Participants 2017

Joshua Eric Romulo B. Uyheng

Name  

Joshua Eric Romulo B. Uyheng

University

Ateneo de Manila University

Supervisor

Prof. Stefan Riezler

Workgroup  

Statistical Natural Language Processing Laboratory

Project

Consistency of Human Feedback for Bandit Learning

Text

Bandit learning refers to a stochastic optimization framework from partial feedback which has been used in the evaluation of machine translation systems. The evaluation of a machine-generated translation involves two important dimensions: the adequacy of the translation (whether it accurately conveys the correct meaning) and its fluency (whether it is a fluent statement in the target language). Without access to expensive gold standards, crowd-sourcing allows for inexpensive and large-scale evaluation of machine translations. However, members of the crowd often do not provide consistent evaluations, either with respect to their own previous evaluations or the evaluations of others. Drawing insights from psychometric techniques, our research aims to develop a crowd-sourcing task that assesses and maximizes the consistency of human feedback on the adequacy and fluency of machine translations using bandit learning.

Feng Jiang

Name  

Feng Jiang

University

Beijing University of Posts and Telecommunications

Supervisor

Prof. Dr. Jürgen Hesser

Workgroup  

Experimental Radiation Oncology

Project

Gesture Control for Handling a Mobile Visite Tool

Text

Touch-less hand gesture recognition systems are becoming important in automotive user interfaces as they improve safety and comfort. The aim of the project was to develop a framework that is able to detect hands, track them in realtime and carry out some dynamic gesture recognition results. It is to be done with simple image processing techniques with a regular laptop web-camera.  The whole project was configured on MacOS via C++ with toolbox OpenCV. The pre-processing stages mainly use the built-in functions, which consists of skin segmentation, contour extraction, calculating  bounding rectangle, motion tracking. The direction of motion of the gesture is determined by two steps. Firstly, it checks a hand is moving vertically or horizontally, i.e. along x-axis or y-axis. Then to specifically determine the direction movement by comparing the current position and the mid-position point throughout the eight frames. Finally, it could classify four moving directions. In this way, it can prevent over-recognition, especially some slight but meaningless shake.

Vema Sundeep

Name  

Vema Sundeep

University

Indian Institute of Technology(IIT) Delhi

Supervisor

Dr. Thomas Carraro

Workgroup  

Numerical methods for multiscale models

Project

Development of Multi-radii Neumann P2D models for Lithium-ion batteries

Text

Lithium-ion batteries have extensive applications in portable devices and lately stationary energy storage. It is important to understand the internal potential and concentration of Li-ion distributions to understand the aging process. My work was focused on extending the already existing one electrode model to couple with separator and finally simulate a complete battery i.e with both electrodes and separator in between. I coupled thermal equation to find out the temperature distribution within the battery. This is similar to Neumann P2D model employed in simulating Li-ion cells but with a provision to include multi-radii active material distribution in the electrodes. We have got some interesting results which we need to verify with experimental results and find an appropriate reasoning. The next step would be to perform real microstructure simulations and compare its results with this model. This is an important comparison since real microstructure simulations require high computational power, if the results are comparable then we can use this as a substitute for real microstructure simulations. Being from a chemical engineering background, I initially faced difficulty in understanding the FEM methods, but the friendly nature of my group members made the learning process very enjoyable. I have learned a lot and made some life-long friends during my stay.

Anirudh Sivakumar

Name  

Anirudh Sivakumar

University

D.G Vaishnav College

Supervisor

PD Dr Ahmad Hujeirat

Workgroup  

Computational Astrophysics

Project

Glitches: The exact quantum signature of pulsars

Text

Pulsars are born with an embryonic ,rigid-body ,rotating super baryon at its center surrounded by dissipative nuclear medium. When the pulsars rotate, there is a continuous loss in their rotational energy. But the energy curve shows glitches, which are caused due to the transition of the super baryon between rotational energy levels. While rotating this core will expel vortices as it transitions, which will then reconnect and dissipate as kelvin waves.  The objective of the project is to produce a simulation of quantum vortices and see how they behave in an ambient medium, in the form of rotation, bending and reconnecting with each other, In this case we place a ambient medium inside a cylindrical container and a spherical solid core. We also study how the core plays a role in modifying the dynamics of these vortices by adding perturbations to the surface of the sphere.The code to produce all of this was developed by Andrew Baggaley from School of Mathematics and Statistics, Newcastle University. It includes fortran and MATLAB source files which help perform the calculations and output the simulation respectively. We obtain the state of the superfluid, i.e it's wave function by solving the Gross-Pitaevskii equation. The code allows us to alter the parameters like the trapping potential, the grid points in the domain, initial condition , dimensions of the vortex lattice etc.But this simulation does not focus on the expulsion phase of vortices and also the formation of kelvin waves, it also considers only one superfluid whereas a more accurate simulation demands another superfluid in the boundary layer between the ambient medium and the baryon core. The study is under modification to enable random motion of vortices , and ejection of vortices from central core as dictated by Onsager-Feynman equation, which should then be absorbed by ambient medium and lead to spin up of the outer crust.

Zehranaz Dönmez

Name  

   Zehranaz Dönmez

University

   Bilkent University

Supervisor

   Dr. Michael Winckler


Project

   Spring Motion Simulation

Text      

During my research at IWR, I have worked on simulation of spring motion. As basic physics suggest, there are some main shapes the spring can get eventually after waving. Via Java, locations of masses as they form a spring are defined to simulate change in the location as time passes.Their instant locations are used to make graphs via Excel tools to simulate that they will move according to the theoretical laws suggest. It was a great pleasure to get this opportunity to improve my academic skills in Past-Bachelor Program.

Kira Vinogradova

Name  

Kira Vinogradova

University

Moscow Institute of Physics and Technology

Supervisor

Prof. Dr. Fred Hamprecht

Workgroup  

Image Analysis and Learning

Project

Deep Learning for Connectomics

Text

During my internship in the group, I trained convolutional neural nets and also retrieved Random Forest's predictions from ilastik, which is a software tool for interactive image analysis, segmentation, and classification. This fast-working and easy-to-use tool was developed in the Image Analysis and Learning lab to help scientists speed up the image analysis. The idea of my project was quite similar.Let's imagine we have a relatively small data set of raw biological images with so-called answers (for example, images with some structures labeled in ilastik), but we don't have ground truths for them because manual labeling is very time-consuming or we can produce only a few ground truths. We aimed to build a model which would be able to predict the answers for another set of very similar images without the manual labeling. The idea was to train U-Net architecture, known for its great performance on segmentation tasks, on ilastik's predictions.I preprocessed the retrieved predictions and the  images from ISBI 2012, 2D segmentation challenge of neuronal processes in EM images, in several ways and trained U-Net on various data subsets and on different combinations of ilastik's predictions and ground truth samples. The result was surprising: U-Net improved Random Forest's predictions without the requirement of ground truth. On the contrary, when a small amount of ground truth was added the samples lead to a decrease in the model's performance.

Navdha Malhotra

Name  

Navdha Malhotra

University

Delhi Technological University

Supervisor

Prof. Dr. Ekaterina Kostina

Workgroup  

Numerical Optimization

Project

l1- Norm Constrained Least Squares Optimization Methods

Text

This project involved studying and understanding of various approaches to find a solution for the l2 norm minimization problem with respect to l1 constraints. This problem arises from the estimation of linear models, and is of great interest and applicability in numerous fields due to its utilization in predictive control and variable selection. We studied the least squares form of the problem and obtained a differentiable convex objective function for the problem, and examined five methods for optimization for this type of problem, including two classic ones, one based on a sign- constrained formulation and the other on a formulation with linear constraints using non-negative variables, and popular methods like Interior-Point, Active-Set and modified LARS. We implemented the aforementioned methods for a small – sized sample problem (with 10 predictor variables and 440 instances). We used a standardized version of the dataset, and obtained results for a large range of values of the tuning parameter δ (corresponding to the l1 constraint). On examining the results, we found them to be coherent with the sparsity inducing property of the l1 constraint. Sparsity of the solution increased with lower values of the parameter δ. All algorithms converged fairly fast, with comparable CPU times. As we were dealing with a small problem, our results were comparable for all the algorithms, but the variability of these methods might be evident while dealing with larger problem sizes.