2017 IACS Research Day Presenters

Automating the Penguin Census Pipeline

Wildlife censuses can be time consuming, logistically challenging and prohibitively expensive, however they are vital for our understanding of population level dynamics and responses to changing climate.  This is particularly true of penguin populations in the Antarctic, where rates of warming are among the highest in the world, and weather, rapidly shifting sea ice and the sheer scale of penguin colonies make censuses challenging.

We use a novel combination of computer vision, machine-learning, and statistical point pattern analysis to automate the census process. This allows us to rapidly count hundreds of thousands of nesting birds from both modern drone imagery and historic aerial photography surveys, expanding our records of abundance both spatially and temporally and allowing the construction of more accurate continental scale population models. Additionally, we can use the vast amounts of high-resolution spatial data, not available through traditional manual census techniques, to construct individual based models that should give us a new insight into the processes that occur within a colony and how these processes may be driving observed declines in penguin numbers.

Bio:  

Phil McDowall received his undergraduate and Master's degrees in quantitative ecology from St Andrews University in Scotland. He is currently a PhD candidate in the Ecology and Evolution department at Stony Brook where he studies the spatio-temporal dynamics of penguin colonies. While not doing fieldwork on the Western Antarctic Peninsula, Phil works on novel methods for tackling ecological problems in 3D, and the development of drone based personal-scale remote sensing platforms.

The Nascent Merger Between Remote Sensing and Computer Vision and its Impact on the Future of Spatial Ecology

As the spatial resolution of available satellite imagery has shrunk, the traditional tools of imagery classification have been slowly replaced with approaches deriving from machine learning and computer vision. Increasingly, spatial ecologists are able to map, model, and track populations of plants and animals across the globe, sometimes down to the scale of individual organisms. In this talk I will describe some of my lab’s work to use UAV’s, satellites, and other geospatial technologies to push the limits on our ability to study wildlife and their responses to changing environmental conditions. I will also highlight some of the ways in which this new frontier in spatial ecology will require interdisciplinary collaborations between ecologists, computer scientists, and mathematicians.

Heather Lynch, IACS Faculty
Ecology & Evolution

Bio:

Dr. Heather Lynch is an Associate Professor of Ecology & Evolution and an affiliated faculty member of IACS. She received her B.A. in Physics from Princeton University, an M.A. in Physics from Harvard University, and a Ph.D. in Organismic and Evolutionary Biology from Harvard University. Her research is focused on the distribution and abundance of Antarctic seabirds and marine mammals, with a particular focus on understanding the dynamics of populations in response to climate change and fisheries.

Performance of Applications of Parallel Markov Chain Monte Carlo Methods

We analyze a massive search space for finding optimal scalable parallelization strategies, including mixing patterns, and mixing periods, for rapid convergence of the parallel Markov chain Monte Carlo methods in optimization. We expect to find out the optimal time and ways for multiple Markov chains to communicate. Also, we adjust the sequential parameter, temperature, to fit for the parallel method. We examine the performance of our strategies by testing three applications: a benchmark optimization problem, Rastrigin function, a classic mathematical problem, Lennard-Jones problem, and a real world problem, mobile route recommendation problem. We find that our method outperforms most of the existing methods in all three problems and our parallel size can be scaled up to 384 for nearly 100% speedup. We believe there exists a scheme for these strategies leading to the optimal parallelization.

Bio: 

Zeyang Ye is originally from Guangdong in China and did some of his undergraduate study in Wuhan University. Midway through his undergrad years, however, he transferred to SBU where he double majored in math and applied math. His research focuses on trying to find a way through computation for a host of various problems to be solved as efficiently as possible.  Zeyang has presented his work at ICDM IEEE Conference 2015 on Data Mining, and he was awarded a $1700 travel grant from the National Science Foundation to attend SC15 where he presented his work on “Parallel Markov Chain Monte Carlo Methods” during the workshop on “Chinese HPC Research Toward New Platforms and Real Applications.” He also has one paper published in the ICDM workshop proceedings, and he has submitted two more journal papers, all as the first author. 

Computer-aided Design of Robust Performance-portable Algorithms and Implementations

Bio: 

Can Machine Learning Settle the Debate of the Dual Microscopic Character of Water?

25 years ago it was hypothesized that liquid water may be composed of two distinct types of molecular environments: one with molecules forming mostly tetrahedral structures via hydrogen bonds and a distorted one with interpenetrating hydrogen bond shells. Since then there have been many theoretical, experimental and computational studies hinting in this direction. However, the complexity of the microscopic structures that emerge in water make this a challenging question to answer and, to date, the debate is still open in the community. Powerful machine learning algorithms designed to extract features from complex data present a very promising tool to analyze the output data of computer simulations. In this talk I will motivate this long-standing debate, introduce the methods and challenges to perform computer simulations of liquid water and finally go over our latest findings about the proposed dual character of water using machine learning. 

Bio:

Adrián Soto is a PhD student in the Department of Physics and Astronomy working under the supervision of Mariví Fernández-Serra. His research advances the understanding of materials and their applications by mixing together 3 ingredients: theoretical physics, computational methods and machine learning. At the moment he is working on machine learning methods to better understand the local structure of liquid water and on studying electronic properties for dark matter detectors. He graduated in 2010 with a BS in Physics at the University of Valencia, Spain. He joined Stony Brook in 2010. Beyond academics, he has served in several committees and organizations to promote diversity in research and to approach science to the general public.

Towards Adaptive Hybrid Multigrid Method for PDE Based Systems

Modern applications often lead to large sparse linear systems arising from various PDE discretizations. Solving these linear systems is an integral part of most scientific simulations. Our earlier study showed that multigrid methods as preconditioners give very promising results in terms of robustness and efficiency. The Classical algebraic multigrid method (AMG) can give very efficient results with fine-tuned parameters, smoothers, interpolation operators, and coarsening strategies. However, the parameter tuning of AMG can be tricky due to its problem dependency. The geometric and P-multigrid multigrid methods are very effective and robust and require less parameter tuning, but they are also PDE dependent. We propose an adaptive, hybrid multigrid framework for PDE based systems, which will combine the ideas of algebraic multigrid with tuned parameters, geometric multigrid and p-multigrid. The framework is adaptive, in that it can adjust the parameters automatically based on the type of the problem using machine learning techniques. This self-adapting framework can potentially improve the productivity and efficiency for PDE-based modeling. We will report some preliminary results along this direction.

Bio:

Understanding Water/Solid Functional Interfaces for Photocatalysis and Electrochemical Applications

In this talk I will review our current efforts on understanding the physics of liquid water and the interaction of water with functional semiconductor and metallic surfaces using ab initio molecular dynamics methods. I will present the state of the art of current simulations and the challenges we face, focusing on two specific problems: the description of aqueous solvated electrode surfaces and the simulation of polar surfaces in aqueous environments. In particular I will show how the physics of ferroelectric and dielectric superlattices is strongly related to the physics that describes the behavior of water at polar interfaces. 

Marivi Fernandez-Serra,
IACS Faculty
Physics & Astronomy

Bio:

Marivi Fernandez-Serra is an associate professor in the department of Physics and Astronomy and a core faculty member of  IACS. She received her PhD in 2005 from the University of Cambridge and after 3 years of post-doc at the Ecole Normale Superieure (ENS) of Lyon she joined the Physics and Astronomy department at Stony Brook University in 2008. She is a computational condensed matter physicist with focus in the development and applications of electronic structure methods to study complex systems at the interface of physics, chemistry and materials science.