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!
Pruning image super-resolution models by removing unnecessary ReLU activations.
Type of Work:
- Guided Research
- Master
Keywords:
- Deep Learning
- Image Processing
- Image Super-Resolution
Description:
This work investigates the optimization of image super-resolution neural network architectures by removing ReLU and other noise-canceling activation layers. The resulting method should combine convolution layers surrounding the removed activation layers into a single convolution layer, reducing redundancy and improving computational efficiency. As a starting point, the selection of ReLU layers for removal will be based on an analysis of their activations distributions (non-noise-canceled vs noise-canceled) using a representative dataset.
Neural ODEs for Adaptive GAN Training
Type of Work:
- Bachelor
- Master
Keywords:
- GANs
- Neural Ordinary Differential Equations
Description:
The goal of this work is to integrate Neural Ordinary Differential Equations (Neural ODEs) into the training of Generative Adversarial Networks (GANs). While GANs are powerful and effective, they are notoriously difficult to train due to instability and mode collapse, stemming from the adversarial nature of the training framework. At the same time, Neural ODEs have demonstrated parameter efficiency by modeling data transformations as a continuous process. This project aims to leverage this property to enable GANs to dynamically adjust the required function evaluations during training, allowing the model to adapt as the generator improves.
- [1] Generative Adversarial Networks, https://arxiv.org/abs/1406.2661
- [2] Neural Ordinary Differential Equations, https://arxiv.org/abs/1806.07366
- [3] Training Generative Adversarial Networks by Solving Ordinary Differential Equations, https://arxiv.org/abs/2010.15040
Latent Generative Adversarial Networks
Type of Work:
- Master
Keywords:
- GAN
- Image Generation
- VAE
Description:
This project aims to train a Generative Adversarial Network (GAN) in latent space, following the principles of latent transformer and diffusion models for image synthesis. Instead of directly modeling pixels, which is computationally expensive and prone to collapse, this approach operates in a lower-dimensional latent space, where data is compressed into meaningful representations. The GAN learns to generate images by synthesizing these latent representations, rather than individual pixels, which simplifies the training process and reduces resource demands.
- StyleGAN-T Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis, https://arxiv.org/abs/2301.09515
- Taming Transformers for High-Resolution Image Synthesis, https://arxiv.org/abs/2012.09841
- High-Resolution Image Synthesis with Latent Diffusion Models, https://arxiv.org/abs/2112.10752
Combining Dynamic Attention-Guided Diffusion and Wavelet-Based Diffusion for Image Super-Resolution
Type of Work:
- Guided Research
- Master
Keywords:
- deep learning
- single image super-resolution
- vision transformer
Description:
This thesis focuses on merging two techniques developed in our group [1, 2]. The first component, Dynamic Attention-Guided Diffusion, allows selective diffusion across regions of interest in the image, driven by time-dependent attention mechanisms. This method ensures that only certain parts of the image are diffused at specific time-steps, enhancing focus on critical image regions. The second component, Wavelet-based Diffusion, introduces image processing in the frequency domain via discrete wavelet transforms (DWT). Instead of working in the pixel domain, this method applies diffusion in the frequency domain, effectively capturing and enhancing multiscale image details. By combining these approaches, this work will explore the synergy of frequency-domain wavelet transforms with dynamic, time-based attention in diffusion models. The research aims to produce sharper, high-resolution images by diffusing across relevant areas in both the spatial and frequency domains, leading to more efficient and accurate SR results.
- [1] Waving Goodbye to Low-Res: A Diffusion-Wavelet Approach for Image Super-Resolution
- [2] Dynamic Attention-Guided Diffusion for Image Super-Resolution
Semantic Segmentation with Efficient Pixel-Transformers
Type of Work:
- Master
Keywords:
- efficient ML
- semantic segmentation
- vision transformer
Description:
Have you ever wondered how machines can “see” the world like we do, assigning labels to every single detail in an image? This field, called semantic segmentation, is crucial for tasks like self-driving cars and medical image analysis.
Transformers have conquered every aspect of machine learning. In particular in computer vision ViT [1] and variants of it are used. Because of their $\mathcal O(N^2)$ computational complexity in the length $N$ of the input sequence, transformers struggle with very long inputs. To combat this problem, an image is usually not fed into the transformer as is, but it is first cut up into $16 \times 16$ pixel segments, the so-called patches, which form the input sequence. This can cause problems for the task of semantic segmentation, since the goal is to capture pixel-level labels. Currently the workaround is to use complicated decoder networks to get back the pixel-level information from the patch-level output of a transformer model [2].
This project tackles this head-on, with the goal to use recent efficient transformer models that scale linearly $\mathcal O(N)$ or $\mathcal O(N \log N)$ to circumvent the decoding-problem by making the transformer use pixels instead of patches as the input. This way no complicated decoder is needed; a simpler, more powerful approach to segmentation using transformer models!
- [1] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- [2] Semantic Segmentation using Vision Transformers: A survey
Dataset Distillation for Fast Proxy Evaluation of Generative Models
Type of Work:
- Master
Keywords:
- dataset distillation
- generative models
- text-to-image evaluation
Description:
In recent years, text-to-image generative models have advanced significantly. Traditionally, evaluating these models relies on generating thousands of images and comparing against a large dataset of real image. Evaluating these models is crucial but the substantial computational resources make regular performance monitoring challenging. Dataset distillation offers a solution by condensing the information into a smaller set of synthetic samples. This project aims to use dataset distillation to create a compact dataset that serves as a proxy for full dataset evaluation.
- [1] FlashEval: Towards Fast and Accurate Evaluation of Text-to-image Diffusion Generative Models, https://arxiv.org/abs/2403.16379
- [2] Latent Dataset Distillation with Diffusion Models, https://arxiv.org/abs/2403.03881
Image Super-Resolution in both ways
Type of Work:
- Bachelor
- Guided Research
Keywords:
- auto-encoder
- deep learning
- single image super-resolution
Description:
The goal of this project is to develop and evaluate a novel dual-decoder architecture for image super-resolution (SR) [1]. This architecture will utilize a single encoder to extract features from an input image, followed by two decoders: one trained to map the features to a low-resolution (LR) output, and the other to map the features to a high-resolution (HR) output. This approach aims to enhance the SR performance by leveraging the complementary learning objectives of both decoders. The goal of the work is to try different architectures and to analyze different loss formulations as well as the feature space learned by the encoder.
Sherlock Holmes goes AI - Generative comics art of detective scenes and identikits
Type of Work:
- Master
Keywords:
- Bias in image generation models
- Deep Learning Frameworks
- Frontend visualization
- Speech-To-Text, Text-to-Image Models
- Transformers, Diffusion Models, Hugging Face
Description:
Sherlock Holmes is taking the statement of the witness. The witness is describing the appearance of the perpetrator and the forensic setting they still remember. Your task as the AI investigator will be to generate a comic sketch of the scene and phantom images of the accused person based on the spoken statement of the witness. For this you will use state-of-the-art transformers and visualize the output in an application. As AI investigator you will detect, qualify and quantify bias in the images which are produced by different generation models you have chosen.
Note:
This work is embedded in the DFKI KI4Pol lab together with the law enforcement agencies. The stories are fictional you will not work on true crime.
Requirements:
- German level B1/2 or equivalent
- Outstanding academic achievements
- Motivational cover letter
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.
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.
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.
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.