Date/Time : 12 Aug 2020 16:00
Venue : via Zoom
Speaker : Dong Bin, Peking University

Join Zoom Meeting https://nus-sg.zoom.us/j/93373644793?pwd=Y2VrdnJRazZnRXllN2N5K1BneWE1UT09 Meeting ID: 933 7364 4793 Password: 711028 In this talk, I will start with a brief review of the dynamics and optimal control perspective on deep learning (including supervised learning, reinforcement learning, and meta-learning), especially the so-called unrolled dynamics approach and its applications in medical imaging. Then, I will present some of our recent studies on how this new approach may help us to advance CT imaging and image-based diagnosis further. Specifically, I will focus on our thoughts on how to combine the wisdom from mathematical modeling with ideas from deep learning. Such combination leads to new data-driven image reconstruction models and new data-driven scanning strategies for CT imaging, and with a potential to be generalized to other imaging modalities.

Add to calendar :

Date/Time : 22 Jan 2020 16:00
Venue : S17 #05-11 SR5
Speaker : Lucas Lestandi, Nanyang Technological University

Date/Time:22 Jan 2020 16:00 Venue: S17 #05-11 SR5 Speaker: Lucas Lestandi, Nanyang Technological University Breaking the curse of dimensionality : from large data to reduced order models. It is known to every scientist who works with numerical simulation – that increasing complexity in the computing task leads to exploding computing times. This can then be materialized in many ways e.g. very fine meshes, extremely small time steps, multiscale or multiphysics problems and most importantly, large number of dimensions. The last issue (large of dimensions) is often referred as the curse of dimensionality. It is very common in fields as diverse as finance, chemistry or “simple” mechanical engineering optimization problem, parametric studies, etc. In this talk, we will cover some of the essential aspects in building reduced order models that can overcome some of the above difficulties. First, we introduce low rank approximation of high dimensional data. These tools enable both data reduction (by several orders) and yield reduced bases that are used to build the actual reduced order model (ROM). Then we will see how common ROMs are built and the issues that remain to be solve. Finally, we will have a look at the future of ROM, in which deep learning is set to play a large part. Add to calendar: 

Add to calendar :

Date/Time : 03 Jan 2020 16:00
Venue : S17 #04-06 SR1
Speaker : HAN Jiequn, Princeton University

Date/Time:03 Jan 2020 16:00 Venue: S17 #04-06 SR1 Speaker: HAN Jiequn, Princeton University Deep BSDE: A Unified Framework for Solving High-Dimensional PDEs, FBSDEs, and Control Problems Developing algorithms for solving high-dimensional partial differential equations (PDEs), forward-backward stochastic differential equations (FBSDEs), and control problems has been an exceedingly difficult task for a long time, due to the notorious difficulty known as the curse of dimensionality. In this talk we introduce the “deep BSDE method” as a unified framework to solve general high-dimensional PDEs and FBSDEs. Starting from the BSDE formulation, we approximate the unknown component by neural networks and design a proper loss function to optimize parameters. Numerical results of a variety of examples, including applications in game theory and eigenvalue problem, demonstrate that the proposed algorithm is quite effective in high- dimensions, in terms of both accuracy and speed. We furthermore provide a theoretical error analysis to illustrate the validity and property of the objective function. Add to calendar: 

Add to calendar :

Date/Time : 26 Dec 2019 16:00
Venue : S17 #05-11 SR5
Speaker : Hu Jingwei, Purdue University

Date/Time:26 Dec 2019 16:00 Venue: S17 #05-11 SR5 Speaker: Hu Jingwei, Purdue University A second-order asymptotic-preserving and positivity-preserving exponential Runge-Kutta method for a class of stiff kinetic equations We introduce a second-order time discretization method for stiff kinetic equations. The method is asymptotic-preserving (AP) — can capture the Euler limit without numerically resolving the small Knudsen number; and positivity-preserving — can preserve the non- negativity of the solution which is a probability density function for arbitrary Knudsen numbers. The method is based on a new formulation of the exponential Runge-Kutta method and can be applied to a large class of stiff kinetic equations including the BGK/ES-BGK equation (relaxation type), the Fokker-Planck equation (diffusion type), and even the full Boltzmann equation (nonlinear integral type). Furthermore, we show that when coupled with suitable spatial discretizations the fully discrete scheme satisfies an entropy-decay property. Various numerical results are provided to demonstrate the theoretical properties of the method. This is joint work with Ruiwen Shu (University of Maryland). Add to calendar: 

Add to calendar :

Date/Time : 13 Nov 2019 16:00
Venue : S17 #05-11 SR5
Speaker : Li Ji, National University of Singapore

Date/Time:13 Nov 2019 16:00 Venue: S17 #05-11 SR5 Speaker: Li Ji, National University of Singapore Numerical Algorithms of Phase Retrieval Phase retrieval is of paramount importance in a broad range of applications in engineering and imaging sciences. Mathematically, it amounts to solving a system of nonlinear quadratic equations, which is generally NP-hard. In this talk, we will give a numerical review of the problem and emphasize the different interests between the applied mathematicians and physicians. In other words, much research achievement of applied mathematicians has not intrigue the interests of physicians. One such reason is the simpler landscape of the resulting formulation. For example, the ambiguity solutions for practical phase retrieval is more complicated than phase retrieval from random sensing measurements. Thus the sound good solvers for random phase retrieval fails in locating a satisfying solution for practical phase retrieval. After we review the numerical algorithms for phase retrieval in literature, we introduce our recent work on phase retrieval, which tries its best to close the gap of random and practical phase retrieval. Our work gives some implications for designing the numerical algorithm for practical phase retrieval. Add to calendar: 

Add to calendar :

Date/Time : 30 Oct 2019 16:00
Venue : S17 #05-11 SR5
Speaker : Luminita Danaila, University of Rouen Normandy

Date/Time:06 Nov 2019 16:00 Venue: S17 #05-11 SR5 Speaker: Ionut Danaila, University of Rouen Normandy Numerical methods and finite-element toolboxes for the simulation of Bose-Einstein condensates We present several numerical tools using classical finite elements with mesh adaptivity for solving different models used for the study of Bose-Einstein condensates. The programs are written as toolboxes for FreeFem++ (www.freefem.org), a free finite-element software, allowing to easily implement various numerical algorithms [1]. For solving the stationary (imaginary-time) Gross-Pitaevskii equation, we use two robust and optimised numerical methods: a steepest descent method based on Sobolev gradients and a minimization algorithm based on the state-of-the-art optimization library IPOPT. A very recent conjugate-gradient method using concepts of Riemannian optimization is also presented [2]. For the Bogoliubov-de Gennes system, representing a linearisation of the Gross-Pitaevskii equation, a Newton method and a fast algorithm based on ARPACK for the calculation of eigenvalues are available. For the real-time Gross-Pitaevskii equation, classical splitting and relaxation methods were implemented and intensively tested. Validations and illustrations are presented for computing difficult configurations with vortices observed in physical experiments: single-line vortex, Abrikosov lattice, giant vortex, dark/anti-dark solitons in one or two-component Bose-Einstein condensates [3]. [1] G. Vergez, I. Danaila, S. Auliac, F.Hecht, A finite-element toolbox for the stationary Gross-Pitaevskii equation with rotation, Computer Physics Communications, 209, p. 144-162, 2016. [2] I. Danaila, B. Protas, Computation of Ground States of the Gross-Pitaevskii Functional via Riemannian Optimization, SIAM J. Sci. Comput. Vol. 39, No. 6, pp. B1102-B1129, 2017. [3] I. Danaila, M. A. Khamehchi, V. Gokhroo, P. Engels and P. G. Kevrekidis, Vector dark-antidark solitary waves in multicomponent Bose-Einstein condensates, Phys. Rev. A 94, 053617, 2016. Add to calendar: 

Add to calendar :

Date/Time : 09 Oct 2019 16:00
Venue : S17 #05-11 SR5
Speaker : CAI Yongyong, Beijing Computational Science Research Center

Date/Time:09 Oct 2019 16:00 Venue: S17 #05-11 SR5 Speaker: CAI Yongyong, Beijing Computational Science Research Center Numerical methods for Bogoliubov excitations of Bose-Einstein condensates We study the analytical properties and the numerical methods for the Bogoliubov-de Gennes equations (BdGEs) describing the elementary excitation of Bose-Einstein condensates around the mean field ground state, which is governed by the Gross-Pitaevskii equation (GPE). Three numerical methods are proposed to solve the BdGEs, including sine-spectral method, central finite difference method and compact finite difference method. Extensive numerical tests are provided to validate the algorithms and confirm that the sine-spectral method has spectral accuracy in spatial discretization, while the central finite difference method and the compact finite difference method are second-order and fourth-order accurate, respectively. Finally, sine spectral method is extended to study the elementary excitations under optical lattice potential and solve the BdGEs around the first excited states of the GPE. Add to calendar: 

Add to calendar :

Date/Time : 02 Oct 2019 16:00
Venue : S17 #05-11 SR5
Speaker : SHAO Sihong, Peking University

Date/Time:02 Oct 2019 16:00 Venue: S17 #05-11 SR5 Speaker: SHAO Sihong, Peking University Lovász extension and graph cut A firm bridge between discrete data world and continuous math field should be tremendously helpful in data analysis. Along this direction, the Lovász extension provides a both explicit and equivalent continuous optimization problem for a discrete optimization problem, for instance, the Cheeger cut problem. In this talk, we report a set- pair Lovász extension which provides not only an answer to the dual Cheeger cut, anti-Cheeger cut, and max 3-cut problems, all of which cannot be handled by the Lovász extension, but also works for the Cheeger cut and maxcut problems. In particular, the set-pair Lovász extension enlarges the feasible region of resulting equivalent continuous optimization problems from a half space (resulted from the Lovász extension ) to the entire space for the Cheeger cut and maxcut problems. On the other hand, it provides new possibilities for designing continuous optimization algorithms for combination problems on the practical side, too. As an illustration, we propose a simple continuous iterative algorithm for the maxcut problem which converges to a local optimum in finite steps. Here ‘simple’ means the inner subproblem is solved analytically and thus no optimization solver is called. Preliminary results on G-set demonstrate that the ratio between the best cut values achieved by the simple algorithm without any local breakout techniques and the best known ones is of at least $0.986$. Add to calendar: 

Add to calendar :

Date/Time : 11 Sep 2019 16:00
Venue : S17 #05-11 SR5
Speaker : Krishna Balasubramanian, UC Davis

Date/Time:11 Sep 2019 16:00 Venue: S17 #05-11 SR5 Speaker: Krishna Balasubramanian, UC Davis Applications of Gaussian Stein’s Identity to Non-Convex Optimization and Sampling Optimization and sampling are arguably the computational backbones of Frequentist and Bayesian statistics respectively. In this talk the use of Gaussian Stein’s identity for performing stochastic Zeroth-Order (ZO) non-convex optimization and sampling will be discussed. Specifically, we consider the case when the function (or density) under consideration is not available to us analytically, rather we are able to only obtain noisy evaluations. Using Gaussian Stein’s identity, techniques for estimating the gradient and Hessian of a function (or density) will be introduced. Based on this, the following algorithms and the corresponding theoretical results will be discussed: (i) ZO stochastic gradient method in high-dimensions (ii) ZO Newton’s method for escaping saddle points (iii) ZO Euler and Ozaki discretized (Kinetic) Langevin Monte Carlo sampling. Add to calendar: 

Add to calendar :