- You are here:
- GT Home
- Home
- News & Events

Monday, November 14, 2016 - 14:05 ,
Location: Skiles 005 ,
Dr. Maryam Yashtini ,
Georgia Tech Mathematics ,
Organizer: Martin Short

Many real-world problems reduce to optimization problems that are solved

by iterative methods. In this talk, I focus on recently developed

efficient algorithms for solving large-scale optimization problems that

arises in medical imaging and image

processing. In the first part of my talk, I will introduce the Bregman

Operator Splitting with Variable Stepsize (BOSVS) algorithm for solving

nonsmooth inverse problems. The proposed algorithm is designed to handle

applications where the matrix in the fidelity

term is large, dense, and ill-conditioned. Numerical results are provided

using test problems from parallel magnetic resonance imaging. In the

second part, I will focus on the Euler's Elastica-based model which is

non-smooth and non-convex, and involves high-order

derivatives. I introduce two efficient alternating minimization methods

based on operator splitting and alternating direction method of

multipliers, where subproblems can be solved efficiently by Fourier

transforms and shrinkage operators. I present the analytical

properties of each algorithm, as well as several numerical experiments

on image inpainting problems, including comparison with some existing

state-of-art methods to show the efficiency and the effectiveness of the

proposed methods.

by iterative methods. In this talk, I focus on recently developed

efficient algorithms for solving large-scale optimization problems that

arises in medical imaging and image

processing. In the first part of my talk, I will introduce the Bregman

Operator Splitting with Variable Stepsize (BOSVS) algorithm for solving

nonsmooth inverse problems. The proposed algorithm is designed to handle

applications where the matrix in the fidelity

term is large, dense, and ill-conditioned. Numerical results are provided

using test problems from parallel magnetic resonance imaging. In the

second part, I will focus on the Euler's Elastica-based model which is

non-smooth and non-convex, and involves high-order

derivatives. I introduce two efficient alternating minimization methods

based on operator splitting and alternating direction method of

multipliers, where subproblems can be solved efficiently by Fourier

transforms and shrinkage operators. I present the analytical

properties of each algorithm, as well as several numerical experiments

on image inpainting problems, including comparison with some existing

state-of-art methods to show the efficiency and the effectiveness of the

proposed methods.

Monday, November 7, 2016 - 14:05 ,
Location: Skiles 005 ,
JD Walsh ,
GA Tech Mathematics, doctoral candidate ,
Organizer: Martin Short

The boundary method is a new algorithm for solving semi-discrete

transport problems involving a variety of ground cost functions. By

reformulating a transport problem as an optimal coupling problem, one

can construct a partition of its continuous space whose boundaries allow

accurate determination of the transport map and its associated

Wasserstein distance. The boundary method approximates region boundaries

using the general auction algorithm, controlling problem size with a

multigrid discard approach. This talk describes numerical and

mathematical results obtained when the ground cost is a convex

combination of lp norms, and shares preliminary work involving other

ground cost functions.

transport problems involving a variety of ground cost functions. By

reformulating a transport problem as an optimal coupling problem, one

can construct a partition of its continuous space whose boundaries allow

accurate determination of the transport map and its associated

Wasserstein distance. The boundary method approximates region boundaries

using the general auction algorithm, controlling problem size with a

multigrid discard approach. This talk describes numerical and

mathematical results obtained when the ground cost is a convex

combination of lp norms, and shares preliminary work involving other

ground cost functions.

Tuesday, November 1, 2016 - 14:05 ,
Location: Skiles 006 ,
Dr. Mehdi Vahab ,
Florida State University Math ,
Organizer: Martin Short

An adaptive hybrid level set moment-of-fluid method is developed to study

the material solidification of static and dynamic multiphase systems. The

main focus is on the solidification of water droplets, which may undergo

normal or supercooled freezing. We model the different regimes of freezing

such as supercooling, nucleation, recalescence, isothermal freezing and

solid cooling accordingly to capture physical dynamics during impact and

solidification of water droplets onto solid surfaces. The numerical

simulations are validated by comparison to analytical results and

experimental observations. The present simulations demonstrate the ability

of the method to capture sharp solidification front, handle contact line

dynamics, and the simultaneous impact, merging and freezing of a drop.

Parameter studies have been conducted, which show the influence of the

Stefan number on the regularity of the shape of frozen droplets. Also, it

is shown that impacting droplets with different sizes create ice shapes

which are uniform near the impact point and become dissimilar away from it.

In addition, surface wettability determines whether droplets freeze upon

impact or bounce away.

the material solidification of static and dynamic multiphase systems. The

main focus is on the solidification of water droplets, which may undergo

normal or supercooled freezing. We model the different regimes of freezing

such as supercooling, nucleation, recalescence, isothermal freezing and

solid cooling accordingly to capture physical dynamics during impact and

solidification of water droplets onto solid surfaces. The numerical

simulations are validated by comparison to analytical results and

experimental observations. The present simulations demonstrate the ability

of the method to capture sharp solidification front, handle contact line

dynamics, and the simultaneous impact, merging and freezing of a drop.

Parameter studies have been conducted, which show the influence of the

Stefan number on the regularity of the shape of frozen droplets. Also, it

is shown that impacting droplets with different sizes create ice shapes

which are uniform near the impact point and become dissimilar away from it.

In addition, surface wettability determines whether droplets freeze upon

impact or bounce away.

Monday, October 24, 2016 - 14:05 ,
Location: Skiles 005 ,
Prof. Lars Ruthotto ,
Emory University Math/CS ,
Organizer: Martin Short

Image registration is an essential task in almost all areas involving

imaging techniques. The goal of image registration is to find

geometrical correspondences between two or more images. Image

registration is commonly phrased as a variational problem that is known

to be ill-posed and thus regularization is commonly used to ensure

existence of solutions and/or introduce prior knowledge about the

application in mind. Many relevant applications, e.g., in biomedical

imaging, require that plausible transformations are diffeomorphic, i.e.,

smooth mappings with a smooth inverse.

imaging techniques. The goal of image registration is to find

geometrical correspondences between two or more images. Image

registration is commonly phrased as a variational problem that is known

to be ill-posed and thus regularization is commonly used to ensure

existence of solutions and/or introduce prior knowledge about the

application in mind. Many relevant applications, e.g., in biomedical

imaging, require that plausible transformations are diffeomorphic, i.e.,

smooth mappings with a smooth inverse.

This talk will present and compare two modeling strategies and numerical

approaches to diffeomorphic image registration. First, we will discuss

regularization approaches based on nonlinear elasticity. Second, we will

phrase image registration as an optimal control problem involving

hyperbolic PDEs which is similar to the popular framework of Large

Deformation Diffeomorphic Metric Mapping (LDDMM). Finally, we will

consider computational aspects and present numerical results for

real-life medical imaging problems.

Monday, October 17, 2016 - 14:05 ,
Location: Skiles 005 ,
Prof. Yanzhao Cao ,
Auburn University Mathematics ,
Organizer: Martin Short

A nonlinear filtering problem can be classified as a stochastic Bayesian

optimization problem of identifying the state of a stochastic dynamical

system based on noisy observations of the system. Well known numerical

simulation methods include unscented Kalman filters and particle

filters. In this talk, we consider a class of efficient numerical

methods based on forward backward stochastic differential equations.

The backward SDEs for nonlinear filtering problems are similar to the

Fokker-Planck equations for SDEs. We will describe the process of

deriving such backward SDEs as well as high order numerical algorithms

to solve them, which in turn solve nonlinear filtering problems.

optimization problem of identifying the state of a stochastic dynamical

system based on noisy observations of the system. Well known numerical

simulation methods include unscented Kalman filters and particle

filters. In this talk, we consider a class of efficient numerical

methods based on forward backward stochastic differential equations.

The backward SDEs for nonlinear filtering problems are similar to the

Fokker-Planck equations for SDEs. We will describe the process of

deriving such backward SDEs as well as high order numerical algorithms

to solve them, which in turn solve nonlinear filtering problems.

Monday, October 3, 2016 - 14:00 ,
Location: Skiles 005 ,
Prof. Qin Li ,
UW-Madison ,
qinli@math.wisc.edu ,
Organizer: Molei Tao

Many kinetic equations have the corresponding fluid limits. In the zero limit of the Knudsen number, one derives the Euler equation out of the Boltzmann equation and the heat equation out of the radiative transfer equation. While there are good numerical solvers for both kinetic and fluid equations, it is not quite well-understood when the two regimes co-exist. In this talk, we model the layer between the fluid and the kinetic using a half-space equation, study the well-posedness, design a numerical solver, and utilize it to couple the two sets of equations that govern separate domains. It is a joint work with Jianfeng Lu and Weiran Sun.

Monday, September 12, 2016 - 14:05 ,
Location: Skiles 005 ,
Prof. Jacob Eisenstein ,
GA Tech School of Interactive Computing ,
Organizer: Martin Short

Language change is a complex social phenomenon, revealing

pathways of communication and sociocultural influence. But while language

change has long been a topic of study in sociolinguistics, traditional

linguistic research methods rely on circumstantial evidence, estimating the

direction of change from differences between older and younger speakers. In

this research, we use a data set of several million Twitter users to track

language changes in progress. First, we show that language change can be

viewed as a form of social influence: we observe complex contagion for

``netspeak'' abbreviations (e.g., lol) and phonetic spellings, but not for

older dialect markers from spoken language. Next, we test whether specific

types of social network connections are more influential than others, using

a parametric Hawkes process model. We find that tie strength plays an

important role: densely embedded social ties are significantly better

conduits of linguistic influence. Geographic locality appears to play a

more limited role: we find relatively little evidence to support the

hypothesis that individuals are more influenced by geographically local

social ties, even in the usage of geographical dialect markers.

pathways of communication and sociocultural influence. But while language

change has long been a topic of study in sociolinguistics, traditional

linguistic research methods rely on circumstantial evidence, estimating the

direction of change from differences between older and younger speakers. In

this research, we use a data set of several million Twitter users to track

language changes in progress. First, we show that language change can be

viewed as a form of social influence: we observe complex contagion for

``netspeak'' abbreviations (e.g., lol) and phonetic spellings, but not for

older dialect markers from spoken language. Next, we test whether specific

types of social network connections are more influential than others, using

a parametric Hawkes process model. We find that tie strength plays an

important role: densely embedded social ties are significantly better

conduits of linguistic influence. Geographic locality appears to play a

more limited role: we find relatively little evidence to support the

hypothesis that individuals are more influenced by geographically local

social ties, even in the usage of geographical dialect markers.

Monday, August 8, 2016 - 14:00 ,
Location: Skiles 006 ,
Prof. Yunho Kim ,
UNIST, Korea ,
Organizer: Sung Ha Kang

Inspired by the usefulness of difference of convex functions in some problems, e.g. sparse representations, we use such an idea of difference of convex functions to propose a method of finding an eigenfunction of a self-adjointoperator. In a matrix setting, this method always finds an eigenvector of a symmetric matrix corresponding to the smallest eigenvalue without solving Ax=b. In fact, such a matrix A is allowed to be singular, as well. We can apply the same setting to a generalized eigenvalue problem. We will discuss its convergence as well.

Wednesday, June 22, 2016 - 11:00 ,
Location: Skiles 006 ,
Dr. Ha Quang, Minh ,
Istituto Italiano di Tecnologia (Italy) ,
Organizer: Sung Ha Kang

Symmetric positive definite (SPD) matrices play important roles in numerous areas of mathematics, statistics, and their applications in machine learning, optimization, computer vision, and related fields. Among the most important topics in the study of SPD matrices are the distances between them that can properly capture the geometry of the set of SPD matrices. Two of the most widely used distances are the affine-invariant distance and the Log-Euclidean distance, which are geodesic distances corresponding to two different Riemannian metrics on this set. In this talk, we present our recently developed concept of Log-Hilbert-Schmidt (Log-HS) distance between positive definite Hilbert-Schmidt operators on a Hilbert space.This is the generalization of the Log-Euclidean distance between SPD matrices to the infinite-dimensional setting. In the case of reproducing kernel Hilbert space (RKHS) covariance operators, we obtain closed form formulas for the Log-HS distance, expressed via Gram matrices. As a practical example, we demonstrate an application of the Log-HS distance to the problem of image classification in computer vision.

Monday, April 11, 2016 - 14:05 ,
Location: Skiles 005 ,
Byungmoon Kim ,
Adobe Research ,
Organizer: Yingjie Liu

This talk will tell the story on using simulation for painting. I will tell a few of

projects that had simulation and painting involved. One is iPad-based ultra-low-cost real

time simulation of old photography process to compute effects that modern day users may

find interesting. The other is more full-blown fluid simulation for painting using

highest-end GPU. Even with massive processing power of GPU, real time high fidelity

painting simulation is hard since computation budget is limited. Basically we should deal

with large errors. It may sound odd if someone says that very low-accuracy simulation is

interesting - but this is very true. In particular, we tried to pull most pressure effect

out from about 10 Jacobi iterations that we could afford. I would like to share my

experience on improving fixed number of fixed point iterations.

projects that had simulation and painting involved. One is iPad-based ultra-low-cost real

time simulation of old photography process to compute effects that modern day users may

find interesting. The other is more full-blown fluid simulation for painting using

highest-end GPU. Even with massive processing power of GPU, real time high fidelity

painting simulation is hard since computation budget is limited. Basically we should deal

with large errors. It may sound odd if someone says that very low-accuracy simulation is

interesting - but this is very true. In particular, we tried to pull most pressure effect

out from about 10 Jacobi iterations that we could afford. I would like to share my

experience on improving fixed number of fixed point iterations.