Topics
Topic 1: Boolean Matrix Factorisation for analyzing genotype datasets (Sven Rahmann, Vu Lam Dang)
Boolean Matrix Factorisation aims to approximate a binary matrix as the boolean product of two low-dimensional matrices[^1]. Given a matrix $X\in \{0,1\}^{m\times n}$ and a rank $k\ll m,n$, the goal is to find 2 matrices $W\in \{0,1\}^{m\times k}$ and $H\in \{0,1\}^{k\times n}$ such that they minimize the objective function:
$$ ||X-W\circ H|| := \sum_{i,j} X_{ij} \oplus (W\circ H)_{ij}, $$
where $\oplus$ denotes the element-wise XOR operation, and $W\circ H$ denotes the boolean matrix product
$$ (W\circ H)_{ij} = \bigvee_{p=1}^k W_{ip} \wedge H_{pj} $$
for all $i\in \{1,\dots, m\}$ and $j \in \{1,\dots, n\}$.
There are several existing algorithms to find approximate solutions. Some of them are discrete by nature, for example, GRECOND+[^2] ASSO [^3], and others are based on surrogate functions in the continuous domain, for example, C-SALT [^4]. The performance of these methods on relatively large genotype matrices is unknown. Genotype matrices are sets of records on which version of a particular SNP presented in each sample. They are binary, unlike gene expression matrices, which contain continuous values. We are looking for students to assess these existing methods and possibly invent their own algorithm for BMF.
[^1]: Miettinen, P., & Neumann, S. (2020). Recent Developments in Boolean Matrix Factorization (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2012.03127
[^2]: Belohlavek, R., & Trnecka, M. (2018). A new algorithm for Boolean matrix factorization which admits overcovering. In Discrete Applied Mathematics (Vol. 249, pp. 36–52). Elsevier BV. https://doi.org/10.1016/j.dam.2017.12.044
[^3]: Miettinen, P., Mielikainen, T., Gionis, A., Das, G., & Mannila, H. (2008). The Discrete Basis Problem. In IEEE Transactions on Knowledge and Data Engineering (Vol. 20, Issue 10, pp. 1348–1362). Institute of Electrical and Electronics Engineers (IEEE). https://doi.org/10.1109/tkde.2008.53
[^4]: Hess, S., & Morik, K. (2017). C-SALT: Mining Class-Specific ALTerations in Boolean Matrix Factorization. In Machine Learning and Knowledge Discovery in Databases (pp. 547–563). Springer International Publishing. https://doi.org/10.1007/978-3-319-71249-9_33
Topic 2: Cell Image Extraction (Franziska Lautenschläger, Erbara Gjana)
We image cells for long time periods under fluorescence microscope. On one side, to receive good images/movies, we increase the laser power. On the other side, light kills the cells. It causes phototoxicity on cells and they die. We need cells alive for long time movies. Also, fluorescense bleaches over time due to laster light and we get more and more noizy images. We ask for someone who can deal with noise in the images to extract the cell shape, and shape of fluorescently labeled proteins from those noisy images. In the end we want to generate images/movies with better quality, even in presence of few photons or low fluorescent signal.
Topic 3: Smart City (Michael Schmitz, Brian Willems)
The municipality of Wadgassen has started several initiatives to collect relevant data as a basis for new services. At the beginning of the year around 550 Radiators with intelligent radiator thermostats based on LoRaWAN technology have been installed in administration buildings, schools and day care centers. The thermostats are able to measure and send the current temperature and receive time plans with target temperatures to achieve in the corresponding rooms. Whenever the current temperature measured by a thermostat is below the target temperature, the thermostat would open the heater and start heating. The question here is now, how to optimize the heating cycles in the corresponding buildings in order to minimize energy consumption and provide the desired temperatures at the same time.
Topic 4: Transfer Learning (Frank Mücklich, Martin Müller)
In der Materialwissenschaft stehen für das maschinelle Lernen oft keine großen Datenmengen zur Verfügung, sodass die Technik des Transfer Learning weit verbreitet ist. Größtenteils erfolgt ein pre-training auf der ImageNet Datenbank und anschließend das fine-tuning des CNN anhand der annotierten materialwissenschaftlichen Daten. Trotz des signifikanten Domänengaps zwischen ImageNet und Materialwissenschaft funktioniert dieser Ansatz gut. Allerdings ist nicht klar, ob durch ein pretraining auf materialwissenschaftlichen Daten eine noch bessere Performance möglich wäre oder weniger annotierte Daten für das fine-tuning benötigt würden. Kürzlich hat die NASA einen Datensatz an mikroskopischen Bildern generiert und damit ein pretraining der gängigen CNN-Architekturen durchgeführt („MicroNet“). Diese stehen auf Github zum Download zur Verfügung.
In diesem Projekt werden verschiedene annotierte Datensätze aus der Materialwissenschaft zur Verfügung gestellt. Die Studierenden sollen jeweils ein allgemein anwendbares Python-Skript für eine CNN-Klassifizierung sowie eine CNN-Segmentierung programmieren, um ein CNN mit MicroNet pretraining zu laden und anschließend ein fine-tuning mit den materialwissenschaftlichen Daten durchzuführen. Danach soll die Performance der Klassifizierung bzw. Segmentierung zwischen MicroNet pretraining, ImageNet pretraining und random initialization verglichen werden. Außerdem soll untersucht werden, ob durch MicroNet pretraining auch mit weniger annotierten Daten eine gute Performance erreicht werden kann.
Topic 5: Hybride Klassifizierungsansätze (Frank Mücklich, Martin Müller)
Werden in der Materialwissenschaft Bilddaten generiert, z.B. Mikroskopaufnahmen der Mikrostruktur eines Materials, fallen sog. Metadaten an. Dabei handelt es sich zum Beispiel um Bildaufnahmeparameter des Mikroskops wie die gewählte Vergrößerung oder den verwendeten Kontrastmechanismus. Darüber hinaus stehen weitere Probeninformationen, z.B. die chemische Zusammensetzung des Materials oder Prozessparameter der Herstellung zur Verfügung. Bei der ML-Klassifizierung der Bilddaten werden diese Informationen bisher nicht weiter berücksichtigt. Allerdings ist davon auszugehen, dass diese Metadaten in bestimmten Anwendungsfällen die Performance der Klassifizierung verbessern könnten.
Aufgabe der Studierenden ist es, ein Python-Skript für eine hybride Klassifizierung, die als Input sowohl Bilddaten als auch tabellarische Daten verwendet, zu programmieren. Es wird ein annotierter Beispiel-Datensatz zur Verfügung gestellt. Dieser enthält ca. 950 mikroskopische Aufnahmen von Stahl-Mikrostrukturen. Als Metadaten stehen die Vergrößerung und der Kontrastmechanismus des Rasterelektronenmikroskops sowie die Wärmebehandlungsparameter der Herstellung zur Verfügung. Die Klassifizierungsgenauigkeit der hybriden Klassifizierung soll der Genauigkeit gegenübergestellt werden, wenn nur die Bilddaten zur Klassifizierung verwendet werden.
Topic 6: Object Detection/Instance Segmentation (Frank Mücklich, Martin Müller)
Eine weit verbreitete Aufgabe in der Materialprüfung ist das Bewerten des Bruchverhaltens eines Materials. Dies geschieht noch häufig manuell und subjektiv, allerdings wurde bereits gezeigt, dass Machine Learning die Klassifizierung von Bruchflächen verbessern kann.
Üblicherweise werden mehrere Proben nacheinander gebrochen und dann eine Übersichtsaufnahme aller Bruchflächen gemacht. Erster Schritt einer automatisierten ML-basierten Pipeline und das Ziel dieses Projektes ist demnach das Identifizieren der einzelnen Bruchflächen in dieser Übersichtsaufnahme. Dies soll sowohl mittels eines object detection Ansatzes als auch eines instance segmentation Ansatzes realisiert werden. Dafür stehen bereits annotierte Daten zur Verfügung.
Topic 7: Ensemble and Fusion Methods (Andreas Schütze, Payman Goodarzi)
Our Machine Learning Toolbox currently lacks ensemble or fusion methods based on existing feature extraction, feature selection, and classification/regression methods. The goal of this mini-project is to implement ensembling/classifier fusion methods for the toolbox in the context of predictive maintenance and condition monitoring.
Topic 8: Conformal Predictions for Traffic Sign Recognition (Georg Schneider, Tim Kehl, Patrick Trampert, Fabian Woitschek)
Traffic sign recognition is essential for autonomous driving because it enables the vehicle to interpret and respond to road signs such as speed limits, stop signs, yield signs, and other regulatory signals. Without accurate and reliable classification of these road signs, autonomous vehicles would struggle to navigate roads safely and efficiently, potentially resulting in accidents, traffic congestion, or violation of traffic rules.
AI-based traffic sign recognition can help autonomous vehicles to identify and analyze road signs in real-time, allowing the vehicle to adjust its speed and driving behavior accordingly. It is a critical component of autonomous driving systems that ensures safe and effective navigation on roads.
In our project, students will develop and implement a conformal prediction approach for traffic sign recognition that should provide statistical confidence measures for each classification. This framework should work for any generic class-dependent classification problem.
Topic 9: Network Science and Nonlinear Dynamics (Giovanna Morigi, Frederic Folz)
Networks are ubiquitous in our everyday life: power grids deliver electricity, networks of airports connect distant places around the world and urban transport systems allow people to travel between different metro stations, for example. Hence, understanding the emergence and characteristics is an important task for an efficient use of resources. Detailed analyses have shown that the essential properties of most real-world networks are not captured by the Erdős–Rényi model and extensions thereof. This has lead to the development of alternative models, such as the Barabási-Albert and the Bianconi-Barabási network models, that can exhibit key behaviors such as first-mover advantage, fit-get-rich, winner-takes-all as a function of the fitness properties of the nodes distribution [Barabasi].
This project aims at analyzing the scaling properties of the Bianconi-Barabási networks with the number of nodes for different fitness distributions.
The students will use an algorithm developed in [Bianconi et al.] to generate the networks. They will write the code for the algorithm in C++ or Python. The generated networks will then be characterized as a function of the key parameters of the algorithm, going beyond the analysis of [Bianconi et al] by determining the degree distribution of the networks. In a final stage, important parameters of the algorithm can be made time-varying, simulating the dynamical behavior of several real-world implementations. To visualize the networks, the disparity filter [Serrano et al] will be used which allows to extract the backbone of the networks. The students will write the code for the disparity filter and for the visualization of the networks.
Literature:
- Barabási: http://networksciencebook.com/
- Bianconi et al.: https://iopscience.iop.org/article/10.1209/epl/i2001-00260-6
- Serano et al.: https://www.pnas.org/doi/10.1073/pnas.0808904106
Topic 10: Robustness of Zero-Noise Extrapolation in Quantum Computing Through Neural Networks (Peter P. Orth)
Current quantum computing hardware is inherently noisy, which means that gates have a finite infidelity and errors that occur during the computation cannot be corrected in real-time. The current era is therefore called NISQ Computing, which stands for Noisy Intermediate-Scale Quantum Computing. One possibility to correct for quantum errors, however, is via classical postprocessing, which is referred to as quantum error mitigation. One of the most successful quantum error mitigation methods is zero-noise extrapolation (ZNE) in which the noise is first deliberately increased during the calculation, for example by inserting gates that add up to the identity. The zero-noise limit is then obtained via extrapolation from the results obtained at different noise strengths. One of the open questions, which will be addressed in this project, is how to choose the best extrapolation protocol. One of the challenges is that the exact scaling factor of the noise strength is typically unknown, and the zero-noise extrapolation result depends on the form of the fit function (e.g. polynomial versus exponential). In this project, we will use a neural network for this task and compare to other approaches such as Richardson extrapolation. The central goal of the project is to increase the robustness of ZNE by identifying a network that is trained on small system sizes, yet generalizes well to larger system sizes. The data will be obtained using noisy simulators of quantum computers via the quantum computing toolkit Qiskit.
Topic 11: Reinforcement Learning for Quantum State Preparation (Peter P. Orth)
Reinforcement learning (RL) has witnessed recent applications to a variety of tasks in quantum programming. This project investigates the capability of RL for quantum state preparation, which is a fundamental task in quantum programming. This project builds upon previous work in my group (https://arxiv.org/abs/1912.12002) that has shown how to model this task as a discrete Markov Decision Process (MDP) and used this description for state preparation of a single qubit. The goal of this project is to extend the previous work to a larger number of qubits and to find optimal gate sequences in the presence of realistic quantum noise. One important open question that will be addressed is the choice of reward. The group will implement and compare different RL methods to obtain the optimal policy such as policy iteration, on-policy TD control (SARSA) and off-policy Q-learning.
Topic 12: Reinforcement Learning for Modular Racetrack (Jörg Hoffmann, Timo P. Gros)
In recent years, several papers published by this faculty, e.g. [1-4], have used the Racetrack benchmark. Racetrack is, in a nutshell, a gross simplification of autonomous driving: ignoring the dimensionality and physics of the vehicle, the agent needs to steer the car in a two-dimensional grid towards a goal. In every step, the car can only accelerate or decelerate within a given range, making foresighted planning/decisions necessary.
In the context of reinforcement learning, this gives a very sparse reward environment: the agent receives a positive reward when it reaches the goal, and it gets punished, i.e., receives a negative reward, if it crashes into a wall, but no further intermediate rewards are given. Thus, Racetrack is a complicated task for reinforcement learning, especially if the considered maps are large.
In this project, the group shall implement a modular Racetrack version. Each map shall be composed of any number of pre-composed building blocks, but only a limited number of the pre-composed blocks exist. Further, the group should apply existing reinforcement learning algorithms to the new benchmark. Especially the choice of state representation will be of interest. The modularity of the benchmark should enable training agents on small maps but deploy them on larger ones.
[1] Baier et al.: "Lab Conditions for Research on Explainable Automated Decisions."
[2] Christakes et al.: "Automated Safety Verification of Programs Invoking Neural Networks."
[3] Gros et al.: "Deep statistical model checking."
[4] Hartmanns and Klauck: "The Modest State of Learning, Sampling, and Verifying Strategies."
More topics will be added as they come in.