Department of Computer Science

Ongoing Theses

Transformers are all you need: Multimodal data generation

In her Master's thesis, Kathrin Seßler addresses the issue of the multimodal data generation using self-attention-based models, such as Transformers. The Master's thesis is supervised by Gjergji Kasneci and supported by Vadim Borisov in an advisory capacity.


Finished Theses

A Post-hoc Explainability Python Framework for Computer Vision Tasks using Superpixel Perturbations

In his Master's thesis, Felix Sieghörtner addressed the issue of the black-box model explainability. The Master's thesis was supervised by Gjergji Kasneci and supported by Vadim Borisov in an advisory capacity.

The abstract: 
Machine learning algorithms and deep neural networks see a significant rise in popularity for all kinds of problem-solving. Whether it is auto-correction in our cell phones or route-finding in our navigation systems, machine learning plays a key role in our daily lives. Some of them may have a heavy impact on human life, such as assisting doctors in selecting the appropriate treatment for patients. For this reason, the Explainability of Artificial Intelligence (XAI) is becoming an increasingly important field of study. Being able to explain decisions made by an AI, helps humans to understand its choices and is crucial for ensuring a wide acceptance among the public.
This thesis presents a Python framework for explainable AI. The key elements of it are two novel algorithms for generating and evaluating feature attributions for computer vision tasks. The developed perturbation- and permutation-based feature attribution generation method is model-agnostic and without the need to train any surrogate machine learning model. The evaluation method replaces superpixels with other superpixels and reevaluates the image afterward. In addition to a visual and empirical evaluation, these methods are compared to the current state-of-the-art feature attribution approaches, resulting in the novel methods performing on par or, in some cases, even better than existing ones.

Missingness in Evolving Data Streams

In his Master's thesis, Stefan Zürn addressed the concept of "missingness" in data streams. In the context of local explanation methods, "missingness" describes the absence of discriminative information and is usually represented by a specific baseline. The aim of the thesis was to develop an efficient adaptation of the "neutral" baseline for local explanation methods and online feature selection models in dynamic data streams. The Master's thesis was supervised by Gjergji Kasneci and supported by Johannes Haug in an advisory capacity.

Augmenting Local Attribution Methods with a Measure of Uncertainty

In his Bachelor's thesis, Manuel Weiß was working on an extension of the local explanation model LICON. The aim of the work was to adapt LICON for use in Bayesian neural networks. This step allows us to quantify the uncertainty with respect to the generated feature attributions, which makes explanations more robust and meaningful. The Bachelor's thesis was supervised by Gjergji Kasneci and supported by Johannes Haug in an advisory capacity.

Feature Selection for Multinomial Classification and Regression in Data Streams

In May 2021 Maximilian Bertsch finished his Bachelor's thesis about "Feature Selection for Multinomial Classification and Regression in Data Streams". The thesis was supervised by Gjergji Kasneci and advised by Johannes Haug. The abstract in german is available as an image on the right.

amorf - A Multi-Output Regression Framework

David Hildner is working on master's thesis, project entitled "amorf - A Multi-Output Regression Framework".

Multi-output regression is the problem of learning a mapping from a multi-dimensional real-valued input space to a multi-dimensional real-valued output space. So far no framework specifically designed for multi-output regression has been developed in Python. In this thesis a number of different multi-output regression methods are presented, evaluated and implemented into to the Python framework amorf. A novel approach, the autoencoder assisted multi-target regression, is introduced.

A Predictive Model for Classification in Data Streams

In March 2020 Alireza Izadchenas finished his Master's thesis about "A Predictive Model for Classification in Data Streams". The thesis was supervised by Gjergji Kasneci and advised by Johannes Haug. Information about the content can be found in the thesis announcement (image on the right).

Feature Selection on Streaming Data in Python

In January 2020 Lars-Christian Achauer finished his Master's thesis about "Feature Selection on Streaming Data in Python". The thesis was supervised by Gjergji Kasneci and advised by Johannes Haug. The abstract is available as an image on the right.

Truth Discovery with Uncertain Entity Reference

In July 2019 Daniel Sacher finished his Master's Thesis about "Truth Discovery with Uncertain Entity Reference". The thesis was supervised by Gjergji Kasneci. The abstract is available as an image on the right.

A Generative Graphical Model for Temporal Clustering of Multivariate Time Series Data

In March 2019 Garima Mittal finished her Master's Thesis about "A Generative Graphical Model for Temporal Clustering of Multivariate Time Series Data". The thesis was supervised by Gjergji Kasneci. The abstract is available as an image on the right.