Participants 2016

Aliaksandra Shysheya

Name  

Aliaksandra Shysheya

University

Belarusian State University

Supervisor

Prof. Dr. Fred Hamprecht

Workgroup  

Image Analysis and Learning

Project

Deep Learning

Text

I was working on Deep Learning project, on speeding up convolutional neural networks in particular. Nowadays CNNs (convolutional neural networks) are highly efficient being used in wide variety of applications in image and video recognition, recommender systems and natural language processing. Unfortunately, such a powerful mechanism as CNNs can not be embedded in a real-life product, as a deep neural network requires uttermost time to be trained from scratch on an ordinary CPU. So, my goal was to accelerate CNNs in order to make them more accessible to use in restricted environment that doesn't have GPUs. I was making CNNs faster by approximating an ordinary convolution layer with two sequential layers that are using horizontal and vertical filters only. As a result, I got about 4 times speedup on a CPU.

Panasun Manorot

Name  

Panasun Manorot

University

Chiang Mai University, Thailand

Supervisor

Prof. Dr. Peter Bastian

Workgroup  

Project Representation of soil water dynamic

Text

This project involves representation of soil water dynamics by using Geophysics Mathematics and computers to construct the groundwater model and compute numerical solution. Mathematical concepts of fluid dynamics are important knowledge that used to investigate dynamics of soil water. We use Darcy’s law, Conservation of mass, and Richards’s equation to represents two phase flow fluid (water and air) in porous median on 2D vertical confine aquifer domain with lakes source. Soil water dynamics model have been solved and simulate by finite element method and implemented by Distributed and Unified Numerical Environment (DUNE).Moreover, the simulation of groundwater will help us to understand about pressure behavior and water flow in porous median. This project can extend to represents groundwater dynamics with river or 3D lakes non-homogeneous domain.

Muhammad Qasim

Name  

Muhammad Qasim

University

COMSATS Institute of Information Technology, Islamabad, Pakistan

Supervisor

Prof. Dr. Peter Bastian

Workgroup  

Error measurement and FEM benchmark for phase field modeling

Project

Parallel Computing

Text

During my stay at IWR, university of Heidelberg, I have carried out research on the project “Error estimate and FEM solution of phase-field model”. For numerical implementation, I have used the Distributed and Unified Numerics Environment (Dune) software. The summer school has provided me an opportunity to learn different discretization techniques and their implementation in Dune-PDELab for numerical solution of the proposed model. The computed results in one and two dimensions have excellent agreement with existing numerical results. I have also computed the error estimates for computed solution. Besides, I also took part in Summer School program of university of Heidelberg in which I learned basic German language and culture as well with interaction with fellow from other countries and have enjoyed their wonderful friendships . As a whole my stay at university of Heidelberg was very fruitful both academically and socially.

Sothy Brokorb
Kao Sethkosa

Name  

Kao Sethkosal & Sothy Brokorb

University

Royal University of Phnom Penh

Supervisor

Dr. Michael Winckler

Workgroup  

3D Modeling

Project

Virtual Museum Prototype

Text

The project involves creating an interactive 3D prototype of the National Museum in Oman. We have researched on how to create a version that is mainly focused on functionality and extensibility of the prototype. For functional implementation, we have used Unity game engine which has a flexible editor, rich visualization and multiplatform deployment. We were building the models of the museum and designing the experience for the users. During the research, we have worked on solutions on how to get maximum performance in rendering the scenes and objects in the Museum, because the scanned models of statues and artifacts from Oman provided by IWR are in very high details. We implemented a technique called LOD (Level of Details), which constantly adjusts the most suitable version of objects by measuring the distance from objects to the visitor of Museum. Moreover, we implemented the objects by allowing users to be able to interactively view the statues or artifacts in detail. There are further features that could be worked on such as interactive map of the museum, using machine learning to study behavior patterns of users in which retrieved data can be used to provide more personalized experience.