Events

February 25, 2019

Towards anonymous and metadata private communication at Internet scale

Albert Kwon PhD Candidate

CSG - CSAIL - MIT

9:30A

- 10:30A

Location

32-G449 (KIVA)

Add to Calendar 2019-02-25 9:30:00 2019-02-25 10:30:00 America/New_York Towards anonymous and metadata private communication at Internet scale As the world becomes more connected, privacy is becoming harder to maintain. From socialmedia services to Internet service providers to state-sponsored mass-surveillance programs,many outlets collect sensitive information about the users and the communication betweenthem – often without the users ever knowing about it. In response, many Internet users haveturned to end-to-end encryption, like Signal and TLS, to protect the content of the communication. Unfortunately, these works do little to hide the metadata of the communication,such as when and with whom a user is communicating. In scenarios where the metadata aresensitive, encryption alone is not sufficient to ensure users’ privacy.Most prior systems that provide metadata private communication fall into one of twocategories: systems that (1) provide formal privacy guarantees against global adversariesbut do not scale to large numbers of users, or (2) scale easily to a large user base butdo not provide strong guarantees against global adversaries. In this thesis, I will presentthree systems that aim to bridge the gap between the two categories to enable privatecommunication with strong guarantees for many millions of users. First, I will present Atom,a horizontally scalable anonymous broadcast system that defends against a global adversarywho monitors the entire network and controls a significant fraction of the servers whilescaling easily by adding more servers to its network. Then, I will present Quark, anotherhorizontally scalable anonymous broadcast system that trades bandwidth for latency toachieve more than an order of magnitude speed-up over Atom under the same threat model.Finally, I will present XRD, which provides metadata private communication between twohonest users against the same adversary using a novel cryptographic primitive. 32-G449 (KIVA) Belfer sarah_donahue@hks.harvard.edu

BLAS-on-Flash: An alternative to large-scale ML training and inference?

Suhas Jayaram Subramanya

Microsoft Research India

10:30A

- 11:30A

Location

32-G882

Add to Calendar 2019-02-25 10:30:00 2019-02-25 11:30:00 America/New_York BLAS-on-Flash: An alternative to large-scale ML training and inference? Abstract: Many large scale machine learning training and inference tasks are memory-bound rather than compute-bound i.e. on large data sets, the working set of these algorithms does not fit in memory for jobs that could run overnight on a few multi-core processors. This often forces an expensive redesign of the algorithm for distributed platforms such as parameter servers and Spark. BLAS-on-flash provides an inexpensive and efficient alternative based on the observation that many ML tasks admit algorithms that can be programmed with linear algebra subroutines. Our library supports a BLAS and sparseBLAS interface on large SSD-resident matrices, enabling multi-threaded code to scale to industrial scale datasets on a single workstation. Using BLAS-on-flash, we are able to process 10x larger models on 10x larger inputs in the same memory envelope in two key production pipelines: training large scale topic models and inference for extreme multi-label learning. This suggests that our approach could be an efficient alternative to expensive distributed big-data systems for scaling up structurally complex machine learning tasks.In this talk, we will take a look at the BLAS-on-flash API, design and implementation of the runtime and the above mentioned case-studies in detail.Relevant Paper:Subramanya, Suhas Jayaram, et al. "BLAS-on-flash: An Efficient Alternative for Large Scale ML Training and Inference?" NSDI 2019. Belfer sarah_donahue@hks.harvard.edu

Hydra: a federated resource manager for data-center scale analytics

Konstantinos Karanasos

Microsoft CISL

1:00P

- 2:00P

Location

Seminar Room G882 (Hewlett Room)

Add to Calendar 2019-02-25 13:00:00 2019-02-25 14:00:00 America/New_York Hydra: a federated resource manager for data-center scale analytics Abstract:Microsoft’s internal data lake processes exabytes of data over millions of cores daily on behalf of thousands of tenants. Scheduling this workload requires 10x to 100x more decisions per second than existing, general-purpose resource management frameworks are known to handle. In 2013, we were faced with a growing demand for workload diversity and richer sharing policies that our legacy system could not meet. In this talk, I will present Hydra, the resource management infrastructure we built to meet these requirements.Hydra leverages a federated architecture, in which a cluster is comprised of multiple, loosely coordinating subclusters. We built Hydra by leveraging, extending, and contributing our code to Apache Hadoop YARN. Hydra is currently the primary big-data resource manager at Microsoft. Over the last few years, Hydra has scheduled nearly one trillion tasks that manipulated close to a Zettabyte of production data. The talk will be based on our upcoming NSDI paper, as well as previous research efforts from our group that led up to Hydra, and will include findings from both open-sourcing Hydra and pushing it to production.Bio:Konstantinos Karanasos is a Principal Scientist at the Cloud and Information Services Lab (CISL) at Microsoft, an applied research group that is part of the Azure Data org. His work at Microsoft has focused on resource management for the company's production analytics clusters and on query optimization for large-scale analytics. More recently, his work has focused on optimizing machine learning workloads. Konstantinos is also a member of the Project Management Committee (PMC) of Apache Hadoop. Within Hadoop, Konstantinos has worked on adding support to YARN for opportunistic containers and for rich placement constraints. Previously, he was a postdoctoral researcher at IBM Almaden Research Center, where he was a member of the Big Data analytics group, working on problems related to query optimization. Konstantinos holds a PhD from Inria and the University Paris-Sud, France, and a Diploma in Electrical and Computer Engineering from the National Technical University of Athens, Greece. Seminar Room G882 (Hewlett Room) Belfer sarah_donahue@hks.harvard.edu

February 27, 2019

Opinion formation, stochastic gradient descent, and gradient-like systems

Fang-Yi Yu

U Michigan

4:00P

- 5:00P

Location

32-G575

Add to Calendar 2019-02-27 16:00:00 2019-02-27 17:00:00 America/New_York Opinion formation, stochastic gradient descent, and gradient-like systems Abstract: We study opinion dynamics on networks with two communities. Eachnode has one of two opinions and updates its opinion as a ``majority-like"function of the frequency of opinions among its neighbors. The networks weconsider are weighted graphs comprised of two equally sized communitieswhere intracommunity edges have weight $p$, and intercommunity edges haveweight $q$. Thus $q$ and $p$ parameterize the connectivity between the twocommunities. We prove a dichotomy theorem about the interaction of the two parameters:1) the ``majority-like" update function, and 2) the level of intercommunityconnectivity. For each set of parameters, we show that either: the systemquickly converges to consensus with high probability in time $\Theta(n\log(n))$; or, the system can get ``stuck" and take time $2^{\Theta(n)}$ toreach consensus. Technically, we achieve this the fast convergence result by exploiting theconnection between a family of reinforced random walks and dynamicalsystems literature. Our main result shows if the systems are a reinforcedrandom walk with a gradient-like function, it converges to an arbitraryneighborhood of a local attracting point in $O(n\log n)$ time with highprobability. This result adds to the recent literature on saddle-pointanalysis and shows a large family of stochastic gradient descent algorithmconverges to local minimal in $O(n\log n)$ when the step size $O(1/n)$. 32-G575 Belfer sarah_donahue@hks.harvard.edu

March 04, 2019

Why use Artificial Intelligence in Big Data Analysis in Neurosurgery?

Juri Kivelev

3:00P

- 4:00P

Location

32-D463 Star

Add to Calendar 2019-03-04 15:00:00 2019-03-04 16:00:00 America/New_York Why use Artificial Intelligence in Big Data Analysis in Neurosurgery? In 2019, we estimate severity of intracranial hemorrhage by the Hunt-Hess scale established in 1968. Moreover, in 2019, we estimate the level of consciousness in a severe brain trauma by the Glasgow Coma Scale established in 1974. These are just a couple of examples of what neurosurgeons are still using in clinical practice worldwide. These scales were established before the computed tomography era and are based on the very primitive statistical analysis of limited clinical data.Now, in Finland we have access to data encompassing billions of bits of clinical parameters from consecutive unbiased patient series. This is due to reliability of the Finnish social insurance system and a novel automated way to collect the data from electronic patient records.Our department has a comprehensive dataset of almost 4,000 patients having a brain aneurysm treated between 2000 and 2018 in our district. The dataset includes nearly 3,000,000 laboratory exam results and 230,000 clinical events with no drop-outs. To our knowledge, this kind of dataset is absolutely unique in the world and worthy of the best possible modern analysis.We need artificial intelligence to exploit these data for scientifically justified decision-making, thus providing more reliable patient care. The essential path of development in neurosurgery seems to be the symbiosis of machine learning and clinical practice. 32-D463 Star Belfer sarah_donahue@hks.harvard.edu

On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization

Jason D. Lee

University of Southern California

4:00P

- 5:00P

Location

32-G449, Patil/Kiva

Add to Calendar 2019-03-04 16:00:00 2019-03-04 17:00:00 America/New_York On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization AbstractDeep Learning has had phenomenal empirical successes in many domains including computer vision, natural language processing, and speech recognition. To consolidate and boost the empirical success, we need to develop a more systematic and deeper understanding of the elusive principles of deep learning.In this talk, I will provide theoretical analysis of several elements of deep learning including non-convex optimization, overparametrization, and generalization error. First, we show that gradient descent and many other algorithms are guaranteed to converge to a local minimizer of the loss. For several interesting problems including the matrix completion problem, this guarantees that we converge to a global minimum. Then we will show that gradient descent converges to a global minimizer for deep overparametrized networks. Finally, we analyze the generalization error by showing that a subtle combination of SGD, logistic loss, and architecture combine to promote large margin classifiers, which are guaranteed to have low generalization error. Together, these results show that on overparametrized deep networks SGD finds solution of both low train and test error.BioJason Lee is an assistant professor in Data Sciences and Operations at the University of Southern California. Prior to that, he was a postdoctoral researcher at UC Berkeley working with Michael Jordan. Jason received his PhD at Stanford University advised by Trevor Hastie and Jonathan Taylor. His research interests are in statistics, machine learning, and optimization. Lately, he has worked on high dimensional statistical inference, analysis of non-convex optimization algorithms, and theory for deep learning. 32-G449, Patil/Kiva Belfer sarah_donahue@hks.harvard.edu

March 06, 2019

Dertouzos Distinguished Lecture

Prof. John Hennessy

4:30P

- 5:30P

Location

32-123

Stata Center, Kirsch Auditorium

Add to Calendar 2019-03-06 16:30:00 2019-03-06 17:30:00 America/New_York Dertouzos Distinguished Lecture Abstract: After 40 years of remarkable progress in VLSI microprocessors, a variety of factors are combining to lead to a much slower rate of performance growth in the future. These limitations arise from three different areas: IC technology, architectural limitations, and changing applications and usage. The end of Dennard scaling and the slowdown in Moore's Law will require more efficient architectural approaches than we have relied on. Although progress on general-purpose processors may hit an asymptote, domain specific architectures may be one attractive path for important classes of problems. Such an approach will pose new challenges for software and chip designers, as well as increase the need for more advanced design tools.Bio: John L. Hennessy, Professor of Electrical Engineering and Computer Science, served as President of Stanford University from September 2000 until August 2016. In 2017, he initiated the Knight-Hennessy Scholars Program, the largest fully endowed graduate-level scholarship program in the world, and he currently serves as Director of the program. Hennessy joined Stanford’s faculty in 1977. In 1981, he drew together researchers to focus on a technology known as RISC (Reduced Instruction Set Computer), which revolutionized computing by increasing performance while reducing costs. Hennessy helped transfer this technology to industry cofounding MIPS Computer Systems in 1984. He served as chair of Computer Science, dean of the School of Engineering, and university provost before being appointed as Stanford’s 10th president. As president he focused on increasing financial aid and on developing new initiatives in multidisciplinary research and teaching. He was the founding board chair of Atheros Communications, one of the early developers of WiFi technology, and has served on the board of Cisco and Alphabet (Google’s parent company). He is the coauthor of two internationally used textbooks in computer architecture.His honors include the 2012 Medal of Honor of the Institute of Electrical and Electronics Engineers and the ACM Turing Award (jointly with David Patterson). He is an elected member of the National Academy of Engineering, the National Academy of Science, the American Academy of Arts and Sciences, The Royal Academy of Engineering, and the American Philosophical Society. Hennessy earned his bachelor’s degree in electrical engineering from Villanova University and his master’s and doctoral degrees in computer science from the Stony Brook University. 32-123 Belfer sarah_donahue@hks.harvard.edu

March 07, 2019

JP Morgan Tech Talk: Privacy Preserving Predictions

Antigoni Polychroniadou

JP Morgan

12:00P

- 1:00P

Location

Star D463

Add to Calendar 2019-03-07 12:00:00 2019-03-07 13:00:00 America/New_York JP Morgan Tech Talk: Privacy Preserving Predictions While big data technology offers great promise, it introduces a challenge for privacy. Can we achieve the best of both worlds, extracting the benefits that big data offers while providing privacy to the data owner? Essentially, we would like to perform complex computations while preserving privacy. In this talk I am going to talk about privacy preserving computation and show how cryptography, streaming data and machine learning can support new ways of organizing predictive analytics. Presenter Bio:Antigoni Polychroniadou is a researcher at J.P. Morgan AI research and cryptographer lead in ROAR Data at J.P. Morgan. Antigoni was a junior Simons fellow, awarded by the Simons Society of Fellows, at Cornell Tech and a postdoctoral researcher in the Computer Science Department at Cornell University, hosted by Rafael Pass and Elaine Shi. Antigoni completed her Ph.D. at Aarhus University under the supervision of Ivan Damgård. She interned at IDC Herzliya, the Technion, University of California, Berkeley and IBM Research Thomas J. Watson. She holds an M.Sc. in mathematics of cryptography and communications from Royal Holloway University of London and B.Sc. in Computer Science and Economics from University of Macedonia, Greece.Please register OR email Callie Mathews, cmathews@csail.mit.edu to confirm attendance. Belfer sarah_donahue@hks.harvard.edu

AI for Imperfect-Information Game Settings

Noam Brown

Carnegie Mellon University

4:00P

- 5:00P

Location

32-G449 Patil/Kiva

Add to Calendar 2019-03-07 16:00:00 2019-03-07 17:00:00 America/New_York AI for Imperfect-Information Game Settings Abstract: The field of artificial intelligence has had a number of high-profile successes in the domain of perfect-information games. But real-world strategic interactions are not like chess or Go where all participants know the exact state of the world. Instead, they typically involve hidden information, such as in negotiations, cybersecurity attacks, and financial markets. Techniques used in perfect-information games fall apart when applied to such imperfect-information games, with poker serving as the classic example. Libratus is an AI that, in a 120,000-hand competition, decisively defeated four top professionals in heads-up no-limit Texas hold’em poker, the leading benchmark for imperfect-information games and a long-standing challenge problem for AI in general. In this talk I will explain why past techniques intended for perfect-information multi-agent and imperfect-information single-agent settings break down both in theory and in practice in imperfect-information multi-agent settings, and the advances in Libratus and my later work that overcome those challenges. In particular, I will describe new general methods I developed for state-of-the-art equilibrium finding and real-time planning in imperfect-information games. These techniques all have theoretical guarantees in addition to strong empirical performance, and are domain-independent. I will conclude by discussing applications of this work and future research directions in the area of multi-agent artificial intelligence. Host: Tommi Jaakkola.Bio: Noam Brown is a PhD candidate in computer science at Carnegie Mellon University advised by Tuomas Sandholm and a Research Scientist at Facebook AI Research. His research combines computational game theory and machine learning to develop AI systems capable of strategic reasoning in large imperfect-information multi-agent settings. He has applied this research to creating Libratus, the first AI to defeat top humans in no-limit poker, which was one of 12 finalists for Science Magazine's Scientific Breakthrough of the Year. Noam received a NeurIPS Best Paper award in 2017, the 2017 Allen Newell Award for Research Excellence, an Outstanding Paper Honorable Mention at AAAI 2019, and the 2019 Marvin Minsky Medal for Outstanding Achievements in AI. His PhD was supported by an Open Philanthropy Project AI fellowship and a Tencent AI Lab fellowship. 32-G449 Patil/Kiva Belfer sarah_donahue@hks.harvard.edu

March 11, 2019

EECS Special Seminar: Using Computer Vision to Study Society: Methods and Challenges

Timnit Gebru

Google

4:00P

- 5:00P

Location

32-G449

Patil/Kiva

Add to Calendar 2019-03-11 16:00:00 2019-03-11 17:00:00 America/New_York EECS Special Seminar: Using Computer Vision to Study Society: Methods and Challenges Abstract: Targeted socio-economic policies require an accurate understanding of a country's demographic makeup. To that end, the United States spends more than 1 billion dollars a year gathering census data such as race, gender, education, occupation and unemployment rates. Compared to the traditional method of collecting surveys across many years which is costly and labor intensive, data-driven, machine learning driven approaches are cheaper and faster--with the potential ability to detect trends in close to real time. In this work, we leverage the ubiquity of Google Street View images and develop a computer vision pipeline to predict income, per capita carbon emission, crime rates and other city attributes from a single source of publicly available visual data. We first detect cars in 50 million images across 200 of the largest US cities and train a model to determine demographic attributes using the detect cars. To facilitate our work, we used a graph based algorithm to collect a challenging fine-grained dataset consisting of over 2600 classes of cars comprised of images from Google Street View and other web sources. Our prediction results correlate well with ground truth income (r=0.82), race, education, voting, sources investigating crime rates, income segregation, per capita carbon emission, and other market research. Data mining based works such as this one can be used for many types of applications--some ethical and others not. I will finally discuss work (inspired by my experiences while working on this project), on auditing and exposing biases found in computer vision systems. Using recent work on exposing the gender and skin type bias found in commercial gender classification systems as a case study, I will discuss how the lack of standardization and documentation in AI is leading to biased systems used in high stakes scenarios. I will end with the concept of AI datasheets for datasets, and model cards for model reporting to standardize information for datasets and pre-trained models, to push the field as a whole towards transparency and accountability. Host: Antonio Torralba.Bio: Timnit Gebru is a research scientist in the Ethical AI team at Google and just finished her postdoc in the Fairness Accountability Transparency and Ethics (FATE) group at Microsoft Research, New York. Prior to that, she was a PhD student in the Stanford Artificial Intelligence Laboratory, studying computer vision under Fei-Fei Li. Her main research interest is in data mining large-scale, publicly available images to gain sociological insight, and working on computer vision problems that arise as a result, including fine-grained image recognition, scalable annotation of images, and domain adaptation. She is currently studying the ethical considerations underlying any data mining project, and methods of auditing and mitigating bias in sociotechnical systems. The New York Times, MIT Tech Review and others have recently covered her work. As a cofounder of the group Black in AI, she works to both increase diversity in the field and reduce the negative impacts of racial bias in training data used for human-centric machine learning models. 32-G449 Belfer sarah_donahue@hks.harvard.edu

March 13, 2019

Discover[i]: Component-based Parameterized Reasoning for Distributed Applications

Roopsha Samanta

Purdue University

11:00A

- 12:00P

Location

Seminar Room G882 (Hewlett Room)

Add to Calendar 2019-03-13 11:00:00 2019-03-13 12:00:00 America/New_York Discover[i]: Component-based Parameterized Reasoning for Distributed Applications Distributed systems are hard to get right. There have been many notable efforts in formal reasoning for distributed systems: these efforts have focused on language design, automated or semi-automated verification, and, more recently, on automated synthesis for systems with a fixed number of processes. Our Discover[i] project seeks to automate aspects of programming distributed applications by developing new foundations, algorithms and tools for synthesis/verification of distributed applications built using (verified) components and parameterized by the number of processes.In this talk, I will present an approach for parameterized synthesis of coordination for distributed systems that use consensus protocols, such as Paxos, as a building block to provide higher-level functionality. Our approach is based on (1) encapsulating the details of consensus into a simple atomic primitive, (2) a decidability result for parameterized verification of safety properties of systems with such consensus primitives, and (3) decision procedures for parameterized synthesis. We have used our tool to synthesize coordination for several examples of distributed applications, including a model of the Small Aircraft Tracking System.Bio: Roopsha Samanta is an Assistant Professor in the Department of Computer Science at Purdue University. Her research mission is to make it easier for people to build provably reliable programs. Her research focuses on developing algorithms and tools for automated program repair and synthesis, and targets diverse application domains such as concurrent and distributed systems, personalized education and machine learning.Roopsha completed her Ph.D. at The University of Texas at Austin in 2013 and was a postdoctoral researcher at the Institute of Science and Technology Austria (IST Austria) from 2014-2016. Seminar Room G882 (Hewlett Room) Belfer sarah_donahue@hks.harvard.edu

March 18, 2019

EECS Special Seminar: The Next Billion Medical Devices

Edward Wang

University of Washington

4:00P

- 5:00P

Location

32-G449

Patil/Kiva

Add to Calendar 2019-03-18 16:00:00 2019-03-18 17:00:00 America/New_York EECS Special Seminar: The Next Billion Medical Devices Abstract:Medical care today revolves mainly around the clinic, with devices and care designed with accuracy as the main metric of success. Many medical tests can be accomplished with just a small vial of blood, and in the ICU we can probe a variety of physiological parameters continuously. In the search for accuracy, accessibility, however, has often taken a backseat. In the majority of situations outside of clinical care, there will not be access to that vial of blood or the myriads of sensors on an ICU monitor. Grounded in this premise, my research looks to invent the next generation of "mobile medical devices" that enables medical sensing beyond the current clinical paradigm in two major directions. First, I will discuss how we can re-appropriate existing mobile sensor infrastructure to enable low-cost and widespread medical sensing. The second major theme of my work explores novel wearable sensors to perform continuous health tracking for passive, long-term insights to a person’s health. Through this talk, I will discuss how I have enabled smartphones to perform physiological measurements comparable to their medical device counterparts and developed novel wearable devices that continuously track blood pressure and user activity in a low-power and highly wearable fashion. My work has taken me from the lab bench building prototypes, clinical validations against state-of-the-art medical devices, to working with actual users and patients at villages in the jungles of South America. Finally, I will discuss my future explorations into largescale data-driven development of accessible mobile health systems that will aim to negotiate scale, human-in-the-loop effects, privacy, and accuracy. Bio:Edward Wang is a Ph.D. candidate in Electrical and Computer Engineering in the Ubiquitous Computing (UbiComp) Lab at the University of Washington, advised by Shwetak Patel. His research work explores practical solutions to address real-world medical needs drawn from collaborations with clinicians and world health organizations, but solved in new, creative ways that leverage state-of-the-art applied machine learning, embedded systems, and mobile sensors. He was awarded the NSF Graduate Fellowship, the ARCS Foundation Fellowship, and 3 paper awards during his time at the University of Washington. He received a B.Sci in Engineering from Harvey Mudd College in 2012, where he was advised by Elizabeth Orwin and specialized in biomedical engineering. 32-G449 Belfer sarah_donahue@hks.harvard.edu

April 26, 2019

Communicating gradients for learning using activity dynamics

Blake Richards

Canadian Institute for Advanced Research (CIFAR)

4:00P

- 5:00P

Location

Singleton Auditorium (46-3002)

MIT Bldg 46-3002, 43 Vassar Street, Cambridge MA 02139

Add to Calendar 2019-04-26 16:00:00 2019-04-26 17:00:00 America/New_York Communicating gradients for learning using activity dynamics Abstract: Theoretical and empirical results in the neural networks literature demonstrate that effective learning at a real-world scale requires changes to synaptic weights that approximate the gradient of a global loss function. For neuroscientists, this means that the brain must have mechanisms for communicating loss gradients between regions, either explicitly or implicitly. Here, I describe our research into potential means of communicating loss gradients using the dynamics of activity in a population of neurons. Using a combination of computational modelling and two-photon imaging data, I will present evidence suggesting that the neocortex may encode loss gradients related to motor learning and sensory prediction using the temporal derivative of activity in populations of pyramidal neurons. Singleton Auditorium (46-3002) Belfer sarah_donahue@hks.harvard.edu