Skip to:

Past Events

View events by calendar year
Enter one or more keywords to search for
Filter by event type
Monday, February 29, 2016 -
4:30pm to 5:30pm

Speaker: Austin Benson

Thursday, February 25, 2016 -
4:30pm to 5:45pm

Speaker: Michael Friedlander, UC Davis

Title: Level-set methods for convex optimization
Abstract: Convex optimization problems in a variety of applications have favorable objectives but complicating constraints, and first-order methods are not immediately applicable.  We propose an approach that exchanges the roles of the objective and constraint functions, and instead approximately solves a sequence of parametric problems.  We describe the theoretical and practical properties of this approach for a broad range of problems, including low-rank semidefinite optimization.
Joint work with A. Aravkin, J. Burke, D. Drusvyatskiy, S. Roy.
Tuesday, February 23, 2016 -
12:30pm to 1:30pm

Speaker: Gill Bejerano, Stanford University

About the Speaker: Gill Bejerano is an Associate Professor of Developmental Biology, of Computer Science and of Pediatrics (Genetics).

The Bejerano Lab studies genome function in human and related species. In the lab, Gill and his colleagues are deeply interested in the following broad questions: Mapping genome sequence (variation) to phenotype (differences) and extracting specific genetic insights from deep sequencing measurements. The lab takes a particular interest in gene cis regulation. They use their joint affiliation to apply a combination of computational and experimental approaches. They collect large scale experimental data; write computational analysis tools; run them massively to discover the most exciting testable hypotheses; which they proceed to experimentally validate. Gill and his colleagues work in small teams, in house or with close collaborators of experimentalists and computational tool users who interact directly with our computational tool builders. 

Monday, February 22, 2016 -
4:30pm to 5:30pm

Speaker: Fayadhoi Ibrahima

About the Speaker: Fayadhoi Ibrahima is a PhD candidate in the Institute for Computational and Mathematical Engineering (ICME) working under the supervision of Professor Hamdi Tchelepi in the Energy Resources Engineering (ERE) department on uncertainty quantification for oil recovery in reservoir engineering. Prior to starting his PhD research, Fayadhoi has been putting efforts in building a coherent understanding of numerical analysis and probability at a strong level acquired in 3 universities: the University Pierre and Marie Curie (UPMC), the Ecole Centrale Paris (ECP) [both in France] and Stanford University. Fayadhoi was born and raised in France, but his parents are from the Comoros Islands. Fayadhoi enjoys traveling, teaching and running.

Friday, February 19, 2016 -
4:00pm to 5:00pm

Speaker: Fred Luskin 

About the Speaker: Fred Luskin, Ph.D., is the Director of the Stanford University Forgiveness Projects, a Senior Consultant in Health Promotion at Stanford University, and a Professor at the Institute for Transpersonal Psychology, as well as an affiliate faculty member of the Greater Good Science Center. He is the author of Forgive for Good: A Proven Prescription for Health and Happiness (HarperSanFrancisco, 2001) and Stress Free for Good: Ten Proven Life Skills for Health and Happiness (HarperSanFrancisco, 2005), with Kenneth Pelletier, Ph.D.

Thursday, February 18, 2016 -
4:30pm to 5:45pm

Speaker: Chao Yang, Lawrence Berkeley National Lab

Title: Fast Numerical Methods for Electronic Structure Calculations
Abstract: The Kohn-Sham density functional theory (KSDFT) is the most widely used theory for studying electronic properties of molecules and solids. It reduces the task of solving a many-body Schrodinger's equation to that of solving a system of single electron equations coupled by the electron density.  These equations can be viewed as a nonlinear eigenvalue problem. Although they contain far fewer degrees of freedom, they are more difficult to solve because of the nonlinear nature of the problem. I will give an overview of efficient algorithms for solving this type of problem, and present recently developed numerical algebraic techniques for reducing the computational complexity of the KSDFT calculation. These techniques exploit special properties of the Kohn Sham map, which is a nonlinear fixed point map of the ground state electron density to itself.
Tuesday, February 16, 2016 -
12:30pm to 1:30pm

Speaker: Doug James

About the Speaker: Doug L. James is a Full Professor of Computer Science at Stanford University since June 2015, and was previously an Associate Professor of Computer Science at Cornell University from 2006-2015. He holds three degrees in applied mathematics, including a Ph.D. in 2001 from the University of British Columbia. In 2002 he joined the School of Computer Science at Carnegie Mellon University as an Assistant Professor, before joining Cornell in 2006. His research interests include computer graphics, computer sound, physically based animation, and reduced-order physics models. Doug is a recipient of a National Science Foundation CAREER award, and a fellow of both the Alfred P. Sloan Foundation and the Guggenheim Foundation. He recently received a Technical Achievement Award from The Academy of Motion Picture Arts and Sciences for "Wavelet Turbulence," and the Katayanagi Emerging Leadership Prize from Carnegie Mellon University and Tokyo University of Technology. He was the Technical Papers Program Chair of ACM SIGGRAPH 2015, and is currently a consulting Senior Research Scientist at Pixar Animation Studios.

Friday, February 12, 2016 -
4:00pm to 5:00pm

Join ICME faculty for a discussion on “All About Advisors”:

  • Finding one
  • Managing one
  • Managing more than one


Thursday, February 11, 2016 -
4:30pm to 5:45pm

Speaker: Eugene Vecharynski, Lawrence Berkeley National Laboratory

Eugene Vecharynski is a postdoctoral research fellow with the Scalable Solvers Group in the Computational Research Division at the Lawrence Berkeley National Laboratory. He is currently working on numerical solution of extreme-scale eigenvalue problems arising in electronic structure calculations. 
Eugene's research is primarily in the field of preconditioned iterative solvers for large algebraic systems, including linear and nonlinear systems and eigenvalue problems. He has contributed to various aspects of matrix computations, such as development of new solvers, theoretical analysis of Krylov subspace methods, and novel preconditioning techniques. Results of his work have been used in a number of applications in physical sciences and data analytics.
Before joining LBNL, Eugene was a postdoc with Yousef Saad and Masha Sosonkina at the Department of Computer Science and Engineering, University of Minnesota, and a Visiting Assistant Professor at the Department of Mathematics and Statistics, Georgia State University.I have a Ph.D. from the Department of Mathematical and Statistical Sciences, University of Colorado Denver, advised by Andrew Knyazev.
Tuesday, February 9, 2016 -
12:30pm to 1:30pm

Speaker: Virgina Williams

About the Speaker: Virginia William is an Assistant Professor of Computer Science at Stanford University. Her research applies combinatorial and graph theoretic tools to various computational domains. Her recent work has focused on the following domains:

(i) designing algorithms for shortest paths, pattern detection and other computational problems in graphs and matrices
(ii) reducing fundamental computational problems to one another in a fine-grained way, sometimes showing equivalences
(iii) studying how much graph distance information can be compressed, and computational issues in social choice: when and how can one efficiently manipulate elections, tournaments and competitions, how to measure the quality of a voting rule, etc.

Thursday, February 4, 2016 -
4:30pm to 5:45pm

Speaker: Paul T. Boggs, Sandia National Laboratories (Retired)

Title: Adaptive, Limited-Memory BFGS Algorithms for Unconstrained Optimization

Abstract: The limited-memory BFGS method (L-BFGS) has become the workhorse optimization strategy for many large-scale nonlinear optimization problems. A major difficulty with L-BFGS is that, although the memory size M can have a great effect on performance, it is difficult to know which size will work best. Importantly, a larger M does not necessarily improve performance, but may in fact degrade it. There is no guidance in the literature on how to choose M. In this talk, we briefly review L-BFGS and then suggest two computationally efficient ways to measure the effectiveness of various memory sizes, thus allowing us to adaptively choose a different M at each iteration. The numerical success of these two adaptive strategies suggested ways to extend them, which we briefly consider. Our numerical results illustrate that our approach improves the performance of the L-BFGS method, and indicate some further directions for research

Tuesday, February 2, 2016 - 2:00pm

Oral Examination
Student: Saman Ghili
Advisor: Prof. Iaccarino

Title: Polynomial and Rational Approximation Techniques for Non-intrusive Uncertainty Quantification

Abstract: With the growth of computing power in recent decades, uncertainty quantification (UQ) for numerical simulations of engineering systems has attracted significant attention. Most non-intrusive UQ methods are concerned with running simulations with several values of the input uncertain/design parameters, and using the output to construct an accurate and efficient surrogate that describes the behavior of the quantity of interest as a function of the sources of uncertainty. The fundamental underlying problem is that of approximating a function from its values at a discrete set of points in its domain.

In the first portion of the talk, we will discuss a non-intrusive polynomial chaos expansion (PCE) technique, in which we use weighted least squares to construct a multivariate polynomial surrogate. The quality of the approximation depends crucially on the location of the points at which the function is evaluated. We present a novel optimization based method for finding the best points for this type of approximation. In the second portion, we discuss the problem where we are not free in choosing the points at which the function will be evaluated. For this scenario, we introduce an efficient rational interpolation scheme based on the Floater-Hormann rational interpolation. We present theoretical guarantees regarding the accuracy and stability of this scheme, and verify its efficiency when compared to similar methods in the literature.

Location: 520-131
Time: 2 p.m.

Tuesday, February 2, 2016 -
12:30pm to 1:30pm

Speaker: Jack Poulson

About the Speaker: Jack Poulson is an Assistant Professor of Mathematics and Member of the Institute of Computational and Mathematical Engineering at Stanford University and completed his Ph.D. in Computational and Applied Mathematics at The University of Texas at Austin at the end of 2012. His current research is focused on developing efficient distributed-memory algorithms for conic Interior Point Methods (especially for Second-Order Cone Programs) and lattice reduction techniques (such as BKZ 2.0). Said research is publicly performed within the library Elemental (

Monday, February 1, 2016 -
4:30pm to 5:30pm

Speaker: Anil Damle

About the Speaker: Anile Damle is a PhD candidate in the Institute for Computational and Mathematical Engineering (ICME) at Stanford University. His general interests include numerical linear algebra, non-linear approximations, matrix analysis, and fast algorithms for structured matrices. Anil's current research projects focus on localization of Kohn-Sham orbitals, updating of certain tree based matrix factorizations, and non-negative matrix factorizations. Visit Anil's personal webpage at:

Friday, January 29, 2016 -
4:00pm to 5:00pm

Speaker: Margot Gerritsen, ICME Director

Dealing with Stress in Grad School
Part 1 - The Happy Imposter
Many of your peers, as well as researchers and faculty, will at one point or another deal with the so-called Imposter Syndrome. The Imposter Syndrome is the feeling that you do not belong, or should not really be where you are, as you are not as good as people think; and that you will one way or another, disappoint. A few years ago, we conducted a study of the Imposter Syndrome on campus. The results may surprise, and comfort, you. You are not alone. 
Margot will present this study and discuss possible ways to deal with the Imposter Syndrome, and turn it from a debilitating state to something more productive. The rest of the hour is for discussion, sharing of stories and best practices.
Tuesday, January 26, 2016 -
12:30pm to 1:30pm

Speaker: Jure Leskovec

About the Speaker: Jure Leskovec is an Assistant Professor of Computer Science at Stanford University, where he is a member of the InfoLab and Artificial Intelligence Lab. Jure also works as a Chief Scientist at Pinterest, where he focuses on machine learning problems. He co-founded a machine learning startup called Kosei, which was acquired by Pinterest.

In 2008/09, Jure was a postdoctoral researcher at Cornell University, working with Jon Kleinberg and Dan Huttenlocher. He completed his Ph.D. in Machine Learning at the School of Computer Science at Carnegie Mellon University under the supervision of Christos Faloutsos in 2008. His undergraduate career in Computer Science was completed at the University of Ljubljana, Slovenia in 2004.
Monday, January 25, 2016 -
4:30pm to 5:30pm

Speaker: Victor Minden

About the Speaker: Victor Minden is a PhD student at the Institute for Computational and Mathematical Engineering (ICME) at Stanford University, where he works with Lexing Ying. His research concerns fast algorithms for scientific computing, in particular fast linear algebra on rank-structured matrices.

More broadly, Victor's research interests include: 
- numerical linear algebra
high-performance computing
signal processing
network theory
numerical optimization
Thursday, January 21, 2016 -
4:30pm to 5:45pm

Speaker: Ange Toulougoussou, The French Institute for Research in Computer Science and Automation and Pierre and Marie Curie University

Title: All-at-once dual-primal Schur domain decomposition method for saddle-point problems: Application to Stokes

Abstract: We consider the solution of large sparse linear systems in saddle-point form. Iterative methods are preferred to direct methods for memory reasons. The rate of convergence of iterative methods depends strongly on the spectrum of the system matrix. Indefiniteness is a significant challenge. Our goal is to introduce an iterative method that circumvents indefiniteness and takes advantage of parallel high performance computing. We present two efficient domain decomposition methods for symmetric positive definite matrices, and we discuss the extension of FETI and BDD to symmetric positive semidefinite problems. These extensions are key ingredients for applying the method to singular saddle-point problems.
FETI and BDD are very similar methods but they differ in the definition of interfaces. We show that a single definition of interfaces can be used in both methods. We then combine FETI and BDD to introduce a dual-primal Schur domain decomposition method for saddle-point problems. We apply the new method to the linear system arising from the discretization of Stokes by P1bubbleP1 mixed finite elements (also called Mini elements). We present numerical results to show the efficiency of the method.
Tuesday, January 19, 2016 -
12:30pm to 1:30pm

Speaker: Percy Liang

About the Speaker: Percy Liang is an Assistant Professor of Computer Science and, by courtesy, of Statistics at Stanford University. He is affiliated with Stanford Artificial Intelligence Lab and the Stanford Natural Language Processing Group. His two aims in research include: (i) creating a software that allow humans to communicate with computers and (ii) developing algorithms that can infer latent structures from raw data. Percy broadly identifies with the machine learning (ICML, NIPS) and natural language processing (ACL, NAACL, EMNLP) communities. Percy is also a strong proponent of efficient and reproducible research.  He develops CodaLab Worksheets in collaboration with Microsoft Research, a new platform that allows researchers to maintain the full provenance of an experiment from raw data to final results. 

Thursday, January 14, 2016 -
4:30pm to 5:45pm

Speaker: Dr. Laura Grigori

About the Speaker: Dr. Laura Grigori received a Ph.D. in Computer Science (December 2001) from Université Henri Poincaré, France, INRIA Lorraine. After spending two years at UC Berkeley and LBNL as a postdoctoral researcher, she joined INRIA in January 2004. Laura was a member of Sage group at INRIA Rennes and Grand-Large group at INRIA Saclay - Ile de France and LRI, Paris 11 University. Since January 2013, she has been leading Alpines, a joint group between INRIA Paris - Rocquencourt and J.L. Lions Laboratory, UPMC. Laura's research interests include numerical linear algebra, high performance computing for scientific applications, sparse matrix computations, combinatorial scientific computing, and mathematical software.

Tuesday, January 12, 2016 -
12:30pm to 1:30pm

Speaker: Nick Henderson

Join Nick Henderson, Research Associate at ICME, in this informal discussion of research, teaching, and fun times in ICME!

About the Speaker: Nick Henderson is a Research Associate and Instructor at ICME. He is also affiliated with the CUDA Center of Excellence, where he collaborates with Stanford Linear Accelerator Center (SLAC) and High Energey Accelerator Research Organization (KEK) in developing GPU-based algorithms and codes to accelerate the simulation of particles travelling through and interacting with material. More of Nick's current work can be found on his webpage:

Monday, January 11, 2016 -
4:30pm to 5:30pm

Speaker: Sergio Camelo

About the Speaker:  Sergio is a PhD student in Computational Mathematics at Stanford (ICME). His interests cover the broad fields of convex optimization and statistics, particularly machine learning and approximation algorithms. Sergio holds a Master's degree in Mathematics from Universidad de los Andes in Colombia, where he worked under the advice of Professor Mauricio Velasco on copositive optimization and semidefinite relaxations to find the independence number of graphs.Sergio also holds a Bachelor’s degree in Mathematics and Economics. His thesis focused on optimal bandwidths for kernel classification under the advice of Prof. Adolfo Quiroz.

For two years, Sergio was a Junior Researcher at Quantil Applied Mathematics working on machine learning. Some of his projects included developing models to predict patterns of crime in Colombian cities through the estimation of non-homogeneous Poisson processes; developing software that detects fraud in the healthcare system, measuring the cross-entropy of reports that insurers give to the government; and evaluating the design of auctions in the centralized dispatch of energy through estimation of the marginal costs of energy producers and simulation of the auction as a big linear program.
Thursday, January 7, 2016 -
4:30pm to 5:45pm

Speaker: Miles Lubin, PhD candidate in Operations Research, MIT

About the Speaker: Miles is a fourth-year Ph.D. candidate in Operations Research at MIT. He received his B.S. in Applied Mathematics and M.S. in Statistics from the University of Chicago in 2011. After graduating, he spent a year as a researcher at Argonne National Laboratory before starting at MIT. His research interests span diverse areas of mathematical optimization, with a unifying theme of developing new methodologies for large-scale optimization drawing from motivating applications in renewable energy. Miles has published work in chance constrained optimization, mixed-integer conic optimization, robust optimization, stochastic programming, algebraic modeling, automatic differentiation, numerical linear algebra, and parallel computing techniques for large-scale problems. 

Tuesday, January 5, 2016 -
12:30pm to 1:30pm

Speaker: TBD

Monday, January 4, 2016 -
4:30pm to 5:30pm

Speaker: TBD

Thursday, December 3, 2015 -
4:30pm to 5:45pm

Speaker: Ming Gu, Professor in the Department of Mathematics at the University of California, Berkeley

Title: Spectrum-revealing Matrix Factorizations

Abstract: Low-rank matrix approximations have become of central importance in the era of big data. Efficient and effective methods for such approximations have been proposed in statistics, theoretical computer science, and optimization. In this talk, we establish spectrum-revealing matrix factorizations, a new framework for efficient and effective matrix approximations. These factorizations are variations of the more classical LU, QR, and Cholesky factorizations with row (and/or) column permutations but are competitive with the best matrix approximations in both theory and computational effectiveness. We also discuss extensions of these factorizations for efficient computations of the truncated SVD and solutions of nuclear norm minimization problems. We demonstrate the effectiveness of our approaches with numerical experiments with both synthetic and real data.

Tuesday, December 1, 2015 -
12:30pm to 1:30pm

Speaker: Sanjeeb Bose, Consulting Assistant Professor, ICME at Stanford University

About the Speaker: Sanjeeb Bose holds a BS degree in Mechanical Engineering from the California Institute of Technology and a PhD in Mechanical Engineering from Stanford University. He was a recipient of the Department of Energy’s Computational Science Graduate Fellowship as a graduate student. He specializes in subgrid-scale and wall modeling for LES and algorithms for large-scale parallel computing. Sanjeeb also contributes to the development of the core infrastructure and numerical methods in the Cascade flow solvers at Cascade Technologies.

Monday, November 30, 2015 -
4:30pm to 5:30pm

Speaker: Saman Ghili, ICME at Stanford University

Title: Scattered Data Interpretation via Weighted L1 Minimization

About the Speaker: Saman Ghili is a graduate student at Stanford University. He works on low-rank separated representations, a promising approach to handle computational models affected by a large number of uncertainties. He is focused on the study of the uncertainty introduced in reactive flows by imprecise specification of the reaction rates. 

Monday, November 30, 2015 -
2:30pm to 3:30pm

Speaker: Torben Andersen, Northwestern University

Title: Pricing Short-Term Market Risk: Evidence from Weekly Options

Abstract: We study short-term market risks implied by weekly S&P 500 index options. The introduction of weekly options has dramatically shifted the maturity profile of traded options over the last five years, with a substantial proportion now having expiry within one week. Economically, this reflects a desire among investors for actively managing their exposure to very short-term risks. Such short-dated options provide an easy and direct way to study market volatility and jump risks. Unlike longer-dated options, they are largely insensitive to the risk of intertemporal shifts in the economic environment, i.e., changes in the investment opportunity set. Adopting a novel general semi-nonparametric approach, we uncover variation in the shape of the negative market jump tail risk which is not spanned by market volatility. Incidents of such tail shape shifts coincide with serious mispricing of standard parametric models for longer-dated options. As such, our approach allows for easy identification of periods of heightened concerns about negative tail events on the market that are not always “signaled” by the level of market volatility and elude standard asset pricing models.

Friday, November 27, 2015 -
8:00am to 5:00pm

No seminars this week due to Thanksgiving break. 

Thursday, November 26, 2015 -
8:00am to 5:00pm

No seminars this week due to Thanksgiving break. 

Wednesday, November 25, 2015 -
8:00am to 5:00pm

No seminars this week due to Thanksgiving break. 

Tuesday, November 24, 2015 -
8:00am to 5:00pm

No seminars this week due to Thanksgiving break. 

Monday, November 23, 2015 -
8:00am to 5:00pm

No seminars this week due to Thanksgiving break. 

Thursday, November 19, 2015 -
4:30pm to 5:45pm

Speaker: Tamara Kolda, Sandia

Title: Optimization Challenges in Tensor Factorization
Abstract: Tensors are multiway arrays, and tensor decomposition is a powerful tool for compression and data interpretation.  We demonstrate the utility of tensor decomposition with several examples and explain the optimization challenges, both theoretical and practical. The optimization problems are nonconvex, but they can typically be solved via an alternating approach that yields convex subproblems. We consider open problems such as determining the model complexity, tensor completion, incorporating symmetries and other constraints, handling ambiguities in scaling and permutation, enforcing structure like sparsity, and considering alternative objective functions.
Thursday, November 19, 2015 -
4:30pm to 5:30pm

Speaker: Andrew Lo, MIT

About the Speaker: Andrew W. Lo is the Charles E. and Susan T. Harris Professor at the MIT Sloan School of Management and director of the MIT Laboratory for Financial Engineering. He received his Ph.D. in economics from Harvard University in 1984. Before joining MIT’s finance faculty in 1988, he taught at the University of Pennsylvania’s Wharton School as the W.P. Carey Assistant Professor of Finance from 1984 to 1987, and as the W.P. Carey Associate Professor of Finance from 1987 to 1988.

He has published numerous articles in finance and economics journals, and has authored several books including The Econometrics of Financial Markets, A Non-Random Walk Down Wall Street, Hedge Funds: An Analytic Perspective, and The Evolution of Technical Analysis. He is currently co-editor of the Annual Review of Financial Economics and an associate editor of the Financial Analysts Journal, the Journal of Portfolio Management, and the Journal of Computational Finance.
His awards include the Alfred P. Sloan Foundation Fellowship, the Paul A. Samuelson Award, the American Association for Individual Investors Award, the Graham and Dodd Award, the 2001 IAFE-SunGard Financial Engineer of the Year award, a Guggenheim Fellowship, the CFA Institute’s James R. Vertin Award, the 2010 Harry M. Markowitz Award, and awards for teaching excellence from both Wharton and MIT.
Tuesday, November 17, 2015 -
12:30pm to 1:15pm

Speaker: Rachael Briggs, Professor of Philosophy

Title: Conditionals 

Abstract: Conditionals (roughly, "if... then..." statements) are useful in ordinary inferences about the world, in decision making and decision theory, in reasoning on the basis of theories, and in philosophy.  I canvass a variety of accounts of conditionals, including selection semantics, a material conditional account, an account on which conditionals lack truth values... and argue that all of them can be understood as instances of the same general template, with different ways of filling in the details.

Monday, November 16, 2015 -
4:30pm to 5:30pm

Speaker: Ohiremen Dibua

Title: Sensitivity Analysis on Model of Electrostatic Actuator In Conducting Liquid
Abstract: Comb-Drive Actuators are well modeled in air. We want to be able to use them for biological applications, such as measuring the mechanical properties of cells. In order to do this, we must operate the actuators in conducting liquid. Our goal is, therefore, to develop increasingly accurate models of the behavior of the comb-drive actuator in conducting liquids. We begin this process by exploring the sensitivity of a simple model of the comb-drive actuator.
Friday, November 13, 2015 -
9:30am to 11:30am

ICME will host four tech workshops on Friday, November 13th as a part of the ICME Xtend annual event.  Whether you are participating in the student recruiting at ICME Xtend or not -- you are welcome to join us for the following tech talks.  

As room is limited, please RSVP by this Wednesday, November 11th to ensure that we have space available:  

9:30- 10:30 a.m.

Choose between the following (RSVP required):

  • Interactive Data Visualization: The Face of Big Data.  HIVE Visualization Center, Huang B050
  • Law, Order, and Algorithms. Huang Building, Room 305


10:30- 11:30 a.m.

Choose between the following (RSVP required)

  • The Center for Financial and Risk Analytics (CFRA).  HIVE Visualization Center, Huang B050
  • Experiments on Networks.  Huang Building, Room 305
Thursday, November 12, 2015 -
4:30pm to 5:45pm

Speaker: Kevin Carlberg, Sandia National Laboratories

Title: Nonlinear Model Reduction: Discrete Optimality and Time Parallelism

Abstract: Large-scale models of nonlinear dynamical systems arise in diverse applications ranging from compressible fluid dynamics to the dynamics of the power grid.  Due to the large computational cost incurred by these models, it is impractical to use them in time-critical scenarios (e.g., control, design, uncertainty quantification), which demand fast-turnaround simulations for rapid decision making.  For simple systems (e.g., linear-time-invariant systems), researchers have developed effective model-reduction techniques that decrease the dimensionality of the dynamical system by projecting the governing equations onto a low-dimensional subspace computed a priori. Such reduced-order models (ROMs) are often equipped with stability guarantees and computable error bounds.  In contrast, model reduction for highly nonlinear dynamical systems remains in its infancy. In fact, the most common approach (POD-Galerkin) is known to be unstable for many practical problems.
This talk will describe several advances that have made nonlinear model reduction viable for a new frontier of problems.  These advances are data driven in nature: they leverage simulation data to drastically reduce the computational resources needed to accurately resolve such models. First, I will introduce the notion of discrete-optimal projection.  In contrast to standard Galerkin-projection strategies, this approach performs optimal projection at the time-discrete level, i.e., after the dynamical system has been discretized in time. Through a set of comparative theoretical and computational studies, I will highlight the practical benefits of discrete-optimal projection over Galerkin projection.
Next, I will discuss one specific discrete-optimal model-reduction technique: the Gauss-Newton with approximated tensors (GNAT) method. This approach realizes computational savings not only through dimensionality reduction via projection, but also by approximating the discrete residual vector such that only a few of its entries must be computed.  This implies that only a small subset of the computational mesh (the 'sample mesh') is required for the GNAT simulation, which enables it to run on a small number of computing cores.
Finally, I will introduce a new approach for data-driven time parallelism.  Because the GNAT model incurs a small computational footprint, parallelizing the computation in the spatial domain quickly saturates; this limits the realizable wall-time speedup for the GNAT ROM.  To address this, we introduce a new method for parallelizing the simulation in time.  The technique relies on a coarse propagator that ensures rapid convergence by leveraging previously available time-domain data.
Thursday, November 12, 2015 - 9:00am to Friday, November 13, 2015 - 1:00pm


ICME Xtend returns for a second year in 2015.

Our main event will take place on Thursday and Friday, November 12-13, 2015.  This one-of-a-kind event brings ICME students and faculty together with our partners from industry and national labs for two days of networking and recruiting, discussions on current trends in our fields, workshops, and mixers.  

For details on special pre-Xtend events, please see the 'more information' section below.  


More Information

Information for ICME students: Click here-

Information for ICME partners: Click here-

Information on joining the ICME partnership programs:  Click here-



Xtend Day 1, November 12, 2015

9:00-11:00 a.m.

Networking Breakfast

Huang basement, Student Commons

11:00 a.m.- 5:00 p.m.


See your schedule for specific appointments and locations

5:00 p.m.-7:30 p.m.


ICME lobby, Huang basement, Suite 060

Xtend Day 2, November 13, 2015

8:30- 9:30 a.m.


ICME Lobby, Huang Basement, Suite 060

9:30- 10:30 a.m.

Choose between the following (RSVP required)

Interactive Data Visualization: The Face of Big Data

HIVE Visualization Center, Huang B050

Law, Order, and Algorithms

Huang Building, Room 305

10:30- 11:30 a.m.

Choose between the following (RSVP required)

The Center for Financial and Risk Analytics (CFRA)

HIVE Visualization Center, Huang B050

Experiments on Networks

Huang Building, Room 305


Tuesday, November 10, 2015 -
12:30pm to 1:15pm

Speaker: Michael Saunders, Professor of Management Science and Engineering at Stanford University

Michael Saunders is a Research Professor in the Systems Optimization Laboratory at Stanford University. He obtained his PhD in Computer Science from Stanford in 1972 (advisor Gene Golub). He is known for contributions to mathematical software. He teaches Large-Scale Numerical Optimization (MS&E318, CME338). Michael Saunders grew up in Christchurch, New Zealand, where he received a BSc (Honors) degree in Mathematics at the University of Canterbury. He was a Scientific Officer at the DSIR in Wellington, New Zealand, for the period 1966-78. He received his MS in 1970 and PhD 1972, both in Computer Science at Stanford University (advisor Gene Golub). He spent two years at the Stanford Operations Research Department in 1975-76, and he rejoined the department as a senior research associate in 1979. He was appointed Professor (Research) in 1987.

The Stanford OR Department evolved into EESOR and finally Management Science and Engineering. Professor Saunders is a member of the Systems Optimization Laboratory (SOL) within the MS&E Department. He was a faculty member in Gene Golub's SCCM Program at Stanford, and is now part of its successor, the Institute of Computational Mathematics and Engineering at Stanford (ICME).
His research interests include numerical optimization, numerical linear algebra, sparse-matrix methods, and portable software. He teaches a class on Large-Scale Numerical Optimization (MS&E 318, CME 338).
Professor Saunders was Associate Editor for ACM TOMS 1982-2004, for SIAM Journal of Optimization 1989-2001, and for Optimization and Engineering 1999-present. He is co-author of numerical software LSQR, LSMR, MINRES, MINRES-QLP, SYMMLQ, LUSOL for sparse linear equations, and MINOS, LSSOL, NPSOL, QPOPT, SQOPT, SNOPT, PDCO for constrained optimization.
Professor Saunders was first recipient of the Orchard-Hays Prize from the Mathematical Programming Society 1985. He was elected Honorary Fellow of the Royal Society of New Zealand in 2007. In 2012 he won the SIAM Linear Algebra Prize (coauthors Sou-Cheng Choi and Christopher Paige) and was inducted into the Stanford Invention Hall of Fame (coinventors Philip Gill, Walter Murray, Bruce Murtagh, and Margaret Wright) for Innovation in the Development of Optimization Software. In 2013 he became a SIAM Fellow.


Thursday, November 5, 2015 -
4:30pm to 5:30pm

Speaker: Thomas Strohmer, Department of Mathematics at UC Davis

Title: Phase Retrieval, Self-Calibration, Random Matrices, and Convex Optimization

Abstract: I will demonstrate how two important but seemingly unrelated problems, namely Phase Retrieval and Self-Calibration, can be solved by using methods from random matrix theory and convex optimization.

Phase retrieval is the century-old problem of reconstructing a function, such as a signal or image, from intensity measurements, typically from the modulus of a diffracted wave. Phase retrieval problems -- which arise in numerous areas including X-ray crystallography, astronomy, diffraction imaging, and quantum physics -- are notoriously difficult to solve numerically. They also pervade many areas of mathematics, such as numerical analysis, harmonic analysis, algebraic geometry, combinatorics, and differential geometry.

Self-calibration is an increasingly important concept, since the need for precise calibration of sensing devices manifests itself as a major roadblock in many scientific and technological endeavors. The idea of self-calibration is to equip a hardware device with a smart algorithm that can compensate automatically for the lack of calibration. I will demonstrate how both phase retrieval and self-calibration can be treated efficiently via convex programming by "lifting" the associated nonlinear inverse problem to an underdetermined linear problem. Using tools from random matrix theory and compressive sensing, we will see that for certain types of random measurements both problems can be solved exactlyvia a convenient semidefinite program. Applications in x-ray crystallography, array calibration, and the Internet-of-Things will be discussed.

Tuesday, November 3, 2015 -
12:30pm to 1:15pm

Speaker: TBD

Monday, November 2, 2015 -
4:30pm to 5:30pm

Speaker: Daniele Schiavazzi, Department of Pediatrics at Stanford University

Title: Assimilation and Propagation of Clinical Data Uncertainty In Cardiovascular Modeling

Abstract: The increasing adoption of computational tools to complement clinical data collection and inform treatment decisions demands a thorough characterization of the confidence in the predicted quantities. This confidence is affected by modeling assumptions as well as variability in clinical data, anatomy, vessel wall material and physiologic state. A wide spectrum of pathologies can be investigated using modern computational tools, including complex fluid-structure interaction phenomena (e.g., valve dynamics), multi-scale coupling between 3D Navier-Stokes solvers and 0D boundary circulation models, and growth and remodeling of vascular tissue. However, the challenge of "learning” model parameters from uncertain patient-specific data remains. This is key to our ability to simulate large patient cohorts and to overcome the limitations of operator-dependent and time consuming "manual" tuning often performed for these models.
Three examples will be discussed with specific focus on the effect of clinical data uncertainty. First, we show an example in the context of Stage I to Stage II single ventricle palliation surgery where distributions of boundary resistances are assimilated from uncertain clinical data and propagated to post-operative results. Second, we use adaptive MCMC to learn the parameters of 0D Norwood circulation models from uncertain blood pressures and flows, following the characterization of structural and practical identifiability metrics and taking advantage of the underlying compartmental model layout. Finally, we conclude by presenting a condensation approach allowing to significantly reduce the computational burden in iterative parameter estimation of full multi-scale models.
Monday, November 2, 2015 -
8:00am to 5:00pm

The inaugural Women in Data Science Conference will be held at Stanford University on Monday, November 2, 2015. 

Click here to view the official event site:

Thanks to amazing response, all tickets for this conference are sold out.  Check back to this website soon for information on how you can be our 'virtual guests' for the conference via webstream.  If you would like to receive wait list notifications, conference updates, and live streaming notifications, sign up for our mailing list.

Women in Data Science Conference

Our aim is to inspire, educate and support women in the field – from those just starting their journey to those who are established leaders in industry, academia, government and NGO’s.    

Join us for this one-day technical conference to:

  • Hear about the latest data science-related research in a number of domains
  • Learn how leading-edge companies are leveraging data science for success
  • Brainstorm during lunchtime breakout session about ways to better support women in data science 
  • Connect with potential mentors, collaborators, funders and others in the eco-system

View the conference presentations here:

Tuesday, October 27, 2015 -
12:30pm to 1:30pm


Speaker: Jenny Suckale

Title: Fire and Ice: Evolving Interfaces in Magma and Ice Flows
Abstract: Multi-phase interactions are fundamental to many questions in Earth science. Given the nonlinear nature of the governing equations, numerical methods play an important role in advancing our basic understanding and predictive capabilities of multi-phase flows. One important challenge in computational approaches for simulating geophysical flows is the accurate representation and tracking of interfaces between different phases. In this study, we focus on evolving interfaces in magmatic and ice flows to illustrate when and how interface dynamics can have profound influence on the overall behavior of the geophysical system.
Monday, October 26, 2015 -
4:30pm to 5:30pm

Speaker: Nicolas Kseib, Mechanical Engineering at Stanford University

Title: Data and Physics Driven Physical Modeling

Abstract: As computing capabilities grow and the amount of experimental and numerical data increases, we can design computational strategies to automatically test and assess different modeling assumptions. We introduce a general data-driven statistical framework that bridges the gap between (numerical or laboratory) experimentation, physical modeling and uncertainty quantification. The framework enables the study of uncertainties and biases in physical models estimated from data. We differentiate between two types of modeling uncertainties and biases, the first one due to physical errors in the models and the second one due to noise introduced by the data-acquisition process. We also present different procedures to build models under different noise assumptions and propose a metric to quantify the quality of the data driven estimations. The framework is tested in the context of combustion science and chemical kinetics and driven by empirical data and simple reactive flow models.


Monday, October 26, 2015 -
4:30pm to 5:30pm

Speaker: Gustavo Schwenkler, Boston University

Title: Efficient Parameter Estimation for multivariate Jump-Discussions

Abstract: This paper develops an unbiased and computationally efficient Monte-Carlo estimator of the transition density of a multivariate jump-diffusion process. The drift, volatility, jump intensity, and jump magnitude are allowed to be state-dependent and non-affine. Most importantly, it is not necessary that the variance-covariance matrix can be diagonalized using a change of variable or change of time. Our density estimator facilitates the parametric estimation of multivariate jump diffusion models based on low frequency data. The parameter estimators we propose have the same asymptotic behavior as maximum likelihood estimators under mild conditions that can be verified using our density estimator. Numerical case studies illustrate our results. Joint work with François Guay.

Thursday, October 22, 2015 -
4:30pm to 5:45pm

Speaker: David Bindel, Assistant Professor in the Department of Computer Science at Cornell University

Title: Model Reduction for Edge-Weighted Personalized PageRank

Abstract: Work on model reduction for fast computation of PageRank for graphs in which the edge weights depend on parameters will be described.  For an example learning-to-rank application, our approach is nearly five orders of magnitude faster than the standard approach. This speed improvement enables interactive computation of a class of ranking results that previously could only be computed offline.  While our approach draws on ideas common in model reduction for large physical simulations, the cost and accuracy tradeoffs for the edge-weighted PageRank problem are different, as we will describe. 

This is joint work with Wenlei Xie, Johannes Gehrke, and Al Demers

Visit David Bindel's website at: