Add to Calendar
2019-04-22 11:00:00
2019-04-22 12:00:00
America/New_York
Imaging a Black Hole with the Event Horizon Telescope
Abstract: This talk will present the methods and procedures used to produce the first results from the Event Horizon Telescope. It is theorized that a black hole will leave a "shadow" on a background of hot gas. Taking a picture of this black hole shadow could help to address a number of important scientific questions, both on the nature of black holes and the validity of general relativity. Unfortunately, due to its small size, traditional imaging approaches require an Earth-sized radio telescope. In this talk, I discuss techniques we have developed to photograph a black hole using the Event Horizon Telescope, a network of telescopes scattered across the globe. Imaging a black hole’s structure with this computational telescope requires us to reconstruct images from sparse measurements, heavily corrupted by atmospheric error.Webcast: http://web.mit.edu/webcast/csail/1917/
32-123
Belfer
sarah_donahue@hks.harvard.edu
Events
April 22, 2019
April 24, 2019
Learning with Generalized Negative Dependence: Probabilistic Models of Diversification for Machine Learning
CSAIL and LIDS
Add to Calendar
2019-04-24 11:00:00
2019-04-24 12:00:00
America/New_York
Learning with Generalized Negative Dependence: Probabilistic Models of Diversification for Machine Learning
This thesis establishes negative dependence as a powerful and computationally efficient framework to analyze machine learning problems that require a theoretical model of diversification. Examples of such problems include experimental design and model compression: subset-selection problems that require carefully balancing the quality of each selected element with the diversity of the subset as a whole. Negative dependence, which models the behavior of “repelling” random variables, provides a rich mathematical framework for the analysis of such problems.The first part of this thesis develops scalable sampling and learning algorithms for determinantal point processes (DPPs), popular negatively dependent measures with many applications to machine learning. We introduce a theoretically-motivated generative deep network for DPP-like samples over arbitrary ground sets. To address the learning problem, we show that maximum likelihood estimation for Dpps is drastically sped up with Kronecker kernels, and is further enriched by negative samples.The second part of this thesis leverages negative dependence for core problems in machine learning. We begin by deriving a generalized form of volume sampling (GVS) based on elementary symmetric polynomials, and prove that the induced measures exhibit strong negative dependence properties. We then show that classical forms of optimal experimental design can be cast as optimization problems based on GVS, for which we derive randomized and greedy algorithms to obtain the associated designs. Finally, we introduce exponentiated strongly Rayleigh measures, which allow for simple tuning of the strength of repulsive forces between similar items while still enjoying fast sampling algorithms. The great flexibility of exponentiated strongly Rayleigh measures makes them an ideal tool for machine learning problems that benefit from negative dependence theory.
32D-463 (Star)
Belfer
sarah_donahue@hks.harvard.edu
Mutational signature analysis and its applications to the clinic
Peter Park
Harvard University
Add to Calendar
2019-04-24 11:30:00
2019-04-24 13:00:00
America/New_York
Mutational signature analysis and its applications to the clinic
Different mutational processes operative in cancer and other diseases leave distinct 'signatures' in the DNA. Mutational signature analysis is an attempt to deconvolve the mutational patterns from cancer sequencing data to better identify the factors that gave rise to cancer. Whereas previous work required a large amount of signal as found in exome and genome sequencing data, our new method SigMA enables accurate detection of mutational signatures even with >100-fold reduction in data size. This allows us to extend signature analysis to gene panels, the common platform used to profile tens of thousands of cancer patients each year. I will describe the methodology behind SigMA and how it can be used to identify patients with deficiency in the homologous recombination DNA repair pathway who should be considered for treatment with PARP inhibitors. This work was led by Dr. Doga Gulhan (PhD in heavy ion physics, MIT)Bio: Dr. Park is Professor of Biomedical Informatics at Harvard Medical School and the director of its Bioinformatics and Integrative Genomics Ph.D. program. His group (http://compbio.hms.harvard.edu) specializes in computational and statistical analysis of high-throughput sequencing data in epigenetics, cancer genetics, and neuroscience. He was originally trained in applied math (B.A., Harvard; Ph.D., Caltech), but he stumbled upon molecular biology and genetics during his postdoctoral studies. He has multiple positions open in his group for students, postdoctoral fellows, and scientific programmers.
32-G575
Belfer
sarah_donahue@hks.harvard.edu
Gauge Fields in Deep Learning
Max Welling
University of Amsterdam
Add to Calendar
2019-04-24 14:00:00
2019-04-24 15:00:00
America/New_York
Gauge Fields in Deep Learning
Abstract:Gauge field theory is the foundation of modern physics, including general relativity and the standard model of physics. It describes how a theory of physics should transform under symmetry transformations. For instance, in electrodynamics, electric forces may transform into magnetic forces if we transform a static observer to one that moves at constant speed. Similarly, in general relativity acceleration and gravity are equated to each other under symmetry transformations. Gauge fields also play a crucial role in modern quantum field theory and the standard model of physics, where they describe the forces between particles that transform into each other under (abstract) symmetry transformations. In this work we describe how the mathematics of gauge groups becomes inevitable when you are interested in deep learning on manifolds. Defining a convolution on a manifold involves transporting geometric objects such as feature vectors and kernels across the manifold, which due to curvature become path dependent. As such it becomes impossible to represent these objects in a global reference frame and one is forced to consider local frames. These reference frames are arbitrary and changing between them is called a (local) gauge transformation. Since we do not want our computations to depend on the specific choice of frames we are forced to consider equivariance of our convolutions under gauge transformations. These considerations result in the first fully general theory of deep learning on manifolds, with gauge equivariant convolutions as the key ingredient.Joint work with Taco Cohen, Maurice Weiler and Berkay KicanaogluBio:Prof. Dr. Max Welling is a research chair in Machine Learning at the University of Amsterdam and a VP Technologies at Qualcomm. He has a secondary appointment as a senior fellow at the Canadian Institute for Advanced Research (CIFAR). He is co-founder of “Scyfer BV” a university spin-off in deep learning which got acquired by Qualcomm in summer 2017. In the past he held postdoctoral positions at Caltech (’98-’00), UCL (’00-’01) and the U. Toronto (’01-’03). He received his PhD in ’98 under supervision of Nobel laureate Prof. G. ‘t Hooft. Max Welling has served as associate editor in chief of IEEE TPAMI from 2011-2015 (impact factor 4.8). He serves on the board of the NIPS foundation since 2015 (the largest conference in machine learning) and has been program chair and general chair of NIPS in 2013 and 2014 respectively. He was also program chair of AISTATS in 2009 and ECCV in 2016 and general chair of MIDL 2018. He has served on the editorial boards of JMLR and JML and was an associate editor for Neurocomputing, JCGS and TPAMI. He received multiple grants from Google, Facebook, Yahoo, NSF, NIH, NWO and ONR-MURI among which an NSF career grant in 2005. He is recipient of the ECCV Koenderink Prize in 2010. Welling is in the board of the Data Science Research Center in Amsterdam, he directs the Amsterdam Machine Learning Lab (AMLAB), and co-directs the Qualcomm-UvA deep learning lab (QUVA) and the Bosch-UvA Deep Learning lab (DELTA).
32-G449 (Stata Center - Patil/Kiva Conference Room)
Belfer
sarah_donahue@hks.harvard.edu
Add to Calendar
2019-04-24 15:00:00
2019-04-24 16:00:00
America/New_York
Thesis Defense: Itay Berman: Information Theoretic Advances in Zero-Knowledge
Abstract:Zero-knowledge proofs have an intimate relation to notions from information�theory. In particular, the class of all problems possessing statistical�zero-knowledge proofs (SZK) was shown to have complete problems characterized by�the statistical distance (Sahai and Vadhan [JACM, 2003]) and entropy difference�(Goldreich and Vadhan [CCC, 1999]) of a pair of efficiently samplable�distributions. This characterization has been extremely beneficial in�understanding the computational complexity of languages with zero-knowledge�proofs and deriving new applications from such languages.��In this thesis, we further study the relation between zero-knowledge proofs and�information theory. Our main results are:��1. Two additional complete problems for SZK characterized by other information� theoretic notions---triangular discrimination and Jensen-Shannon� divergence. These new complete problems further expand the regime of� parameters for which the Statistical Difference Problem is complete for SZK.��2. The hardness of a problem related to the non-interactive variant of the� Entropy Difference Problem implies the existence of a useful cryptographic� primitive called multi-collision resistant hash functions (MCRH).��3. We initiate the study of zero-knowledge in the model of interactive proofs of� proximity (IPP). We show efficient zero-knowledge IPPs for several� problems. We also show that not every efficient IPP can be made� zero-knowledge.�
D507
Belfer
sarah_donahue@hks.harvard.edu
Adversarially Robust Property Preserving Hashes
Rio LaVigne
MIT CSAIL
Add to Calendar
2019-04-24 16:00:00
2019-04-24 17:00:00
America/New_York
Adversarially Robust Property Preserving Hashes
Abstract: Property-preserving hashing is a method of compressing a large input x into a short hash h(x) in such a way that given h(x) and h(y), one can compute a property P(x,y) of the original inputs. The idea of property-preserving hash functions underlies sketching, compressed sensing and locality-sensitive hashing.Property-preserving hash functions are usually probabilistic: they use the random choice of a hash function from a family to achieve compression, and as a consequence, err on some inputs. Traditionally, the notion of correctness for these hash functions requires that for every two inputs x and y, the probability that h(x) and h(y) mislead us into a wrong prediction of P(x,y) is negligible. As observed in many recent works (incl. Mironov, Naor and Segev, STOC 2008; Hardt and Woodruff, STOC 2013; Naor and Yogev, CRYPTO 2015), such a correctness guarantee assumes that the adversary (who produces the offending inputs) has no information about the hash function, and is too weak in many scenarios.We initiate the study of adversarial robustness for property-preserving hash functions, provide definitions, derive broad lower bounds due to a simple connection with communication complexity, and show the necessity of computational assumptions to construct such functions. Our main positive results are two candidate constructions of property-preserving hash functions (achieving different parameters) for the (promise) gap-Hamming property which checks if x and y are “too far” or “too close.” Our first construction relies on generic collision-resistant hash functions, and our second on a variant of the syndrome decoding assumption on low-density parity.Joint work with Elette Boyle and Vinod Vaikuntanathan.
32-G882 (Hewlett Room)
Belfer
sarah_donahue@hks.harvard.edu
CyberWork and the American Dream
Elizabeth Cobbs and Jim Shelley
Add to Calendar
2019-04-24 16:30:00
2019-04-24 18:00:00
America/New_York
CyberWork and the American Dream
In "CyberWork and the American Dream", a new 56-minute documentary, a look at the past gives hope for the future: Throughout history, new technology has always caused problems in labor markets, and humanity always finds innovative ways to persevere. Join Elizabeth Cobb, writer, and Jim Shelley, director of the documentary for a screening and discussion on A.I. and the future of work moderated by Daniela Rus.Light refreshments will be served by 4:15pm, and the program starts at 4:30.
32-G449
Belfer
sarah_donahue@hks.harvard.edu
April 25, 2019
Thesis Defense: Haoyang Zeng, "Machine Learning Models for Functional Genomics and Therapeutic Design"
MIT CSAIL - Gifford Lab
Add to Calendar
2019-04-25 10:00:00
2019-04-25 12:00:00
America/New_York
Thesis Defense: Haoyang Zeng, "Machine Learning Models for Functional Genomics and Therapeutic Design"
Abstract: Due to the limited size of training data available, machine learning models for biology have remained rudimentary and inaccurate despite the significant advance in machine learning research. With the recent advent of high-throughput sequencing technology, an exponentially growing number of genomic and proteomic datasets have been generated. These large-scale datasets admit the training of high-capacity machine learning models to characterize sophisticated features and produce accurate predictions on unseen examples. In this thesis, we attempt to develop advanced machine learning models for functional genomics and therapeutics design, two areas with ample data deposited in public databases and tremendous clinical implications. The shared theme of these models is to learn how the composition of a biological sequence encodes a functional phenotype and then leverage such knowledge to provide insight for target discovery and therapeutic design.First, we design three machine learning models that predict transcription factor binding and DNA methylation, two fundamental epigenetic phenotypes closely tied to gene regulation, from DNA sequence alone. We show that these epigenetic phenotypes can be well predicted from the sequence context. Moreover, the predicted change in phenotype between the reference and alternate allele of a genetic variant accurately reflect its functional impact and improves the identification of regulatory variants causal for complex diseases.Second, we devise two machine learning models that improve the prediction of peptides displayed by the major histocompatibility complex (MHC) on the cell surface. Computational modeling of peptide-display by MHC is central in the design of peptide-based therapeutics. Our first machine learning model introduces the capacity to quantify uncertainty in the computational prediction and proposes a new metric for peptide prioritization that reduces false positives in high-affinity peptide design. The second model improves the state-of-the-art performance in MHC-ligand prediction by employing a deep language model to learn the sequence determinants for auxiliary processes in MHC-ligand selection, such as proteasome cleavage, that are omitted by existing methods due to the lack of labeled data. Third, we develop machine learning frameworks to model the enrichment of an antibody sequence in phage-panning experiments against a target antigen. We show that antibodies with low specificity can be reduced by a computational procedure using machine learning models trained for multiple targets. Moreover, machine learning can help to design novel antibody sequences with improved affinity.
32-G575
Belfer
sarah_donahue@hks.harvard.edu
Fintech@CSAIL Lecture: The Future of Banking
Professor Andrew Lo, Faculty Co-Director
FinTech@CSAIL
Add to Calendar
2019-04-25 15:00:00
2019-04-25 16:30:00
America/New_York
Fintech@CSAIL Lecture: The Future of Banking
Panelists include:Sandeep AroraCOO, Citi MarketsDavid Kadio-MorokroPrincipal, Deputy Innovation Leader, US Financial Services Advisory, EYWith the onset of innovative and disruptive technologies such as distributed ledgers, applications of artificial intelligence, and automation beginning to emerge in the world of finance, what role will banks play in the future? How are banks working with regulators to create a regulatory framework for the future? All of these questions and more will be addressed in our first FinTech@CSAIL lecture series for 2019 presented by FinTech@CSAIL Faculty Advisor Andrew Lo and our panelists of CSAIL Alliances member companies representing different sectors of the banking industry. Panelists will demonstrate how their organizations are already preparing for these changes and helping to pave the way for the “Banks of the Future”.Livestreaming will be available for the lecture. If you plan to attend remotely, please select the ticket type and we will share call-in details the week of the session.
Belfer
sarah_donahue@hks.harvard.edu
Hardness escalation in computational complexity theory
Pritish Kamath
MIT
Add to Calendar
2019-04-25 16:00:00
2019-04-25 17:00:00
America/New_York
Hardness escalation in computational complexity theory
Abstract: We prove new "hardness escalation" results in computational complexity theory; a phenomenon where hardness results against seemingly weak models of computation for any problem can be "lifted" in a completely black box manner, to much stronger models of computation by studying a simple gadget composed version of the original problem.In particular, we will show how lower bounds against the "resolution proof system" can be transformed into lower bounds on the size of "monotone circuits" and the "cutting planes proof system" (all these notions will be defined in the talk). The key point here is that resolution is a significantly weak proof system for which many hardness results are known. On the other hand, there have only been a limited set of techniques to prove lower bounds against monotone circuits and cutting planes proof system. A new result that follows from this approach is that monotone circuits cannot perform Gaussian elimination efficiently; this yields the first exponential monotone circuit size lower bound for a very simple monotone function that is computable by poly-size poly-log-depth non-monotone circuits.Finally, we describe an intimate connection of these results to the complexity class TFNP, the class of all efficiently verifiable "total" search problems. In particular, we will demonstrate equivalences between computational models and communication complexity analogues of sub-classes of TFNP. The hardness escalation results that we describe (and some others in literature) can thus be viewed as "query-to-communication lifting" theorems for these sub-classes of TFNP.
32-G449
Belfer
sarah_donahue@hks.harvard.edu
April 26, 2019
Monocular 3D Pose, Shape and Appearance Modeling
Cristian Sminshisescu
Lund University and Google
Add to Calendar
2019-04-26 14:00:00
2019-04-26 15:00:00
America/New_York
Monocular 3D Pose, Shape and Appearance Modeling
Human sensing has greatly benefited from recent advances in deep learning, parametric human modeling, and the availability of large scale 2d and 3d datasets. However, existing models make strong assumptions about the scene, considering either a single person per image, full views of the person, a simple background or many cameras. In this talk I will describe approaches that combine deep multi-task neural networks and parametric human and scene models, towards an automatic monocular visual sensing system which infers the 2d and 3d pose and shape of multiple people from a single image, automatically integrates scene constraints, and extends to video by optimally solving the temporal assignment and pose smoothness problem while preserving image alignment fidelity. Besides human sensing, the framework supports several modeling capabilities including realistic human appearance transfer as well as human behavior understanding. In this context I will introduce a new, fine-grained action classification and localization task defined on non-staged videos, recorded during robot-assisted therapy sessions of children with autism, and aimed at estimating their valence-arousal, automating therapy and quantitatively measuring progress. Time permitting, I may briefly review ongoing work on deep reinforcement learning for visual recognition as well as our matrix backpropagation methodology that allows the training of deep structured models with layers implementing global operations like SVD, eigen-decomposition, projectors, or relaxations, and supporting, the end-to-end refinement of deep spectral clustering, higher-order pooling or graph-matching models.This is joint work with C. Ionescu, E. Marinoiu, D. Nilsson, M. Zanfir, A. Popa, A. Pirinen, A. Zanfir.BioCristian Sminchisescu is a Professor at Lund University and a Research Scientist leading a team at Google Research. He has obtained a doctorate in computer science and applied mathematics from INRIA and has done postdoctoral research in the AI Lab at the University of Toronto. Prior to Lund he was a faculty in Chicago and Bonn, and has held affiliations at the University of Toronto and the Romanian Academy. He has been an associate editor for IEEE PAMI (since 2010), IJCV (since 2018), and a program chair for ECCV 2018 in Munich. His work has been funded by the NSF, the Swedish and German Science Foundations, the European Commission under a Marie Curie Excellence Grant, and the ERC under a Consolidator Grant. His work (with A. Zanfir) on deep learning for graph matching received a best paper award honorable mention at IEEE CVPR 2018.
Belfer
sarah_donahue@hks.harvard.edu
Visualization for People + Systems
University of Washington
Add to Calendar
2019-04-26 14:00:00
2019-04-26 15:00:00
America/New_York
Visualization for People + Systems
Abstract:Making sense of large and complex data requires methods that integrate human judgment and domain expertise with modern data processing systems. To meet this challenge, my work combines methods from visualization, data management, human-computer interaction, and programming languages to enable more effective and more scalable methods for interactive data analysis and communication.More specifically, my research investigates automatic reasoning over domain-specific representations of visualization and analysis workflows, in order to produce both improved human-centered designs and system performance optimizations. My work on Vega-Lite provides a high-level declarative language for rapidly creating interactive visualizations. Vega-Lite can serve as a convenient representation for tools that generate visualizations. To create effective designs, these tools must also consider perceptual principles of design. My work on Draco provides a formal model of visual encodings, a knowledge base to reason about visualization design decisions, and methods to learn design rules from experiments. Draco can formally reason over the visualization design space to recommend appropriate designs but its applications go far beyond. Draco makes theoretical design knowledge a shared resource that can be extended, tested, and systematically discussed in the research community. The Falcon and Pangloss systems enable scalable interaction and exploration of large data volumes by making principled trade-offs among people’s latency tolerance, precomputation, and approximation of computations.A recurring strategy across these projects is to leverage an understanding of people’s tasks and capabilities to inform system design and optimization.Bio:Dominik Moritz is a Computer Science PhD candidate at the University of Washington. He works with Jeffrey Heer and Bill Howe in the Interactive Data Lab and the Database Group. Dominik’s research develops scalable interactive systems for visualization and analysis. His systems have won awards at premier academic venues and are available as open source projects with significant adoption by the Python and JavaScript data science communities.
32-G882
Belfer
sarah_donahue@hks.harvard.edu
Communicating gradients for learning using activity dynamics
Canadian Institute for Advanced Research (CIFAR)
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
May 01, 2019
Data structures to represent sets of k-mers
Paul Medvedev
Penn State
Add to Calendar
2019-05-01 11:30:00
2019-05-01 13:00:00
America/New_York
Data structures to represent sets of k-mers
The analysis of biological sequencing data has been one of the biggest applications of string algorithms. The approaches used in many such applications are based on the analysis of k-mers, which are short fixed-length strings present in a dataset. While these approaches are rather diverse, storing and querying k-mer sets has emerged as a shared underlying component and there have been many specialized data structures for their representation. In this talk, I will describe the applications of k-mer sets in bioinformatics and motivate the need for specialized data structures. I will give an overview of known approaches and lower bounds, with a focus on unitig-based representations. Finally, I will describe a data structure for representing sets of k-mer sets, called the HowDe Sequence Bloom Tree.Bio:Paul Medvedev is an Associate Professor in the Department of Computer Science and Engineering and the Department of Biochemistry and Molecular Biology and the Director of the Center for Computational Biology and Bioinformatics at the Pennsylvania State University. His research focus is on developing computer science techniques for analysis of biological data and on answering fundamental biological questions using such methods. Prior to joining Penn State in 2012, he was a postdoc at the University of California, San Diego and a visiting scholar at the Oregon Health & Sciences University and the University of Bielefeld. He received his Ph.D. from the University of Toronto in 2010, his M.Sc. from the University of Southern Denmark in 2004, and his B.S. from the University of California, Los Angeles in 2002.
32-G575
Belfer
sarah_donahue@hks.harvard.edu
Deep Internal Learning
The Weizmann Institute of Science
Add to Calendar
2019-05-01 12:00:00
2019-05-01 13:00:00
America/New_York
Deep Internal Learning
In this talk I will show how complex visual inference tasks can be performed using Deep-Learning, in a totally unsupervised way, by exploiting the internal redundancy inside a single image. The strong recurrence of information inside a single natural image provides powerful internal examples which suffice for self-supervision of CNNs, without any prior examples or training data. This new “Deep Internal Learning” paradigm gives rise to true “Zero-Shot Learning”. I will show the power of this approach to a variety of problems, including super-resolution, image-segmentation, transparent layer separation, blind image-dehazing, image-retargeting, and more.
32-D463 Star Conference Room
Belfer
sarah_donahue@hks.harvard.edu
Batch Normalization Causes Gradient Explosion in Deep Randomly Initialized Networks
Greg Yang
Microsoft Research
Add to Calendar
2019-05-01 16:00:00
2019-05-01 17:00:00
America/New_York
Batch Normalization Causes Gradient Explosion in Deep Randomly Initialized Networks
Abstract: Batch Normalization (batchnorm) has become a staple in deep learning since its introduction in 2015. The authors conjectured that “Batch Normalization may lead the layer Jacobians to have singular values close to 1” and recent works suggest it benefits optimization by smoothing the optimization landscape during training. We disprove the “Jacobian singular value” conjecture for randomly initialized networks, showing batchnorm causes gradient explosion that is exponential in depth. This implies that at initialization, batchnorm in fact “roughens” the optimization landscape. This explosion empirically prevents one from training relu networks with more than 50 layers without skip connection. We discuss several ways of mitigating this explosion and their relevance in practice.
32-G575
Belfer
sarah_donahue@hks.harvard.edu
May 03, 2019
The Era of Artificial Intelligence
Kai-Fu Lee
Chairman and CEO, Sinovation Ventures
Add to Calendar
2019-05-03 11:00:00
2019-05-03 12:00:00
America/New_York
The Era of Artificial Intelligence
ABSTRACT:In this talk, I will talk about the four waves of Artificial Intelligence (AI) , and how AI will permeate every part of our lives in the next decade. I will also talk about how this will be different from previous technology revolutions -- it will be faster and be driven by not one superpower, but two (US and China). AI will add $16 trillion to our global GDP, but also cause many challenges that will be hard to solve. I will talk in particular about AI replacing routine jobs -- the consequences, the proposed solutions that don't work (such as UBI), and end with a blueprint of co-existence between humans and AI.BIO:Dr. Kai-Fu Lee is the Chairman and CEO of Sinovation Ventures (www.sinovationventures.com/) and President of Sinovation Venture’s Artificial Intelligence Institute. Sinovation Ventures, managing US$2 billion dual currency investment funds, is a leading venture capital firm focusing on developing the next generation of Chinese high-tech companies. Prior to founding Sinovation in 2009, Dr. Lee was the President of Google China. Previously, he held executive positions at Microsoft, SGI, and Apple. Dr. Lee received his Bachelor degree from Computer Science from Columbia University, Ph.D. from Carnegie Mellon University, as well as Honorary Doctorate Degrees from both Carnegie Mellon and the City University of Hong Kong. He is a Fellow of the Institute of Electrical and Electronics Engineers (IEEE), Times 100 in 2013, WIRED 25 Icons , Asian Business Leader 2018 by Asia House, and followed by over 50 million audience on social media.In the field of artificial intelligence, Dr. Lee built one of the first game playing programs to defeat a world champion (1988, Othello), as well as the world’s first large-vocabulary, speaker-independent continuous speech recognition system. Dr. Lee founded Microsoft Research China, which was named as the hottest research lab by MIT Technology Review. Later renamed Microsoft Research Asia, this institute trained the great majority of AI leaders in China, including CTOs or AI heads at Baidu, Tencent, Alibaba, Lenovo, Huawei, and Haier. While with Apple, Dr. Lee led AI projects in speech and natural language, which have been featured on Good Morning America on ABC Television and the front page of Wall Street Journal. He has authored 10 U.S. patents, and more than 100 journal and conference papers. Altogether, Dr. Lee has been in artificial intelligence research, development, and investment for more than 30 years. His New York Time and Wall Street Journal bestselling book AI Superpowers (aisuperpowers.com) discusses US-China co-leadership in the age of AI as well as the greater societal impacts brought upon by the AI technology revolution.
Kirsch 32-123
Belfer
sarah_donahue@hks.harvard.edu
May 08, 2019
Explainable Artificial Intelligence in Precision Medicine
University of Washington
Add to Calendar
2019-05-08 11:30:00
2019-05-08 13:00:00
America/New_York
Explainable Artificial Intelligence in Precision Medicine
Modern machine learning (ML) models can accurately predict patient progress and outcomes. However, they do not explain why selected features make sense or why a particular prediction was made. For example, a model may predict that a patient will get chronic kidney disease, which can lead to kidney failure. The lack of explanations about which features drove the prediction – e.g., high systolic blood pressure, high BMI, or others – hinders medical professionals in making diagnoses and decisions on appropriate clinical actions. I will briefly describe my group’s efforts to develop interpretable ML techniques for varied medical applications, including treating cancer based on a patient’s own molecular profile, identifying therapeutic targets for Alzheimer’s, predicting kidney diseases, preventing complications during surgery, enabling pre-hospital diagnoses for trauma patients, and improving our understanding of pan-cancer biology and genome biology. My talk will focus in greater detail on: MERGE, which uses ML to target treatment of acute myeloid leukemia, published in Nature Communications (Jan 2018); our explainable artificial intelligence system, Prescience, for preventing hypoxemia in patients under anesthesia, recently featured on the cover of Nature Biomedical Engineering (Oct 2018); and SHAP, our general ML framework on model interpretability, published as a full oral presentation at Neural Information Processing Systems (Dec 2017; cited 150).Short bioProf. Su-In Lee is an Associate Professor in the Paul G. Allen School of Computer Science & Engineering and an Adjunct Associate Professor in the Departments of Genome Sciences, Electrical Engineering, and Biomedical Informatics and Medical Education at the University of Washington. She completed her PhD in 2009 at Stanford University with Prof. Daphne Koller (Stanford Artificial Intelligence Laboratory). Before joining the UW in 2010, Lee was a visiting professor in the Computational Biology Department at Carnegie Mellon University. She has received the National Science Foundation CAREER Award and been named an American Cancer Society Research Scholar. She has received numerous generous grants from the National Institutes of Health, the National Science Foundation, and the American Cancer Society. Lee is currently the PI for the following active grants: NIH/NIA R01, NIH/NLM R21, NIH/NIGMS R35, NSF/BIO INNOVATION, NSF/BIO CAREER, and ACS Research Scholar.
32-G575
Belfer
sarah_donahue@hks.harvard.edu
May 09, 2019
Recent advances in fetal imaging
Joseph Hajnal
King's College London
Add to Calendar
2019-05-09 16:00:00
2019-05-09 17:00:00
America/New_York
Recent advances in fetal imaging
The talk will present recent work on fetal imaging at KCL where wehave been working to develop comprehensive techniques working acrossMRI modalities and seeking to address many challenges associated withmotion, field variations and efficiency by integrating MR physics,data correction strategies and image analysis.
32-D507
Belfer
sarah_donahue@hks.harvard.edu
May 15, 2019
TBA
Tsachy Weissman
Stanford Univerrsity
Add to Calendar
2019-05-15 11:30:00
2019-05-15 13:00:00
America/New_York
TBA
32-G575
Belfer
sarah_donahue@hks.harvard.edu
May 20, 2019
Accelerator-level Parallelism: Mobile SoCs as Harbinger of the Future
Mark D. Hill
University of Wisconsin-Madison
Add to Calendar
2019-05-20 14:00:00
2019-05-20 15:00:00
America/New_York
Accelerator-level Parallelism: Mobile SoCs as Harbinger of the Future
Abstract:This talk will first discuss how computer systems are transitioning from homogeneous parallelism to heterogeneity: ILP, TLP, and new Accelerator-level Parallelism (ALP). It will then discuss systems on a chip (SoCs) for mobile computing, and why they may be a harbinger of computer systems’s ALP future. It will conclude presenting the Gables model largely developed at Google’s Mobile Silicon Group for HPCA 2019’s Industrial Session. Gables seeks to make SoC selection and design more scientific via extending Roofline and bottleneck analysis to provide the first answers, not the final answers. This is joint work with Vijay Janapa Reddi at Harvard.Biography:Mark D. Hill (http://www.cs.wisc.edu/~markhill) is John P. Morgridge Professor and Gene M. Amdahl Professor of Computer Sciences at the University of Wisconsin-Madison, where he also has a courtesy appointment in Electrical and Computer Engineering. His research interests include parallel-computer system design, memory system design, and computer simulation. He is a fellow of IEEE and the ACM. He serves as Chair of the Computer Community Consortium (2018-19) and served as Wisconsin Computer Sciences Department Chair 2014-2017. Hill has a PhD in computer science from the University of California, Berkeley.**refreshments at 1:45 pm
32-G882
Belfer
sarah_donahue@hks.harvard.edu
June 07, 2019
Add to Calendar
2019-06-07 13:30:00
2019-06-07 15:00:00
America/New_York
CSAIL Commencement Reception
Congratulations CSAIL graduates! Please join us for some cake and light refreshments following the commencement exercises on Friday, June 7, 2019 in the R&D Commons.
R&D Commons
Belfer
sarah_donahue@hks.harvard.edu