Alex Bäuerle, M. Sc.
I completed my masters degree at the Institute of Media Informatics 2017 before joining the research group Visual Computing. Since then, I work on visualization in the domain of artificial intelligence.
- Neural Network Visualization
- Explainable AI
- AI Fairness
I supervise(d) the following lectures:
Theses and Projects
I would gladly supervise theses from the field of neural networks. I am especially interested in generating visualizations of networks, training results and training data. We can, however, also discuss other topics in the field of neural Networks via mail or in my office.
Important: The selection of projects that can be found below is only a subset of possibilities. Working on your own ideas is always possible. Also, most of the time, I have some additional projects ready to be tackled. Therefore, it's always worth making a personal appointment.
Concept Activation Vectors (CAVs) are widely used to find out what a network is interested in when making a classification. These are obtained by a manually selected set of examples that are divided in context images and non-context images. However, do these activation vectors really always represent the contexts we are searching for?
Explaining Reinforcement Learning
Reinforcement Learning is really interesting, and gained traction with its application to robots, board games, and even modern video games.
Currently, reinforcement learning lacks a common debugging and visualization tool. Such a tool could provide a unified interface across different tasks and problems.
Neural networks have evolved a lot over time. There have been different architectures used over time and new problems and techniques have emerged. Comparing these is a tedious task, especially when not trained on the same dataset.
This work should give a historic overview of the development of these Networks and make comparing them possible. The interactive visualization should respect different properties of the network and make them easily comparable. It should be publicly accessible and thus web-based.
PointNets are used to train neural networks on pointclouds. While there are many visualization techniques for image-based neural networks, PointNets are lacking such visualizations.
This work aims at providing such visualizations for PointNets. It includes reading up on existing visualization techniques for ImageNets, but should also contain the elaboration of new visualization options for these types of networks.
Respiratory Motion Detection
This master thesis which was submitted by Christoph Baldauf with the title "Convolutional Neural Networks (CNN) Applied to Respiratory Motion Detection" was done in cooperation with the centre for translational imaging. During TAVI procedures, the patient is supervised with imaging technology throughout the whole operation. To support doctors, overlays on these images are used. Since the patient is breathing however, there overlays are not always positioned correctly.
To compensate this, a neural network was trained to recognize and compensate the vertival movement of the chest, and this way reposition such overlays.
In cooperation with the Institute of Protein
Biochemnistry, this project was aimed to automatically detect crossovers in micro-ct scans of fibrils. These crossovers are important for analyzing and reconstructing the scanned fibrils.
To enable this automatic localization of crossovers, a neural network for semantic segmentation of fibril images was trained as a student project.