Viviane Kakerbeck

Portfolio

About Me

Hi, I'm Viviane Kakerbeck, a student of Cognitive Science at the University of Osnabrück. I am interested in how the human brain works and how these findings can be applied in various other fields. Lately I've especially focused on deep learning algorithms, data science, new user interfaces and new ways to convey and learn knowledge. I am very interested in how humans explore, understand and interact with the world and how this knowledge can improve our ways to do so. I wrote my Bachelor's Thesis on eye tracking in VR and am still working at the followup projects as a research assistant. I am currently writing my masters thesis on progressively growing neural networks for scene graph generation from images. On this website you can see some of my recent projects.

B. Sc. in Cognitive Science (2015-2017)
M. Sc in Cognitive Science (2017-2018)

Pursuing a PhD in Computational Cognition (since 12.2018) 

  Twitter: @vkakerbeck

  GitHub: https://github.com/vkakerbeck/

  LinkedIn: www.linkedin.com/in/viviane-kakerbeck

Recent Awards:

 - Award for extraordinary performance during the studies (Given by MLP Finanzdienstleistungen AG) 2018

 - Science Price of Lower Saxony 2018 (

https://www.noz.de/lokales/osnabrueck/artikel/1595577/osnabruecker-studentin-gewinnt-wissenschaftspreis-niedersachsen-2018)

- Award for extraordinary Master Thesis (given by ROSEN Gruppe) 2019


My Projects

Progressively Growing Neural Networks

For my master thesis I implemented neural networks which can progressively grow and learn. As an example task I used image recognition and three common data sets of different sizes and difficulties (MNIST, CIFAR-10 and CIFAR-100). The network can for example start training on 50 classes and the progressively learn more classes by expanding its output layer. This ability of the network to learn even after an initial training is extremely useful in real world applications where new data constantly becomes available after the deployment of a neural network. The code as well as my thesis can be found on GitHub:

https://github.com/vkakerbeck/Progressively-Growing-Networks

  • Growing works the best when re-initializing the last layer with the already trained weights.
  • New classes and nodes can be introduced without much drop in the network performance.
  • The performance drop mainly results from the new class being learned. The old classes stay stable.
  • When training for longer progressively growing archives similar results to a normal training.
  • Freezing the convolutional layer leads to performance reduction. But not much.
  • There are differences in the difficulty of the individual classes.
  • Different orders of class introductions lead to a higher variance in test accuracies.
  • Training with increasing difficulty of the task is significantly better than decreasing difficulty.
  • When introducing classes in increasing difficulty the accuracy exceeds a normal network training.

Hack4Health - Hackathon 2018

During the course of four day we implemented a prototype of a website for disease spread predictions and won the first price with this. The hackathon was organized by the University of Osnabrück and the Robert Koch Institut which also provided us with a big amount of data about disease spread over the past 15 years in Germany. With this data we trained three different networks which were ultimately stacked on top of each other to predict the spread of a disease in Germany for the upcoming weeks. These predictions are then integrated into a website which also collects real time data to make new predictions and provides the Robert Koch Institute with live statistics. Two newspaper articles as well as a radio broadcast reported about our project:



Health Predictions from Big Data
I've been training several neural networks to predicts treatments from diagnosis's of a patient as well as to predict the code sent to the health insurance and the money the hospital receives for the patient. The networks are then integrated into an interactive user interface such that they are easy to handle for doctors, hospitals or patients.


A Quantitative Analysis of Artistic Styles:
In this study project we investigate weather the amount of similarity between art styles can be measured with empirical data. To test this we trained CycleGANs on seven different art styles and use them to create stimuli. These stimuli are photographs modified in the seven different artistic styles. They are then used in a Pop-Out experiment as well as an eye tracking experiment. From the results of the two experiment we hope to be able to create a taxonomy of art styles which is similar to the opinion of art experts.

Data Analysis and Visualization
Here are just some examples of what I've analyzed and how I visualized our results. The full results as well as code will be added to my GitHub soon.


Seahaven - The VR City

During the course of my bachelors thesis as well as my research assistant job I've designed a VR city used in several experiments investigating spatial navigation. Additionally I've implemented eye tracking in VR and a simple analysis pipeline of it.

The following video is a little demonstration of the VR city and the workflow in it. It shows how we do eye tracking in VR and what kind of analysis you can do with the collected data. We hope it can inspire many future research projects. It is a great technology combining the benefits of a more natural environment than conventional screen-experiments and a controlled environment where it is easier to analyze the data than in real world experiments. [Publication Following]


Data Analysis and Visualization
Of course I also spent much time analyzing and visualizing the recorded results. Overall we've recorded more than 200 subjects in VR so far as well as 60 subjects in a comparison condition (map training). Here is a selection of visualizations I enjoyed to create for these results.
Example Code: https://github.com/vkakerbeck/NBP-VR-Lab/blob/master/Analysis/MapAnalysis/MapTrainingAnalysis.ipynb

Apartment Search
This project was a little side idea I had while looking for new apartments. It is very time consuming to flip through pages and pages of apartment offers on different portals so I was wondering if there is a better way to get a quick overview. Currently the data gets crawled from immobilienscout42 and wg-gesucht which are two big German platforms to find an apartment. This data can then be visualized in some graphs depicting overall statistics as well as on an interactive map which overlays a map with the apartment locations and visualizes their size and price-size relation colorfully.
Code: https://github.com/vkakerbeck/AppartementSearch

 
 

 
Sensor Data Visualization in 3D, AR & VR
At the institute for business computer science I've been working on live sensor data visualizations in 3D. For this I use the 3D game engine Unity as well as different AR and VR products such as the HTC Vive, Oculus Rift and the Microsoft HoloLens. The data is managed and indexed in Elastic Search which makes fast retrieval and querying possible. I've played around with different visualization techniques for a high dimensional data set of sensor values recorded from an industrial machine. Additionally I've created a live visualization of sensor data from the Texas Instruments SensorTag in which data such as sensor temperature, room temperature, sensor orientation, humidity and light intensity are intuitively visualized in a 3D and VR compatible scene. This will be compared to other visualizations (sliders & graph) in experiments.


Articles on Medium - Towards Data Science
I've just recently become a writer for Towards Data Science (https://towardsdatascience.com/). Check out my first article about animated graphs in Python here: https://towardsdatascience.com/how-to-create-animated-graphs-in-python-bb619cc2dec1


Sin - The Movie
This is a very personal project unrelated to my studies but still very important to me. Since the beginning of 2017 I've been working on the production of a 90 minute feature film called 'Sin'. We have people from 11 different countries and 6 continents working on this movie. It is about and opioid addicted drifter who gets involved in the criminal underworld with the theft of a precious statue of the Babylonian god 'Sin'. The movie addresses the current opioid epidemic in the United States and the many deaths and destroyed lives due to opioid addictions.

I play the role of Alethea in the movie but have also been helping out with anything else around the set. Aside from that I have designed the movies website https://sin-themovie.com/ and helped with all other promotional and organizational work. I've also researched and created the graphics for the short opioid documentary shown below and have done a big part of the editing of it.

Follow us and stay updated!
https://www.facebook.com/SinTheFilm/
Twitter: @Sin_The_Movie
 

The official trailer to the movie:

The documentary on opioids in the USA: