The following are some examples of topics we offer for Bachelor's and Master's theses. If you cannot find a suitable topic but would like to work on something related to data science, machine learning, or statistical signal processing, with applications in mobile communications, neuroscience, or medicine, please talk to us.

Übrigens: Wir schreiben unsere Themen generell auf Englisch aus. Sie dürfen Ihre Abschlussarbeit bei uns aber auch gerne auf Deutsch schreiben.

Interpretable machine learning techniques for medical data

suitable for: Bachelor’s or Master’s students

Prerequisites: Basic knowledge of machine learning, experience with Python and Tensorflow.
Harsh Panwar et al.
Contact: Maurice Kuschel

Machine learning algorithms have become state-of-the-art in various sorts of prediction and classification tasks. However, their inexplicable behavior poses problems in applications where thorough understanding and accountability of the decision-making algorithm are needed. One such field is that of medicine and diagnostics. Even though an algorithm might be able to accurately diagnose patients based on their health data, the responsible practitioner would like to understand the reasoning behind a prediction to validate the result. Therefore, it is crucial not only to develop accurate machine learning, but also to focus on interpretability and transparency.

The goal of this thesis is to develop and evaluate state-of-the-art methods that explain a decision of a machine learning algorithm. For example, there already are methods like GRAD-CAM [Selvaraju et al., 2017] and Integrated Gradients [Sundararajan et al., 2017] that visualize the attention of a neural network, as well as related variants for time-series data. The focus can be either on image datasets, for example X-Ray images, or wrist-band sensor data. The scope of the thesis can be adjusted according to your prerequisites.

Nonlinear multiview analysis using deep learning-based solutions for biomedicine

suitable for: Bachelor's or Master's students (with adjusted scope)

Prerequisites: Knowledge of statistical signal processing and machine learning. Experience with Python is helpful.
Image Credit:
Contact: Tanuj Hasija

In recent times, we have seen a vast increase in gathering of data where a common set of objects are observed through multiple views. For example, the brain function is now analyzed using multiple modalities like functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) as they provide complementary information. The fMRI has a very good spatial resolution, but a poor temporal resolution, whereas EEG has a high temporal resolution but a poor spatial localization. Similarly, use of multiple modalities such as heart rate and sweat activity have led to an increased detection of epileptic seizures than using each of them separately. Identifying the common latent features (and their relationships across different views) is vital in finding biomarkers for brain functioning and seizure prediction.

It is of urgent interest to improve available tools and theory for nonlinear multiview analysis where each view’s data is generated through a nonlinear mixing of the common latent features. Recent solutions based on deep learning provide completely data-driven way of incorporating the nonlinearities through use of neural networks (especially autoencoders) but are not theoretically well-studied.

In this thesis, you will develop novel deep learning methods which provide theoretical guarantees on feature identifiability. The practical issue of overfitting which stems from the methods to blindly minimize loss functions and discover spurious relationships will also be closely analyzed. The developed methods will then be applied to the data available from either of the two real-world applications mentioned above.

Tensor Decompositions for classification of patients with epilepsy and healthy controls based on medical data

suitable for: Bachelor's students

Prerequisites: Knowledge of linear algebra. Experience with MATLAB or Python is helpful.
3-dimensional tensor with 3 modes: subjects (i.e., patients), time, channels
Contact: Isabell Lehmann

More than 65 million people worldwide, which is nearly 1% of the world population, suffer from epilepsy. Epilepsy is a disease that causes uncontrolled seizures, which are sudden voltage discharges in the brain. These discharges can cause changes in behavior, movements or feelings, and in levels of consciousness. Under physical exhaustion, people with epilepsy (PWE) show an increased risk for cardiac death. Since the heart is under direct control of the brain through the autonomic nervous system, this thesis aims to investigate whether the brain reacts differently in an exhaustive exercise situation for PWE compared to healthy controls.

The goal of this thesis is to detect exercise-related patterns that differentiate between PWE and healthy controls using statistical signal processing and machine learning techniques, in particular, tensor decompositions. Tensors are multidimensional arrays, which preserve the higher-dimensional structure of the data. For example, time-series data collected from different subjects and channels can naturally be stored in a 3-dimensional tensor (see figure on the left). If meaningful patterns are found during the analysis, they can be used as features for classifying PWE and controls using machine learning techniques.

Interference management in wireless communication

suitable for: Master's students

Prerequisites: Knowledge of wireless communications and mathematical tools, i.e., linear algebra and/or optimization methods is required. Depending on your background, scope and focus of the task can be adjusted. Some prior experience with MATLAB is also expected.
Image Credit: Zahir, Talha, et al., IEEE Communications Surveys & Tutorials (2013) and  Yucek, Tevfik, and Huseyin, Arslan, IEEE Communications Surveys & Tutorials (2009)

Interference from other transmitters is the main bottleneck for most wireless communication systems, which are called interference-limited systems. Due to the bandwidth shortage, these systems must tolerate some interference in order to enhance the spectrum usage. There are different approaches to handle interference in wireless systems. Among them is employing improper signaling. In improper signals, the powers of the real and imaginary parts are not equal, and/or there is correlation between the real and imaginary parts of the signal. The capacity-achieving signal in traditional systems is a proper signal; however, it has been shown that improper signaling can increase the rate in interference-limited systems.

In this thesis, you will study the performance of an interference management technique in a wireless communication system. To this end, we first define a scenario in which the system is interference-limited. For instance, we can consider an underlay/overlay cognitive radio (CR) system. In CR systems, licensed or primary users (PU) share the spectrum with unlicensed or secondary users (SU) under the constraint that the PU's communications are not affected by the SU's transmission. In such system, we should design the transmission strategy of SU to get the desired system performance. Moreover, different transmission techniques (e.g., OFDM and single carrier) can be considered in the thesis. You will consider different interference management techniques including improper signaling and interference alignment. Then, you will optimize the performance of the defined scenario by analytical tools. Finally, you will evaluate your results by simulations using MATLAB.