Add to Calendar
2019-09-23 14:00:00
2019-09-23 15:00:00
America/New_York
Tiramisu: A Polyhedral Compiler for Dense and Sparse Deep Learning
Abstract: Tiramisu is a polyhedral compiler for deep learning. It has two unique features: (1) it is the first sparse DNN compiler; and (2) it can express and optimize general RNNs (Recurrent Neural Networks). Tiramisu relies on a flexible representation based on the polyhedral model and has a rich scheduling language allowing fine-grained control of optimizations. We show that Tiramisu matches or outperforms Intel MKL-DNN by up to 5x for sparse DNNs. We also show that Tiramisu allows many new capabilities such as wavefront parallelization for RNNs.Bio: Riyadh Baghdadi is a postdoctoral associate at CSAIL. He got his PhD from ENS Paris. He works on the area of compiler optimization and code generation for high performance architectures. Recently, he has been interested in exploring the intersection of compilers and deep learning.
32-D463 (Star)
Belfer
sarah_donahue@hks.harvard.edu
Events
September 19, 2019
No events scheduled
September 23, 2019
Feedback-Driven Processors
Professor Joseph Devietti
Department of Computer & Information Science at the University of Pennsylvania
Add to Calendar
2019-09-23 15:30:00
2019-09-23 16:30:00
America/New_York
Feedback-Driven Processors
Abstract:Modern processor architectures are complex machines due to decades' worth of performance optimizations. This complexity can lead to subtle performance issues at the microarchitectural level that are very difficult to diagnose. Fortunately, modern processors also contain a rich set of performance monitoring facilities that enable introspection of many aspects of program execution. My research goal is to build automatic systems that can diagnose and repair performance issues in running programs, relieving programmers of the burden of performance tuning.I'll discuss our work on a series of hardware-software systems for C/C++ and Java that use performance counters to diagnose false sharing bugs in multicore programs, and then perform automatic online repair of these bugs to improve performance through careful consideration of the memory consistency model. These repair techniques are able to achieve much of the benefit of hand-written expert fixes, but require no programmer intervention. In ongoing work we are applying the feedback-driven processor methodology to the problem of front-end stalls in the datacenter. I'll also discuss related efforts to improve performance and correctness via static and dynamic analysis of GPU programs and via deterministic execution systems.Bio:Joseph Devietti is an Assistant Professor in the Department of Computer & Information Science at the University of Pennsylvania. His research focuses on making parallel computers easier to program, leveraging techniques across the computing stack, including computer architecture, compilers, runtime systems and programming languages. He was awarded an Intel Early Career Faculty Honor in 2013, and the Radhia Cousot Young Researcher Best Paper Award at SAS 2018. He earned his PhD in Computer Science & Engineering from the University of Washington in 2012.**Refreshments at 3:15
32-G882
Belfer
sarah_donahue@hks.harvard.edu
Dell EMC Tech Talk: Contextualized AI Models for Multiple Industries, and Building Relationships: Interactions and Graph Networks
Michael Shepherd and Ben Fauber
Dell EMC
Add to Calendar
2019-09-23 16:00:00
2019-09-23 17:00:00
America/New_York
Dell EMC Tech Talk: Contextualized AI Models for Multiple Industries, and Building Relationships: Interactions and Graph Networks
As increasingly broader data domains are contextualized to enhance AI models, we find many industries are challenged to handle the sheer scale of this fast-moving technology shift. At Dell Technologies, our purpose is to drive human progress, through greater access to better technology, for people with big ideas around the world. Accomplishing this means we must be practitioners, researchers and leaders in our own right as AI takes a front row seat in everything we do. From manufacturing to predictive maintenance, supply chain optimization to fraud and sales to support we will provide an overview of how Dell Technology Services (DT Services) AI Research team is applying AI to industry wide problems.Michael will first discuss how CTO classifies projects as AI In, AI On, and AI For Dell and how Services manages and optimizes 65 Million devices with predictive models. Ben will then discuss the contextualization of data, through graph networks, and how graph modeling can enable novel applications, and present new challenges in data mining and scalable machine learning.SPEAKER BIOS: With 25 yrs experience, Michael Shepherd is a Distinguished Engineer and recognized technical evangelist who speaks globally on the impact of emerging technologies and AI. He currently serves as the Lead Technologist for the DT Services AI Research team. Michael sits on the Advisory Council for the University of Texas McCombs MSBA program and is on the core team for Xavier Health’s AI Initiative bringing FDA and industry together to advance AI in Healthcare. Michael has been granted thirteen hardware and software patents in eight countries.Ben Fauber is a senior data scientist at Dell Technologies in Austin, Texas, where he leads university research partnerships for the Applied Data Science Research team in DT Services. He earned his B.Sc. from Colorado State University and Ph.D. from The University of Texas at Austin. Prior to joining Dell Technologies, he held various scientific and leadership positions at Stanley Black & Decker, Genentech/Roche, AstraZeneca, and Pfizer. He is a co-inventor and author on more than 70 patents and peer-reviewed publications, including publications in the journals Nature, and Proceedings of the National Academy of Sciences. He currently serves on advisory boards for The University of Texas at Austin Dell Medical School and McCombs School of Business.Refreshments will be provided. Please RSVP or email cmathews@csail.mit.edu to confirm attendance.
Star D463
Belfer
sarah_donahue@hks.harvard.edu
September 24, 2019
Data Feminism
MIT, DUSP
Add to Calendar
2019-09-24 13:00:00
2019-09-24 14:00:00
America/New_York
Data Feminism
AbstractAs data are increasingly mobilized in the service of global corporations, governments, and elite institutions, their unequal conditions of produc- tion, their inequitable impacts, and their asymmetrical silences become increasingly more apparent. It is precisely this power that makes it worth asking: "Data science by whom? For whom? In whose interest? Informed by whose values?" And most importantly, "How do we begin to imagine alternatives for data’s collection, analysis, and communication?" These are some of the questions that emerge from what Lauren Klein and I call Data Feminism (forthcoming from MIT Press in early 2020). Data feminism is a way of thinking about data science and its products that is informed by the past several decades of intersectional feminist activism and critical thought, emerging anti-oppression design frameworks, and scholarship from the fields of Critical Data Studies, Science & Technology Studies, Geography/GIS, Digital Humanities and Human Computer Interaction. An intersectional feminist lens prompts questions about how, for instance, challenges to the male/female binary can also help challenge other binary (and empirically wrong) classification systems. It encourages us to ask how the concept of invisible labor can help to expose the gendered, racialized, and colonial forms of labor associated with data work. And it demonstrates why the data never, ever, speak for themselves. In this talk, I will introduce seven principles for data feminist work: examining and challenging power, rethinking binaries and hierarchies, considering context, embracing pluralism, making labor visible, and elevating emotion. The goal of this work is to transform scholarship into action – to operationalize feminism in order to imagine more ethical and more equitable data practices.BioCatherine D'Ignazio is a scholar, artist/designer and hacker mama who focuses on feminist technology, data literacy and civic engagement. She has run women's health hackathons, designed global news recommendation systems, created talking and tweeting water quality sculptures, and led walking data visualizations to envision the future of sea level rise. Her forthcoming book from MIT Press, Data Feminism, co-authored with Lauren Klein, charts a course for more ethical and empowering data science practices. Her research at the intersection of technology, design & social change has been published in the Journal of Peer Production, the Journal of Community Informatics, and the proceedings of Human Factors in Computing Systems (ACM SIGCHI). In Jan 2020, D'Ignazio will be an assistant professor of Urban Science and Planning in the Department of Urban Studies and Planning at MIT where she is starting the Data + Feminism Lab.
Star Room (32-D463)
Belfer
sarah_donahue@hks.harvard.edu
László Végh: A Strongly Polynomial Algorithm for Linear Exchange Markets
László Végh, London School of Economics
Add to Calendar
2019-09-24 16:00:00
2019-09-24 17:00:00
America/New_York
László Végh: A Strongly Polynomial Algorithm for Linear Exchange Markets
Abstract: In this talk, I will present the first strongly polynomial algorithm for computing an equilibrium in exchange markets with linear utilities. The exchange market model has been extensively studied since its introduction by Walras in 1874. Our starting point is the previous weakly polynomial combinatorial algorithm by Duan and Mehlhorn. We use a variant of this algorithm to gradually reveal new variables that are in the support of every equilibrium. A key subroutine used for re-initializing this algorithm reduces to solving a linear program (LP). Even though we are unable to solve this LP in strongly polynomial time, we show that it can be approximated by a simpler LP with two variables per inequality that is solvable in strongly polynomial time. This turns out to be good enough to make the overall algorithm run in strongly polynomial time.This is based on joint work with Jugal Garg.
Kiva/G449
Belfer
sarah_donahue@hks.harvard.edu
September 25, 2019
Add to Calendar
2019-09-25 14:00:00
2019-09-25 15:00:00
America/New_York
Digitalisation and European Welfare States
Abstract: Rapid technological progress and innovation can destroy jobs and disrupt welfare systems. This is not a new concern. Historically, automation of production processes has led to extraordinary efficiency gains and to the displacement of labour. But history has also shown that, in the longer run, the gains in efficiency pay off and new jobs are created.But the past is not necessarily a guide to the future. Currently, an unprecedented digitalisation of our economy is underway. Artificial intelligence has become a reality and machines are able learn how to outperform humans in some cognitive tasks. The way work is performed is also changing, with jobs allocated via online platforms and people matched to tasks in a way that means they are neither full-time employees, nor self-employed workers in the traditional sense.For welfare systems, which are largely funded by taxes on employment, these changes have significant implications. One of the big challenges of the twenty-first century will be to redefine the nature and functioning of welfare states in the context of the fundamental changes brought about by digitalisation, artificial intelligence and the changing status of workers. If technology has a negative impact on labour income, how will the welfare state be funded? How can workers’ welfare rights be secured?This volumes tackles these questions and provides recommendations to inform the discussion in the European Union.Bio:J. Scott Marcus, Senior Fellow, Breugel (https://bruegel.org/author/scott-marcus/)J. Scott Marcus is a Senior Fellow at Bruegel, a Brussels-based economics think tank, and also works as an independent consultant dealing with policy and regulatory policy regarding electronic communications. His work is interdisciplinary and entails economics, political science / public administration, policy analysis, and engineering.From 2005 to 2015, he served as a Director for WIK-Consult GmbH (the consulting arm of the WIK, a German research institute in regulatory economics for network industries). From 2001 to 2005, he served as Senior Advisor for Internet Technology for the United States Federal Communications Commission (FCC), as a peer to the Chief Economist and Chief Technologist. In 2004, the FCC seconded Mr. Marcus to the European Commission (to what was then DG INFSO) under a grant from the German Marshall Fund of the United States. Prior to working for the FCC, he was the Chief Technology Officer (CTO) of Genuity, Inc. (GTE Internetworking), one of the world's largest backbone internet service providers.Marcus is the author of numerous papers, a book on data network design. He either led or served as first author for numerous studies for the European Parliament, the European Commission, and national governments and regulatory authorities around the world.
Star Room (32-D463)
Belfer
sarah_donahue@hks.harvard.edu
Add to Calendar
2019-09-25 14:00:00
2019-09-25 15:30:00
America/New_York
Digitalisation and European Welfare States
Rapid technological progress and innovation candestroy jobs and disrupt welfare systems. This is nota new concern. Historically, automation ofproduction processes has led to extraordinaryefficiency gains and to the displacement of labour.But the past is not necessarily a guide to the future.Currently, an unprecedented digitalisation of oureconomy is underway.For welfare systems, which are largely funded bytaxes on employment, these changes havesignificant implications. One of the big challenges ofthe twenty-first century will be to redefine thenature and functioning of welfare states in thecontext of the fundamental changes brought aboutby digitalisation, artificial intelligence and thechanging status of workers.
Belfer
sarah_donahue@hks.harvard.edu
October 01, 2019
Add to Calendar
2019-10-01 9:00:00
2019-10-01 16:00:00
America/New_York
CSAIL Alliances Member Samsung Visit
Belfer
sarah_donahue@hks.harvard.edu
Add to Calendar
2019-10-01 9:00:00
2019-10-01 16:00:00
America/New_York
CSAIL Alliances Member Schlumberger Doll-Research Visit
Belfer
sarah_donahue@hks.harvard.edu
October 02, 2019
Quantum Computing: Current Approaches and Future Prospects
Alphabet X, formerly Google X
Add to Calendar
2019-10-02 11:00:00
2019-10-02 12:00:00
America/New_York
Quantum Computing: Current Approaches and Future Prospects
CBMM Special SeminarAbstract: Jack Hidary will take us through the nascent, but promising field of quantum computing and his new book, Quantum Computing: An Applied Approach.Bio: Jack D. Hidary is a research scientist in quantum computing and in AI at Alphabet X, formerly Google X. He and his group develop and research algorithms for NISQ-regime quantum processors as well as create new software libraries for quantum computing. In the AI field, Jack and his group focus on fundamental research such as the generalization of deep networks as well as applied AI technologies.
46-3002
Belfer
sarah_donahue@hks.harvard.edu
Saving Spaceship Earth
Dava Newman
AeroAstro / MIT
Add to Calendar
2019-10-02 16:00:00
2019-10-02 17:00:00
America/New_York
Saving Spaceship Earth
Abstract:From solar electric propulsion to cutting edge life support systems, advanced space suit design, to the first crops grown in space, the journey to Mars is already unfolding in tangible ways today for tomorrow. However, Mars is not ‘Plan B’. Spaceship Earth, our pale blue dot, is the most magnificent planet to inhabit. Professor Dava Newman, former Deputy Administrator of NASA, MIT Apollo Professor of Astronautics at MIT, and an expert in space technology and policy, offers an orbital view of planet Earth’s interconnected systems through supercomputer data visualizations and stories to demonstrate risks, actions and solutions.Biography: Dr. Dava Newman is the Apollo Program Professor of Astronautics at the Massachusetts Institute of Technology (MIT) in Aeronautics and Astronautics and a Harvard–MIT Health, Sciences, and Technology faculty member. She is a MacVicar Faculty Fellow for excellence and outstanding contributions to undergraduate education. She is the MIT Director of the MIT Portugal Program (MPP2030). Her research expertise is in multidisciplinary aerospace biomedical engineering investigating human performance across the spectrum of gravity. She is a leader in advanced space suit design, dynamics and control of astronaut motion, leadership development, innovation and space policy. Newman was the principal investigator on 4 spaceflight missions. The Space Shuttle Dynamic Load Sensors (DLS) experiment measured astronaut-induced disturbances of the microgravity environment on mission STS-62. An advanced system, the Enhanced Dynamic Load Sensors experiment, flew on board the Russian Mir Space Station from 1996–1998. Dr. Newman was a Co-Investigator on the Mental Workload and Performance Experiment (MWPE) that flew to space on STS-42 to measure astronaut mental workload and fine motor control in microgravity. She also developed the MICR0-G space flight experiment to provide a novel smart sensor suite and study human adaptation in extreme environments. She is the MIT PI on the Gravity Loading Countermeasure Suit, or Skinsuit, onboard the International Space Station as an ESA technology demonstration 2015-2017. Best known for her second skin BioSuit™ planetary EVA system, her advanced spacesuits inventions are now being applied to “soft suits/exoskeletons” to study and enhance locomotion on Earth. Recently, she co-founded EarthDNA with partner Guillermo Trotti to accelerate solutions for spaceship Earth’s Ocean, Land and Air subsystems by curating near-space satellite data to make the world work for 100% of humanity. Newman is the author of Interactive Aerospace Engineering and Design, and has published more than 250 papers in journals and refereed conferences, and holds numerous compression technology patents. She has supervised 90 graduate student theses and supervised and mentored over 200 undergraduate researchers.She served as NASA Deputy Administrator from 2015–2017, and along with the NASA Administrator was responsible for articulating the agency's vision, providing leadership and policy direction, and representing NASA to the White House, Congress, international space agencies, and industry. Dr. Newman was the first female engineer and scientist to serve in this role and was awarded the NASA Distinguished Service Medal. She championed the human journey to Mars, technology and innovation, and education. Recent honors include: Lowell Thomas Award, Phi Beta Kappa Visiting Scholar, AIAA Fellow, AIAA Jeffries Aerospace Medicine and Life Sciences Research Award, and Women in Aerospace Leadership Award.
Patil Seminar Room 32-G449
Belfer
sarah_donahue@hks.harvard.edu
October 03, 2019
TBD
Vitaly Feldman
Google Research
Add to Calendar
2019-10-03 16:00:00
2019-10-03 17:00:00
America/New_York
TBD
E18-304
Belfer
sarah_donahue@hks.harvard.edu
October 08, 2019
EBICS DLS: Mina Bissell
Mina Bissell
Lawrence Berkeley Labs
Add to Calendar
2019-10-08 10:00:00
2019-10-08 12:30:00
America/New_York
EBICS DLS: Mina Bissell
Seminar Room G449 (Patil/Kiva)
Belfer
sarah_donahue@hks.harvard.edu
October 16, 2019
David Patterson: Domain Specific Architectures for Deep Neural Networks: Three Generations of Tensor Processing Units (TPUs)
David Patterson
Univ. of California, Berkeley
Add to Calendar
2019-10-16 16:30:00
2019-10-16 17:30:00
America/New_York
David Patterson: Domain Specific Architectures for Deep Neural Networks: Three Generations of Tensor Processing Units (TPUs)
Abstract:The recent success of deep neural networks (DNN) has inspired a resurgence in domain specific architectures (DSAs) to run them, partially as a result of the deceleration of microprocessor performance improvement due to the ending of Moore’s Law. DNNs have two phases: training, which constructs accurate models, and inference, which serves those models. Google’s first generation Tensor Processing Unit (TPUv1) offered 50X improvement in performance per watt over conventional architectures for inference. We naturally asked whether a successor could do the same for training. This talk reviews TPUv1 and explores how Google built the first production DSA supercomputer for the much harder problem of training, which was deployed in 2017. Google’s TPUv2/TPUv3 supercomputers with up to 1024 chips train production DNNs at close to perfect linear speedup with 10X-40X higher floating point operations per Watt than general-purpose supercomputers running the high-performance computing benchmark Linpack.Bio:David Patterson is a Berkeley CS professor emeritus, a Google distinguished engineer, and the RISC-V Foundation Vice-Chair. He received his BA, MS, and PhD degrees from UCLA. His Reduced Instruction Set Computer (RISC), Redundant Array of Inexpensive Disks (RAID), and Network of Workstation projects helped lead to multibillion-dollar industries. This work led to 40 awards for research, teaching, and service plus many papers and seven books. The best known book is ‘Computer Architecture: A Quantitative Approach,’ and the newest is ‘The RISC-V Reader: An Open Architecture Atlas.’ In 2018 he and John Hennessy shared the ACM A.M. Turing Award.
32-123
Belfer
sarah_donahue@hks.harvard.edu
October 17, 2019
TBD
Kamalika Chaudhuri
University of California, San Diego
Add to Calendar
2019-10-17 16:00:00
2019-10-17 17:00:00
America/New_York
TBD
32-D463 (Stata Center, Star Conference Room)
Belfer
sarah_donahue@hks.harvard.edu
October 24, 2019
TBD
David Duvenaud
University of Toronto
Add to Calendar
2019-10-24 16:30:00
2019-10-24 17:30:00
America/New_York
TBD
35-225
Belfer
sarah_donahue@hks.harvard.edu
October 26, 2019
Add to Calendar
2019-10-26 9:00:00
2019-10-26 16:00:00
America/New_York
CSAIL Alliances Member Mercedes-Benz Experience Week at MIT
Belfer
sarah_donahue@hks.harvard.edu
October 27, 2019
Add to Calendar
2019-10-27 9:00:00
2019-10-27 16:00:00
America/New_York
CSAIL Alliances Member Mercedes-Benz Experience Week at MIT
Belfer
sarah_donahue@hks.harvard.edu
October 28, 2019
Add to Calendar
2019-10-28 9:00:00
2019-10-28 16:00:00
America/New_York
CSAIL Alliances Member Mercedes-Benz Experience Week at MIT
Belfer
sarah_donahue@hks.harvard.edu
Beyond Empirical Risk Minimization: the Lessons of Deep Learning
The Ohio State University
Add to Calendar
2019-10-28 12:00:00
2019-10-28 13:00:00
America/New_York
Beyond Empirical Risk Minimization: the Lessons of Deep Learning
Abstract: "A model with zero training error is overfit to the training data and will typically generalize poorly" goes statistical textbook wisdom. Yet, in modern practice, over-parametrized deep networks with near perfect fit on training data still show excellent test performance. This apparent contradiction points to troubling cracks in the conceptual foundations of machine learning. While classical analyses of Empirical Risk Minimization rely on balancing the complexity of predictors with training error, modern models are best described by interpolation. In that paradigm a predictor is chosen by minimizing (explicitly or implicitly) a norm corresponding to a certain inductive bias over a space of functions that fit the training data exactly. I will discuss the nature of the challenge to our understanding of machine learning and point the way forward to first analyses that account for the empirically observed phenomena. Furthermore, I will show how classical and modern models can be unified within a single "double descent" risk curve, which subsumes the classical U-shaped bias-variance trade-off.Finally, as an example of a particularly interesting inductive bias, I will show evidence that deep over-parametrized autoencoders networks, trained with SGD, implement a form of associative memory with training examples as attractor states.
46-3002
Belfer
sarah_donahue@hks.harvard.edu
October 31, 2019
TBD
Stefano Ermon
Stanford University
Add to Calendar
2019-10-31 16:30:00
2019-10-31 17:30:00
America/New_York
TBD
35-225
Belfer
sarah_donahue@hks.harvard.edu
November 05, 2019
Anindya De: Junta correlation is testable.
Anindya De
Add to Calendar
2019-11-05 16:00:00
2019-11-05 17:00:00
America/New_York
Anindya De: Junta correlation is testable.
Abstract: A Boolean function f on the n-dimensional hypercube is saidto be a k-junta if it is dependent only on some k coordinates of theinput. These functions have been widely studied in the context ofPCPs, learning theory and property testing. In particular, a flagshipresult in property testing is that k-juntas can be tested with\tilde{O}(k) queries -- i.e., there is an algorithm which when givena black box to f, makes only \tilde{O}(k) queries and decides betweenthe following two cases:1. f is a k-junta.2. f is at least 0.01 far from every k-junta g in Hamming distance.Surprisingly, the query complexity is completely independent of theambient dimension n. In this work, we achieve the same qualitativeguarantee for the noise tolerant version of this problem. Inparticular, we give a 2^k query algorithm to distinguish between thefollowing two cases.1. f is 0.48-close to some k-junta.2. f is at least 0.49 far from every k-junta.The algorithm and its proof of correctness are simple and modularemploying some classical tools like "random restrictions" withelementary properties of the noise operator on the cube.Joint work with Elchanan Mossel and Joe Neeman.
Patil Kiva G449
Belfer
sarah_donahue@hks.harvard.edu
November 07, 2019
TBD
Jure Leskovec
Stanford University
Add to Calendar
2019-11-07 16:30:00
2019-11-07 17:30:00
America/New_York
TBD
35-225
Belfer
sarah_donahue@hks.harvard.edu
November 18, 2019
Add to Calendar
2019-11-18 9:00:00
2019-11-18 16:00:00
America/New_York
CSAIL Alliances Member BT Insights
Belfer
sarah_donahue@hks.harvard.edu