Thesis Topics

This list includes topics for potential bachelor or master theses, guided research, projects, seminars, and other activities. Search with Ctrl+F for desired keywords, e.g. ‘machine learning’ or others.

PLEASE NOTE: If you are interested in any of these topics, click the respective supervisor link to send a message with a simple CV, grade sheet, and topic ideas (if any). We will answer shortly.

Of course, your own ideas are always welcome!


Context-Aware Recommender Systems for Personal Knowledge Assistants

Type of Work:

  • Bachelor
  • Guided Research
  • Master
  • Project
  • Seminar

Keywords:

  • context-aware recommender systems
  • graph embedding
  • information retrieval
  • knowledge worker actions and scenarios detection
  • personal information management
  • re-finding agents

Description:

One of the ways to assist knowledge workers in the daily tasks and improve their productivity is by providing them with relevant helpful information based on their current context. There are many challenges in developing such a recommender system which could be capable of recommending the right information at the right time to the right person. Understanding the contextual state of the users as precisely as possible, along with detecting their activities and information need play a significant role in enhancing the context-aware recommender systems for personal knowledge assistants. A mixture of knowledge and skills in various fields such as information retrieval, graph embedding, and machine learning is needed to tackle the existing challenges in this area.


Self-supervised Video Object Segmentation

Type of Work:

  • Master

Keywords:

  • self-supervision
  • video object segmentation

Description:

Exploring the potential advantages of integrating global context for self-supervised video object segmentation.


Graph Structure and Node Label Prediction on Circuit Diagrams

Type of Work:

  • Master

Keywords:

  • Autoencoder
  • GAN
  • Graph Neural Network
  • LSTM

Description:

The objective of this master’s thesis is to examine the performance of deep (graph) neural networks for node labeling and structure prediction on graph structures representing electrical circuit diagrams. In the course of the thesis, the following model types should be evaluated: Classification Models should be used to aid the human user to get insights in the circuit’s functional principle. For example, all nodes of a circuit should be classified whether or not they contribute to the power supply of the described device. This part of the thesis is intended to compete (partly) against the existing rule-based approach. Transformative Models should be trained in a natural-language translator-like fashion to map one circuit to another, in which the entire graph structure is mapped to a coding space by an encode an subsequently unwrapped by a decoder model. The output graph should be an alternate version of the input graph which desirable properties. For example, the output graph should represent an energy- optimized version of the input device. Optionally, the process of turning the content of a coding space into a graph structure should be examined here. Biologically inspired approaches should be considered in which a single node divides, and differentiates under the exchange of signals like cells in embryonic tissue (mitosis, apoptosis, release and reception of cytokines). Generative Models (trained e.g. in GAN structures) should allow for the high-dimension interpolation between graphs. This should facilitate the generation of new devices like “a mixture of an AM and an FM radio”, “a slightly more toaster-ish backing oven.”


Graph Extraction/Generation from Diagrams

Type of Work:

  • Bachelor
  • Master
  • Project
  • Seminar

Keywords:

  • cnn
  • cv
  • dynamic system simulation

Description:

I am currently supervising students in multiple topics related to the extraction and simulation of Graph-Based Engineering Drawings from optical sources: Graph Extraction from Printed/Handdrawn Circuit Diagrams/Piping&Instumentation Diagrams, Generation of Hand-Drawn Diagrams, Interactive Circuit Detection and Simulation


Efficient & data type independent CUDA kernels

Type of Work:

  • Bachelor
  • Guided Research
  • Master

Keywords:

  • cuda
  • efficiency
  • gpu
  • ptx assembly

Description:

With typical GPU code all operands must have the same data type. However, the speed of most operations on GPUs is limited not by computation, but by memory bandwidth and latency. In these cases using the smallest possible data types can greatly enhance performance. The standard approach of compiling one version of a kernel for each combination inputs is infeasible though, since the number of combinations, especially for 3 or more inputs, leads to exorbitantly large binaries and very long compilation times. Much more desirable is a flexible kernel that can load inputs of arbitrary data type and converts them to an intermediate representation that can be used to perform the actual computation.

Note: This topic is highly specialized and requires deep knowledge in CUDA C++ programming and PTX assembly. Please only apply if you have previously worked with those languages.


Multi-view Fusion with Redundant and Complementary Satellite Image Time Series (SITS) for Land Use Land Cover (LULC)

Type of Work:

  • Bachelor (A Smaller Version With Respect To Master)
  • Master Thesis

Keywords:

  • data fusion
  • deep learning
  • land use land cover segmentation
  • multi-view/modal learning
  • remote sensing
  • satellite image time series

Description:

Remote Sensing (RS) based applications are increasingly growing in recent years thanks to the use of Artificial Intelligence (AI) technology (mainly based on Deep Learning) that allows powerful solutions to problems of global interest. The use of multiple RS sources (satellites/sensors) to describe a region of interest and develop applications have become a ubiquitous approach, usually called multi-view learning. However, RS is a domain where views are quite heterogeneous. Since each sensor could have a different temporal and spatial resolution than the others used, it poses a challenge. Nevertheless, the information that each view contains could have different noises or be at different levels, where usually there is not a clear understanding of how similar or different are various RS sources.

Data focus:

Feel free to reach out if you have any question regarding the topic or have ideas related.


Innovative Knowledge Graph visualizations for end-users in corporate memories

Type of Work:

  • Bachelor
  • Master

Keywords:

  • corporate memory
  • knowledge graph
  • ontology

Description:

Topic just an example. If you are interested, just ping us, we may find a topic suting your expertise and interest in this field.

Our Corporate Memory CoMem uses knowledge graphs (KG) to represent personal and organizational knowledge such as documents, topics, appointments, contacts, projects, etc. The resulting KG are complex and currently can be browsed by end-users using “classic” views such as dashboards and widgets. In this thesis, new ways of interacting with the KG are sought with focus on innovative visualizations of the graphs. The usual generic graph visualizations will overwhelm users. Therefore, adaptive visualizations in different scenarios shall be investigated and implemented such as exploration of a user’s search process switching between graphs, lists, and tables, exploring a context, or focussing on dedicated questions to be answered in the graph visualization. An inspiration can be taken by the wikidata implementation at https://query.wikidata.org/ CoMem https://comem.ai


Efficient Sampling of Training Patches for Image Super-Resolution

Type of Work:

  • Guided Research
  • Master

Keywords:

  • benchmark
  • deep learning
  • proxy datasets
  • single image super-resolution

Description:

Typically, the Single Image Super-Resolution (SISR) training pipeline decomposes train images into sub-images. It follows the deep learning paradigm of “more data is better.” Be that as it may, it is not necessarily true [1, 2]. This work aims to rethink the image decomposition for the SISR training pipeline and test various sampling methods. The results should be evaluated using state-of-the-art deep learning models on standard datasets.


Attention for video object segmentation

Type of Work:

  • Master

Keywords:

  • reinforcement learning (rl)
  • video object segmentation

Description:

Exploring RL-based methods for temporal attention in video object segmentation


Knowledge Graphs für das Immobilienmanagement

Type of Work:

  • Bachelor
  • Master

Keywords:

  • corporate memory
  • knowledge graph
  • ontologie

Description:

Das Management von Immobilien ist komplex und umfasst verschiedenste Informationsquellen und -objekte zur Durchführung der Prozesse. Ein Corporate Memory kann hier unterstützen in der Analyse und Abbildung des Informationsraums um Wissensdienste zu ermöglichen. Aufgabe ist es, eine Ontologie für das Immobilienmanagement zu entwerfen und beispielhaft ein Szenario zu entwickeln. Für die Materialien und Anwendungspartner sind gute Deutschkenntnisse erforderlich.


Learning Analytics in Education

Type of Work:

  • Master
  • Project

Keywords:

  • affective state
  • cognitive state
  • deep learning
  • machine learning

Description:

None


Exploring the potential of hebbian plasticity (meta-learning) to treat adversarial attacks

Type of Work:

  • Master

Keywords:

  • adversarial attacks
  • deep learning
  • meta learning
  • reinforcement learning
  • robust models

Description:

Adversarial attacks are one of the biggest problems in the application of neural networks in safety-critical areas. The aim of this thesis is to evaluate to what extent dynamic parameters based on hebbian plasticity are suitable for handling these attacks. The aim of this work is to adapt this approach, which was originally introduced for reinforcement learning, for this purpose.


Deep Self-organizing feature maps in time series analysis

Type of Work:

  • Guided Research
  • Master

Keywords:

  • deep learning
  • python
  • self-organizing feature maps
  • sofm
  • time series analysis
  • unsupervised training

Description:

Deep Self-Organising Feature Maps (DSOFM) have shown that they are capable of capturing high-dimensional topologies and are furthermore suitable for image classification on a limited scale. The aim of this project is to apply a stacked version of DSOFM to data from the time analysis domain and to test its suitability for anomaly detection. In addition, advantages and disadvantages, as well as possible improvements, are to be highlighted.


Neural Architecture Search for Diffusion Models

Type of Work:

  • Guided Research
  • Master

Keywords:

  • benchmark
  • deep learning
  • denoising diffusion probabilistic models
  • single image super-resolution

Description:

Denoising Diffusion Probabilistic Models (DDPMs) or Diffusion Models became an effective tool for various tasks and also for Single Image Super-Resolution (https://iterative-refinement.github.io/). However, despite astonishing results, the design of the denoise function, i.e., the denoising neural network, is yet to be fully discovered. In this work, we want to try Neural Architecture Search approaches to explore the possibilities, advantages, and disadvantages of specific designs (https://arxiv.org/abs/1806.09055).


Anomaly detection in time-series

Type of Work:

  • Master
  • Project

Keywords:

  • cnn
  • explainability

Description:

Working on deep neural networks for making the time-series anomaly detection process more robust. An important aspect of this process is explainability of the decision taken by a network.


Understanding and enhancing model robustness against adversarial attacks

Type of Work:

  • Master

Keywords:

  • adversarial training
  • cnn
  • deep learning
  • robust models
  • visual recognition

Description:

None


Generative Adversarial Networks for Agricultural Yield Prediction

Type of Work:

  • Guided Research
  • Master

Keywords:

  • Deep Learning
  • Generative Adversarial Networks
  • Yield Prediction

Description:

Agricultural yield prediction has been an essential research area for many
years, as it helps farmers and policymakers to make informed decisions about
crop management, resource allocation, and food security. Computer vision and machine learning techniques have shown promising results in predicting crop
yield, but there is still room for improvement in the accuracy and precision
of these predictions. Generative Adversarial Networks (GANs) are a type of neural network that has shown success in generating realistic images, which can be leveraged for the prediction of agricultural yields.

References:

  • ‘Goodfellow, Ian, et al. “Generative adversarial networks.” Communications of the ACM 63.11 (2020)': 139-144.
  • ‘Z. Xu, J. Du, J. Wang, C. Jiang and Y. Ren, “Satellite Image Prediction Relying on GAN and LSTM Neural Networks,” ICC 2019 - 2019 IEEE International Conference on Communications (ICC), Shanghai, China, 2019, pp. 1-6, doi’: 10.1109/ICC.2019.8761462.
  • ‘Drees, Lukas, et al. “Temporal prediction and evaluation of brassica growth in the field using conditional generative adversarial networks.” Computers and Electronics in Agriculture 190 (2021)': 106415

Patch-based Image Synthesis of Complex Scenes

Type of Work:

  • Master

Keywords:

  • Adversarial image synthesis
  • GAN

Description:

Models such as the recently proposed InfinityGAN can generate high-resolution images in a patch-by-patch manner conditioned on coordinates. However, they typically only work on landscape datasets with repeatable structure and texture. The goal of this work is to extend the capabilities to more complex and object-centric datasets using additional supervision in the form of segmentation maps. [1] InfinityGAN, https://arxiv.org/abs/2104.03963


Time Series Forecasting Using transformer Networks

Type of Work:

  • Guided Research
  • Project

Keywords:

  • time series forecasting
  • transformer networks

Description:

Transformer networks have emerged as competent architecture for modeling sequences. This research will primarily focus on using transformer networks for forecasting time series (multivariate/ univariate) and may also involve fusing knowledge into the machine learning architecture.


Exploring Self Attention in Transformers

Type of Work:

  • Master
  • Project
  • Seminar

Keywords:

  • language generation
  • nlp
  • transformers

Description:

Transformers have shown a promising new direction in Language generation replacing recurrent neural networks. Different attention mechanisms have been attributed as possible cause for successful performance of these architectures. The goal of this topic would broadly be to develop on and improve the self attention mechanism used in the transformer architectures.


Construction & Application of Enterprise Knowledge Graphs in the E-Invoicing Domain

Type of Work:

  • Bachelor
  • Guided Research Project
  • Master

Keywords:

  • knowledge graphs
  • knowledge services
  • linked data
  • semantic web

Description:

In recent years knowledge graphs received a lot of attention as well in industry as in science. Knowledge graphs consist of entities and relationships between them and allow integrating new knowledge arbitrarily. Famous instances in industry are knowledge graphs by Microsoft, Google, Facebook or IBM. But beyond these ones, knowledge graphs are also adopted in more domain specific scenarios such as in e-Procurement, e-Invoicing and purchase-to-pay processes. The objective in theses and projects is to explore particular aspects of constructing and/or applying knowledge graphs in the domain of purchase-to-pay processes and e-Invoicing.


Fault and Efficiency Prediction in High Performance Computing

Type of Work:

  • Master Thesis

Keywords:

  • deep learning
  • event data modelling
  • survival modelling
  • time series

Description:

High use of resources are thought to be an indirect cause of failures in large cluster systems, but little work has systematically investigated the role of high resource usage on system failures, largely due to the lack of a comprehensive resource monitoring tool which resolves resource use by job and node. This project studies log data of the DFKI Kaiserslautern high performance cluster to consider the predictability of adverse events (node failure, GPU freeze), energy usage and identify the most relevant data within. The second supervisor for this work is Joachim Folz.

Data is available via Prometheus-compatible system:

Reference:

Feel free to reach out if the topic sounds interesting or if you have ideas related to this work. We can then brainstorm a specific research question together. Link to my personal website.

Previous Job Offers
Next Contact