learning to implement a correct reasoning process, within and outside the training distribution. We shall see how graph neural networks exploit the algorithmic structure in reasoning processes to improve the generalization. Each part of the thesis also studies one aspect of the theoretical landscape of learning: the representation power, generalization, extrapolation, and optimization.I: We build powerful graph neural networks. II: We analyze generalization by considering the training algorithm, network structure, and task structure. III: We study how neural networks extrapolate and show implications for learning reasoning outside the training distribution. IV: We show global convergence rates and practically accelerate the training. Thesis Advisor: Stefanie JegelkaThesis Committee: Stefanie Jegelka, Tommi Jaakkola, Antonio Torralbahttps://mit.zoom.us/j/94720991143
https://mit.zoom.us/j/94720991143
Leveraging IoT to ensure equity in aging: Applications to Parkinson's disease and end-of-life counseling.
University of Rochester
Add to Calendar
2021-05-11 15:00:00
2021-05-11 16:00:00
America/New_York
Leveraging IoT to ensure equity in aging: Applications to Parkinson's disease and end-of-life counseling.
Abstract: By 2030, the old will begin to outnumber the young for the first time in recorded history. Population aging is poised to impose a significant strain on economies, health systems, and social structures. However, it also presents a unique opportunity for IoT to introduce personalization and inclusiveness to ensure equity in aging.Vulnerable populations such as older adults learn, trust, and use new technologies differently. Any prediction algorithm that we develop must use high-quality and population-representative input data outside of the clinic and produce accurate, generalizable, and unbiased results. Therefore, the translational path for AI into clinical care needs deeper engagement with all the stakeholders to ensure that we solve a pressing problem with a practical solution that end-users, clinicians, and patients all find value in.In this talk, I will provide some examples of working systems, evaluated by controlled experiments, and could potentially be deployed in the real world to ensure equity and access among the aging population. In particular, I will highlight two specific examples:1. Innovating for Parkinson's, the fastest-growing neurological disability and is linked with aging. 2. Modeling end-of-life communication with terminal cancer patients where their values and preferences are respected as they plan for a deeply personal human experience such as death. Bio: Ehsan Hoque is an associate professor of computer science at the University of Rochester, where he co-leads the Rochester Human-Computer Interaction (ROC HCI) Group. From 2018-2019, he was also the Interim Director of the Goergen Institute for Data Science. Ehsan earned his Ph.D. from MIT Media Lab in 2013, where the MIT Museum highlighted his dissertation—the development of an intelligent agent to improve human ability—as one of MIT's most unconventional inventions. Building on the work/patent, Microsoft released "Presenter Coach" in 2019 to be integrated into PowerPoint. Ehsan is best known for introducing and extensively validating the idea of using AI to train and enhance elements of basic human ability. Ehsan and his students' work has been recognized by NSF CAREER Award, MIT TR35, Young Investigator Award by the US Army Research Office (ARO). In 2017, Science News named him one of the 10 scientists to watch, and in 2020, the National Academy of Medicine recognized him as one of the emerging leaders in health and medicine. Ehsan is an inaugural member of the ACM's Future of Computing Academy
https://mit.zoom.us/j/98781075598
Something Else About Working Memory
Picower Institute for Learning and Memory, BCS Dept., MIT
Add to Calendar
2021-05-11 16:00:00
2021-05-11 17:30:00
America/New_York
Something Else About Working Memory
Abstract: Working memory is the sketchpad of consciousness, the fundamental mechanism the brain uses to gain volitional control over its thoughts and actions. For the past 50 years, working memory has been thought to rely on cortical neurons that fire continuous impulses that keep thoughts “online”. However, new work from our lab has revealed more complex dynamics. The impulses fire sparsely and interact with brain rhythms of different frequencies. Higher frequency gamma (> 35 Hz) rhythms help carry the contents of working memory while lower frequency alpha/beta (~8-30 Hz) rhythms act as control signals that gate access to and clear out working memory. In other words, a rhythmic dance between brain rhythms may underlie your ability to control your own thoughts.---Zoom link: https://mit.zoom.us/j/96121350408?pwd=ZU1seGNLSWkvS2xBTGM3SlhjaDNXQT09Passcode: 405475
Citadel Securities Tech Talk + Recruiting Event for CSAIL Students 5/11, 5pm EST
Dr. Sören Künzel
Sr. Quant Researcher at Citadel Securities, PhD UC Berkeley
Add to Calendar
2021-05-11 17:00:00
2021-05-11 18:00:00
America/New_York
Citadel Securities Tech Talk + Recruiting Event for CSAIL Students 5/11, 5pm EST
Tech Talk+ Recruitment Event on Tuesday, May 11 @ 5pm ESTRegistration: https://mit.zoom.us/meeting/register/tJAuf-CurjkoH9yD0vETpaYvG9Eu0V57-Qxw
May 12, 2021
MASSIVELY MULTIPLEXED, WHOLE-EMBRYO DEVELOPMENTAL GENETICS AT SINGLE-CELL RESOLUTION
Cole Trapnell
Department of Genome Sciences at the University of Washington.
Add to Calendar
2021-05-12 11:30:00
2021-05-12 13:00:00
America/New_York
MASSIVELY MULTIPLEXED, WHOLE-EMBRYO DEVELOPMENTAL GENETICS AT SINGLE-CELL RESOLUTION
Abstract:Vertebrate genomes encode programs of development that are remarkably robust to intrinsic and environmental noise. However, our understanding of mechanisms underlying this developmental robustness is limited. Recent advances in high-throughput genomics and single-cell RNA-sequencing now allow for comprehensive views of the molecular- and cellular-state of cells in whole developing embryos. To resolve the variation between embryos, we adapted single cell combinatorial indexing (sci)-RNA-seq with oligo-hashing to barcode a cell’s embryo of origin. We deployed this technology to characterize developmental variability during (1) normal zebrafish embryonic development, and (2) development consequent to genetic knockout of 22 essential developmental genes. Together these datasets comprise ~1,700 barcoded embryos across 15 timepoints and nearly 4 million cells. These atlases of wild-type and mutant zebrafish development establish the baseline range of variation in cellular states observed during normal embryonic development and demonstrate how genetic knockouts result in specific alterations in cell-type abundances, cell state changes, and increased variance in specific cell types over time. We expect that the scale and resolution afforded by multiplexed sci-RNA-seq will transform our ability to phenotype developmental processes and necessitate new statistical and quantitative models to describe them, together enabling a mechanistic understanding of developmental robustness across cell types and organisms.
https://mit.zoom.us/j/91842580282
Add to Calendar
2021-05-12 15:00:00
2021-05-12 16:00:00
America/New_York
Parsing numbers at a gigabyte per second
******************IMPORTANT NOTE ABOUT REGISTRATION******************- If you have already registered for any previous Fast Code Seminar on Zoom, please use the Zoom link that you have received in the past. This link will stay the same for all subsequent Fast Code seminars on Zoom. Please save it to your calendar!- Zoom does not recognize a second registration, and will not send out the link a second time. The organizers will not be notified of any second registration.- If you have any problems with registration, please contact lindalynch@csail.mit.edu by 2:30pm on the day of the seminar, so that we can try to resolve it before the seminar begins.*********************************************************************Abstract: Back when disks could barely provide megabytes of bandwidth per second, slow data processing software was acceptable. It is now time to revisit our performance expectations. With disks and networks providing gigabytes per second, parsing decimal numbers from strings becomes a bottleneck. We consider the problem of parsing decimal numbers to the nearest binary floating-point value. We present a C++ implementation that is often 4 times faster than standard C library on modern 64-bit systems (Intel, AMD, ARM and POWER9). Our work is available as open source software used by major systems such as Apache Arrow and Yandex ClickHouse. The Go standard library has adopted a version of our approach. The Rust port of our library is much faster than the state-of-the-art lexical library.Bio: Daniel is a computer science professor at the University of Quebec, and long-time blogger. He is @lemire on Twitter, and he blogs at https://lemire.me/ His 75 research papers have been cited about 4000 times. He is among the top 500 GitHub users in terms of follower count: https://github.com/lemire
https://mit.zoom.us/meeting/register/tJUrdOqopj8uHdO4gUyVMnfglOFEqIye_Je0 (Registration required, only if you haven't registered for this series before; please read IMPORTANT NOTE below)
Add to Calendar
2021-05-12 16:00:00
2021-05-12 17:00:00
America/New_York
Beyond Worst-Case Adversaries in Machine Learning
Beyond Worst-Case Adversaries in Machine LearningAbstract. The widespread application of machine learning in practice necessitates the design of robust learning algorithms that can withstand unpredictable and even adversarial environments. The holy grail of robust algorithm design is to provide performance guarantees that do not diminish in presence of all-powerful adversaries who know the algorithm and adapt to its decisions. However, for most fundamental problems in machine learning and optimization, such guarantees are impossible. This indicates that new models are required to provide rigorous guidance on the design of robust algorithms. This talk goes beyond the worst-case analysis and presents such models and perspectives for fundamental problems in machine learning and optimization. I will also present general-purpose techniques that lead to strong robustness guarantees in practical domains and for a wide range of applications, such as online learning, differential privacy, discrepancy theory, and learning-augmented algorithm design.Bio: Nika Haghtalab is an Assistant Professor in the Department of Electrical Engineering and Computer Sciences at University of California, Berkeley. She works broadly on the theoretical aspects of machine learning and algorithmic economics with a special focus on developing foundations for machine learning that accounts for social and strategic interactions. Prior to Berkeley, she was an assistant professor in the CS department of Cornell University in 2019-2020 and a postdoctoral researcher at Microsoft Research, New England in 2018-2019. She received her Ph.D. from the Computer Science Department of Carnegie Mellon University in 2018. Her thesis titled Foundation of Machine Learning, by the People, for the People received the CMU School of Computer Science Dissertation Award and a SIGecom Dissertation Honorable Mention Award.
https://mit.zoom.us/j/97637459251
May 13, 2021
Mining Software Artifacts for use in Automated Machine Learning
José Pablo Cambronero
MIT-CSAIL
Add to Calendar
2021-05-13 10:00:00
2021-05-13 11:00:00
America/New_York
Mining Software Artifacts for use in Automated Machine Learning
Abstract:Successfully implementing classical supervised machine learning pipelinesrequires that users have software engineering, machine learning, and domainexperience. Machine learning libraries have helped along the first twodimensions by providing modular implementations of popular algorithms. However,implementing a pipeline remains an iterative, tedious, and data-dependent taskas users have to experiment with different pipeline designs.To make the pipeline development process accessible to non-experts and moreefficient for experts, automated techniques can be used to efficiently searchfor high performing pipelines with little user intervention. The collection oftechniques and systems that automate this task are commonly termed automatedmachine learning (AutoML).Inspired by the success of software mining in areas such as code search, programsynthesis, and program repair, we investigate the hypothesis that informationmined from software artifacts can be used to build, improve interactions with,and address missing use cases of AutoML. In particular, I will present threesystems -- AL, AMS, and Janus -- that make use of software artifacts. AL minesdynamic execution traces of a collection of programs that implement machinelearning pipelines and uses these mined traces to learn to produce newpipelines. AMS mines documentation and program examples to automaticallygenerate a search space for an AutoML tool by starting from a user-chosen set ofAPI components. And Janus mines pipeline transformations from a collection ofmachine learning pipelines, which can be used to improve an input pipeline whileproducing a nearby variant. Jointly, these systems and their experimentalresults show that mining software artifacts can simplify AutoML systems, maketheir customization easier, and apply them to novel use cases.Thesis Committee: Martin Rinard (advisor), Saman Amarasinghe and Armando Solar-LezamaLink: https://mit.zoom.us/j/91205512750?pwd=Y1ZKRDd4WEtjNXN0a1BQajBiUERxdz09For Password contact: José Pablo Cambronero, josepablocam@gmail.com or Mary McDavitt, mmcdavit@csail.mit.edu
https://mit.zoom.us/j/91205512750?pwd=Y1ZKRDd4WEtjNXN0a1BQajBiUERxdz09
May 14, 2021
Thesis Defense: Machine Learning and Causality, Building Reliable and Efficient Models for Decision Making
Maggie Makar
MIT CSAIL
Add to Calendar
2021-05-14 12:00:00
2021-05-14 13:00:00
America/New_York
Thesis Defense: Machine Learning and Causality, Building Reliable and Efficient Models for Decision Making
Abstract: We explore relationships between machine learning (ML) and causal inference. We focus on improvements in each by borrowing ideas from one another.ML has been successfully applied to many problems, but the lack of strong theoretical guarantees has led to many unexpected failures. Models that perform well on the training distribution tend to break down when applied to different distributions; small perturbations can "fool" the trained model and drastically change its predictions; arbitrary choices in the training algorithm lead to vastly different models; and so forth. On the other hand, while there has been tremendous progress in developing causal inference methods with strong theoretical guarantees, existing methods typically do not apply in practice since they assume an abundance of data. Working at the intersection of ML and causal inference, we directly address the lack of robustness in ML, and improve the statistical efficiency of causal inference techniques.The motivation behind the work presented in this talk is to improve methods for building predictive, and causal models that are used to guide decision making. We focus mostly on decision making in the healthcare context. On the ML for causality side, we use ML tools and analysis techniques to develop statistically efficient causal models that can guide clinicians when choosing between two treatments. On the causality for ML side, we study how knowledge of the causal mechanisms that generate observed data can be used to efficiently regularize predictive models without introducing biases. In a clinical context, we show how causal knowledge can be used to build robust, and accurate models to predict the spread of contagious infections. In a non-clinical setting, we study how to use causal knowledge to train models that are robust to distribution shifts in the context of image classification.Thesis Advisor: John GuttagThesis Committee: John Guttag (MIT), David Sontag MIT, Jenna Wiens (U. Michigan)Location: https://mit.zoom.us/j/95424351092?pwd=Y2NycTFTcFhMVmZKM2puUzR0dTY2QT09Passcode: 796445
https://mit.zoom.us/j/95424351092?pwd=Y2NycTFTcFhMVmZKM2puUzR0dTY2QT09
May 18, 2021
Unitless Fabrication
Raf Ramakers
Hasselt University
Add to Calendar
2021-05-18 13:00:00
2021-05-18 14:00:00
America/New_York
Unitless Fabrication
With the advent of the industrial revolution in the early 19th century came the need for measurement and precision as parts had to fit together and mass-produced according to tolerances. This quest for precision in industry resulted in new fields i.e. metrology and numerous advanced machines and instruments. Over the past decade, several of these technologies, including digital fabrication machines, have become widely available and accessible to makers, DIY enthusiasts, researchers and educators in various domains. While the wide-spread availability of digital fabrication tools empowers many people outside the field of engineering to experiment with new ideas and prototype physical artifacts themselves, these novel technologies also expose users to very precise specifications, used in industry, for which digital fabrication machines and supporting software tools were originally developed. Measurements and precise specifications are however a major source of errors in fabrication workflows. In this talk, I therefore consider research in fabrication, prototyping, and craft from the perspective of measurement and in particular their user aspects. I also present several novel fabrication techniques that demonstrate how one can fabricate without explicit measurements. Raf Ramakers is an assistant professor in computer science at Hasselt University. His research in Human Computer Interaction takes on an engineering perspective and focuses on digital fabrication and physical computing. More specifically, in his research he builds and investigates novel hardware and software tools to facilitate prototyping physical artifacts and sensor systems. He enjoys working across different disciplines. This resulted in ACM UIST and CHI publications that cross the boundaries of human-computer interaction, electronic engineering, product design, and material science. Raf also regularly serves in program committees of these top-tier venues. Raf received his PhD in computer science in 2016 from Hasselt University. His PhD thesis was awarded both the FWO – IBM innovation award and the FWO-Nokia Bell Scientific Prize. These awards are given annually to the best Belgian PhD dissertation in respectively computer science and communication technology.
https://mit.zoom.us/j/94414460536
May 21, 2021
Add to Calendar
2021-05-21 13:00:00
2021-05-21 14:30:00
America/New_York
Fermi Ma: Post-Quantum Succinct Arguments
Abstract: We prove that Kilian’s four-message succinct argument systemis post-quantum secure in the standard model when instantiated withany probabilistically checkable proof and any collapsing hash function(which in turn exist based on the post-quantum hardness of Learningwith Errors).At the heart of our proof is a general-purpose quantum rewindingprocedure that enables a reduction to repeatedly query a quantumadversary for accepting transcripts as many times as desired.Based on joint work with Alessandro Chiesa, Nicholas Spooner, and Mark Zhandry.https://eprint.iacr.org/2021/334.pdf
email dlehto@mit.edu for Zoom Link
May 25, 2021
Private Event
Machine Learning-centric, Energy-Optimized Wireless Systems, VLSI architectures, and Hardware-Efficient Algorithms
University of Michigan, Ann Arbor
Add to Calendar
2021-05-25 15:00:00
2021-05-25 16:00:00
America/New_York
Machine Learning-centric, Energy-Optimized Wireless Systems, VLSI architectures, and Hardware-Efficient Algorithms
Abstract: This talk presents holistic approaches to realize energy-optimized, machine-learning-centric artificial intelligent of things (AIoT). The optimized system integration is a major challenge in AIoT systems with stringent energy, power, cost, and size constraints. A truly energy-optimal AIoT solution is attainable only by a cross-layer optimization that requires a full characterization of the complete end-to-end system. Addressing this critical technical challenge in emerging ML-centric IoT applications, a cross-layer interdisciplinary research that spans deep learning algorithms, wireless communication, digital signal processing, and VLSI hardware architecture is discussed in this talk. Presented research projects aim for ultra-low power and/or energy-aware wireless IoT systems enabled by novel hardware-friendly algorithms and cross-layer optimized VLSI systems.Bio: Hun-Seok Kim is an assistant professor at the University of Michigan, Ann Arbor. Kim received his B.S. degree from the Seoul National University (South Korea) in 2001, and M.S. & Ph.D. degrees from the University of California, Los Angeles (UCLA), all in Electrical Engineering. His research focuses on system analysis, novel algorithms, and efficient VLSI architectures for low-power/high-performance wireless communication, signal processing, computer vision, and machine learning systems. Before joining the University of Michigan, Kim worked as a technical staff member at Texas Instruments (2010 – 2014). He is serving as an associate editor of IEEE Solid State Circuits Letters, IEEE Transactions on Mobile Computing, and IEEE Transactions on Green Communications & Networking. Kim is a recipient of the 2018 Defense Advanced Research Projects Agency (DARPA) Young Faculty Award (YFA) and the National Science Foundation (NSF) Faculty Early Career Development (CAREER) Award 2019.
https://mit.zoom.us/j/98781075598
May 27, 2021
Network Data Remanence Attacks: When the Secret Sharing-based Schemes Fail to Provide Information Theoretic Security
Leila Rashidi
University of Calgary
Add to Calendar
2021-05-27 14:00:00
2021-05-27 15:00:00
America/New_York
Network Data Remanence Attacks: When the Secret Sharing-based Schemes Fail to Provide Information Theoretic Security
Abstract: With progress toward a practical quantum computer has come an increasingly rapid search for quantum-safe, secure communication schemes that do not rely on discrete logarithm or factorization problems. One such encryption scheme, Multi-path Switching with Secret Sharing (MSSS), combines secret sharing with multi-path switching to achieve security as long as the adversary does not have global observability of all paths and thus cannot capture enough shares to reconstruct messages. MSSS assumes that sending a share on a path is an atomic operation and all paths have the same delay. In this paper, we identify a side-channel vulnerability for MSSS, created by the fact that in real networks, sending a share is not an atomic operation as paths have multiple hops and different delays. This channel, referred to as Network Data Remanence (NDR), is present in all schemes like MSSS whosesecurity relies on transfer atomicity and all paths having same delay. We demonstrate the presence of NDR in a physical testbed. We then identify two new attacks that aim to exploit the side-channel, referred to as NDR Blind and NDR Planned, propose an analytical model to analyze the attacks, and demonstrate them using an implementation of MSSS based on the ONOS SDN controller. Finally, we present a countermeasure for the attacks and show its effectiveness in simulations and Mininet experiments.Zoom: Topic: CSAIL Security SeminarTime: This is a recurring meeting Meet anytimeJoin Zoom Meetinghttps://mit.zoom.us/j/97527284254Password: <3securityOne tap mobile+16465588656,,97527284254# US (New York)+16699006833,,97527284254# US (San Jose)Meeting ID: 975 2728 4254US : +1 646 558 8656 or +1 669 900 6833International Numbers: https://mit.zoom.us/u/auBvg4NEVJoin by SIP97527284254@zoomcrc.comJoin by Skype for Businesshttps://mit.zoom.us/skype/97527284254