Open Thesis Topics

Our research group offers a variety of projects (Bachelor theses, Practicals, Master theses) on the following topics:

  • Causality and causal inference
  • Machine learning and causal modeling in cognitive neuroscience
  • Brain-Computer Interfaces (BCIs) for communication and rehabilitation

Below is a list of open topics. If you are interested in a particular topic, please send an email to the contact person listed underneath the project description. If you would like to suggest a topic of your own, please contact moritz.grosse-wentrup@univie.ac.at.

Projects for P1 and/or P2

PyTorch implementation of BunDLe-Net (P1/P2 project)

BunDLe-Net is a state-of-the-art neural network architecture that extracts visual insights from complex neuronal and behavioural data. It has been successfully deployed to interpret high-dimensional time-series neuronal data from varying sources: roundworm, monkey, rat, fish. The architecture and learning is currently implemented in TensorFlow. As part of your project, you will implement BunDLe-Net in PyTorch, optimise the learning process, and make it easily applicable to neuroscience data across a range of modalities.

Contact person: Akshey Kumar

 

Potential Projects in Interpretable Reinforcement Learning (P1)

Recent work from our group (Kumar et al., 2023, link: https://www.biorxiv.org/content/10.1101/2023.08.08.551978v1 ) found meaningful abstractions of neuronal dynamics on a lower-dimensional manifold. The crux to the work is that it preserves only the information relevant for a certain behavioural context and thus is discarding potential noise and irrelevant information. The original data consists of neuronal recordings from a small worm and corresponding behaviour, typically certain kinds of movements.

If you transfer the idea to reinforcement learning (RL) the neuronal data comes from an RL agent’s policy function and the predicted/chosen action serves as the behavioural context. Since for such a pairing the question of why a certain action was chosen by the RL agent is typically not comprehensible, i.e. the action prediction is not interpretable, we aim to apply aforementioned manifold learning to comprehend the RL agent’s actions. While there is no concrete topic formulation, you are invited to approach us with your own ideas regarding the topic or just have a chat with us about potential project ideas, if you are interested in making RL interpretable.

Contact person: Christoph Luther

 

Benchmarking Causal Structure Learning for inference of cond. independencies (P1 or P2)

Causal graphs, typically directed acyclic graphs, are a great way to visualise dependencies between random variables of a probability distribution. They come along with their own representation of statistical independence called d-separation. Under the two standard assumptions that the given distribution satisfies the Markov assumption and is faithful w.r.t. the graph both concepts are equivalent. Since there is a plethora of algorithms to efficiently infer such graphs from data and conditional independence testing directly from data is a challenging problem, in this project you are asked to compare whether the procedure of first estimating such graphs from data and then (efficiently) reading off d-separations can replace explicit conditional independence tests. For the project, you are asked to acquaint yourself with causal structure learning algorithms and apply (already implemented) concepts to data and read off d-separations (conditional independences). You shall then compare the results to those of standard conditional independence tests (like partial correlation tests for Gaussian data) and evaluate which approach is more efficient as well as accurate. In order to have a ground truth to compare to, you would mostly rely on synthetic data sampled from a known distribution.

Depending on the extent of the benchmark and the coverage of theory, the project can be either P1 or P2.

If you are generally interested in causal inference/causal structure learning, you can also look at interesting algorithms and for example implement them yourself and apply them to synthetic data. An interesting direction is the application to (simulated) neuronal data.

Contact person: Christoph Luther

 

Auditory cortex clustering (P1)

This project is in collaboration with the Brain and Language Lab (Narly Golestani) at the Vienna Cognitive Science Hub. The aim is to test different clustering algorithms on a dataset of auditory cortex structural magnetic resonance imaging data, focusing on the Heschl's gyrus (a main “language area” in the brain). The first goal is to cluster a large group of participants in a data driven manner and find an optimal clustering that is relatively consistent across clustering algorithms and can deal with missing data. Once this is done, the clustering can be used to predict demographic and language aptitude information.

Contact person: Jozsef Arato

 

Eye-movement similarity (P1)

The goal is to participate in the development of a python package for eye-movement data analysis (fixations, scanpath). The project is based at the Vienna Cognitive Science Hub and involves working on data from art history, psychology and open eye-movement datasets. The goal is to add new functionalities to the package (e.g., improved time-series analysis), making it more user friendly, and testing the algorithms on different data-sets, to come up with the best default settings, and work toward an eventual publication as open source software.

Contact person: Jozsef Arato

Projects for Bachelor & Master theses

Why correlation does not imply causation – except that it often does.

Reichenbach's principle states that if two variables are correlated, then either one is a cause of the other or the two variables share a common cause. As such, it is clear that correlation does not generally imply a causal relationship. However, in practice, scientists and humans, in general, are usually guided in the discovery of causal relationships by strong correlations. This raises the question of whether strong correlations often indicate a cause-effect relationship? In this project, you will run large-scale simulations of causal graphs and analyze the induced correlations between variables. Based on these simulations, you will study whether strong correlations are more indicative of a causal relationship than weak correlations.

This project is suitable for a Bachelor's thesis as well as a Master's thesis. While a Bachelor's thesis would focus more on the practical aspects, i.e., the simulations, a Master's thesis should also include a more theoretical argument on why strong correlations often imply causal relationships.

Contact person: Moritz Grosse-Wentrup