Topics
Topic 1: Populist imagery and European Elections (Rosa M. Navarrete, Brahmani Nutakki)
One of the most widely accepted definitions of populism characterizes it as a thin ideology, as described by Mudde and Kaltwasser (2017), which views society as divided into two "homogeneous and antagonistic camps": "the people" versus "the corrupt elite." Additionally, populism is often associated with rhetoric centered around the concept of the "general will." Consequently, extensive research has been conducted on populist rhetoric as a political strategy. However, less attention has been paid to populist imagery. In an era where the influence of images and memes on social media is paramount, this project aims to investigate whether populist parties also employ a populist imagery that mirrors their political strategy of appealing to a binary division of society: good versus evil, the pure versus the corrupt, us versus them. Students involved in this project will collect graphic content from the social media accounts of populist parties and politicians, analyzing the elements that define how populist political actors utilize images as a component of their political strategy. The upcoming European Elections provide an ideal context for testing this hypothesis, as political actors will campaign simultaneously across various European countries.
Topic 2: Demonstrations and Online Behaviour (Alex Hartland, Jianlong Zhu)
What effect do demonstrations have on online behaviour? Recent demonstrations against the AfD attracted a lot of participants and media coverage in Germany. Previous research finds that peaceful protest can encourage support for specific causes, while more disruptive or violent protest can be counterproductive. However, the increased salience of far right parties and their issues can also serve to activate latent support for these organisations. Or perhaps such mass mobilisations make little or no difference, and participants would better serve their cause by channelling their energies elsewhere. Did these protests and their relative size increase online engagement with content from the AfD's opponents, or with the AfD and their supporters themselves? Any effects from these activities may also vary along other dimensions, particularly by region and particularly in the German context. Contrasting East-West, Urban-Rural and Border-Inland reactions are therefore likely. We should also consider the flip side to these questions. Do protests in support of the far right drive an opposite and equal reaction? And how does the tone and content of any engagement from either side vary? Answering these questions will bring clarity on many themes which are currently highly relevant in political science research.
Topic 3: Antipluralist discourse in Europe (Giuseppe Carteny, Jianlong Zhu)
Over the past two decades, democratic nations have grappled with the rise of radical figures, particularly those from the radical right-wing populist camp. This surge has raised concerns about the stability and endurance of democratic systems in Europe and beyond, capturing the attention of commentators, scholars, and pundits alike. While much research has delved into the ideological extremism or populist tendencies of these figures, the direct link between these traits and democratic stability remains uncertain. A single issue of radicalism does not necessarily pose a threat to democratic systems, nor does populism invariably undermine democratic processes. So, what exactly constitutes the most worrying characteristic of a political leader or party for a democratic system? According to a novel line of inquiry, the crucial factor lies in the degree of anti-pluralism exhibited by political figures.
Pluralism, as a foundational value system, not only advocates tolerance but also underscores respect for diverse opinions rooted in mutual reciprocity (Sartori, 1997). This entails achieving consensus on the principles of reciprocal tolerance, thereby alleviating conflicts within the democratic framework (Popper, 1945). Grounded in the principle of tolerance, pluralism underscores the democratic tenet that while the majority may wield power, it must do so while respecting minority rights. Consequently, pluralism acknowledges the existence of diverse social groups with varying ideas and interests within societies (Mudde and Kaltwasser, 2013).
By analysing speeches given by heads of government in European democracies and autocratic regimes worldwide, obtained through web-scraping techniques, Maerz and Schneider (2020) have shown that anti-pluralism is a crucial factor in distinguishing between democratic and autocratic figures and systems. Additionally, further research focusing on political parties has revealed a close connection between this trait and democratic backsliding. For instance, Lührmann et al. demonstrate that candidates using anti-pluralist rhetoric before elections are more likely to turn towards autocracy after gaining office. While not all autocrats issue such warnings, many do, emphasising the importance of vigilance in preserving liberal democracy (Lührmann et al., 2021).
Yet, lingering questions persist: are anti-pluralist attitudes and behaviours spreading in public discourse? Does the presence of an anti-pluralist figure spark similar discourses in other domains? How do these spread on social media platforms? And ultimately, how do democratic figures respond to such discourse?
Topic 4: Anti-war movements and Russian propaganda (Giuseppe Carteny, Brahmani Nutakki)
Since time immemorial, warfare has been invariably accompanied by propaganda. Rulers, leaders, and military commanders have consistently utilised persuasive tactics for various objectives: to sway public opinion, rally support, and justify military endeavours before their subjects or citizens. However, propaganda's influence extends beyond shaping internal discourse; a significant objective is also to diminish backing for the enemy's military endeavours. According to analysts and commentators, the Russian-Ukrainian conflict, sparked by Russia's invasion of Ukraine in 2022, is no exception.
Following the setbacks in Russia's initial war plans, such as the failure to swiftly capture Kyiv and control the entire country, the Russian military strategy has included a relentless campaign to undermine public support for the Ukrainian resistance and counter-offensive (Nixey, 2023).
Such tactics are not novel for the Russian regime. For instance, we already know that Russia exploited bot and troll activities on Twitter aimed at manipulating the international audience's opinion on Russian opposition leader Alexei Navalny and broader Russian-Western relations (Alieva et al., 2022). Examining the discourse in the Russian social media sphere concerning the Ukrainian conflict reveals active manipulation by supporters of the Russian invasion (Alieva & Carley, 2022). However, there is a lack of systematic evidence regarding how these dynamics have influenced discourse beyond Russian social media, particularly in countries supporting Ukraine's resistance.
This mini-project aims to address the following research questions: How has Russian "anti-war" propaganda disseminated in European and US public discourse? Which political actors, media outlets, movements, or influencers endorse the Kremlin's strategy? How can we distinguish between pro-Russian actors and genuine advocates for a peaceful solution of the conflict?
Topic 5: Analyzing Party Competition on Gender-Related Issues (Giuseppe Carteny, Vikram Kamath Cannanure)
Over the last decade, gender issues have become a highly contentious and polarised topic in European politics. Scholarly attention to these issues has steadily increased, transforming a largely neglected topic in European politics into a productive site of political research. In contrast to the burgeoning literature on gender and politics for individuals, analyses of party competition on such issues are still lagging behind. In particular, most of the existing work has been case studies, and comparative data for measuring party positions across geographical contexts and over time have been lacking so far. With the UNTWIST EU Horizon project, we collected a multilingual set of election manifestos from six European countries (Denmark, Germany, Hungary, Spain, Switzerland and the United Kingdom) published between 2004 and 2021. These manifestos have been manually annotated according to a novel coding scheme and procedure developed to measure the salience and positioning of parties on gender-related goals, issues and policies. The aim of this project is to fine-tune a multilingual BERT (mBERT) to extend the dataset beyond the national contexts for which annotated data is available. We then replicate our exercise by relying on multilingual applications of supervised machine learning (SML) methods. Substantively, this project would represent one of the first attempts to comparatively analyse political parties gender-politics. Methodologically, it would provide new insights into multilingual text analysis.
Topic 6: Quantum computing (Peter P. Orth, Dr. Prachi Sharma)
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 and gate compilation, which are fundamental tasks in quantum programming. This project builds upon previous work in my group (npj Quantum Inf. 9, 108 (2023)) 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 qudits (i.e. multilevel quantum systems) and to a larger number of qubits. 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 7: AI and Quantum Computing (Peter P. Orth, Dr. Prachi Sharma)
Current quantum computing hardware is inherently noisy, which means that gate operations have a nonzero 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. In this project, you will employ a variety of state-of-the-art ML methods such as linear regression, multi-layer perceptrons, random forests, and graph neural networks to predict the ideal outcome of the quantum calculation. The key idea is to learn the relation between the noisy data and the ideal output on smaller instances and classically simulatable so-called Clifford circuits.
Topic 8: Network dynamics in the presence of noise (Giovanna Morigi, Frederic Folz and Sayan Roy)
This project aims at analyzing network dynamics inspired by the food search of the slime mold Physarum polycephalum - a primitive organism that has demonstrated its ability to solve shortest path problems and to design efficient transport systems. For this purpose, the students will use an algorithm, developed in [Bonifaci et al], to analyze the dynamics of transport networks serving two demands on a grid of 30 x 30 nodes. They can use an existing implementation of the algorithm in Matlab. The network dynamics is mainly governed by a nonlinear activation function, similar to the use of activation functions in machine learning architectures, e.g., sigmoid or tanh. There, it is known that their saturation behavior can limit the training efficiency, which can be overcome by introducing noisy variants of the activation functions [Gulcehre et al]. Inspired by that, the students will modify the provided algorithm for network design by introducing noisy activation functions. They will use the modified algorithm to generate transport networks and characterize them as a function of the key parameters, going beyond the analysis of [Bonifaci et al]. The topology of the resulting networks will be determined by an extension of the disparity filter [Serrano et al] to the stochastic dynamics [Folz et al]. The characterization will be implemented by applying measures for their robustness, transport efficiency, and costs. Further, the students will analyze the impact of the noisy activation function on the convergence speed of the algorithm.
Topic 9: Semi-supervised Learning zur Mikrostrukturauswertung von metallischen Werkstoffen (Frank Mücklich, Martin Müller)
Bei materialwissenschaftlichen Anwendungen ist das Labeln der Daten für das maschinelle Lernen, so wie auch in anderen Disziplinen, oft mit großem Aufwand verbunden und stellt dementsprechend den Bottleneck einer ML-Implementierung dar. Dementsprechend stehen oft nur wenig gelabelte Daten, aber viele ungelabelte Daten zur Verfügung. Hier kommt die Methode des semi-supervised learning ins Spiel. Dabei handelt es sich um eine Art des maschinellen Lernens, bei dem ein Modell sowohl mit gelabelten als auch mit ungelabelten Daten trainiert wird, indem Informationen aus den gelabelten auf die ungelabelten Daten ausgedehnt werden. Dabei kann durch das Verwenden der ungelabelten Daten zum Teil eine Verbesserung der Modell-Performance gegenüber dem regulären supervised learning festgestellt werden.
In diesem Projekt soll ein Workflow für das semi-supervised learning am Beispiel von Mikroskopaufnahmen erstellt werden. Dazu werden vollständig annotierte Datensätze für eine Klassifizierung sowie eine semantische Segmentierung zur Verfügung gestellt. Die Techniken der Label Propagation sowie des Self-Trainings sollen zum Einsatz kommen. Im nächsten Schritt soll das semi-supervised learning mit einem active learning Ansatz kombiniert werden. Beim active learning kann der Lernalgorithmus den menschlichen Benutzer interaktiv auffordern, neue Datenpunkte zu labeln und dem Modell hinzuzugfügen. Der Workflow soll dann dahingehend erweitert werden, mittels “uncertainty sampling” die Datenpunkte auszuwählen, bei denen das aktuelle Modell am wenigsten sicher ist, was die korrekte Ausgabe sein sollte, und diese anschließend dem Modell als neue Trainingsdaten hinzufügen zu können.
Topic 10: Vision Transformers zur Auswertung von Mikroskopaufnahmen (Frank Mücklich, Martin Müller)
Inspiriert von der erfolgreichen Anwendung von Transformer-Modellen in der natürlichen Sprachverarbeitung wurden ähnliche Architekturen auch für die Verarbeitung von Bildern entwickelt, die sog. Vision Transformer (ViT). In der Materialwissenschaft kommen zur Auswertung der Mikroskop-Aufnahmen der Mikrostruktur eines Materials ML-Methoden bereits seit längerem zum Einsatz. Üblicherweise werden dafür Ansätze verwendet, die sich in der Informatik oder anderen Disziplinen bewährt haben und als etablierte Technologien betrachtet werden können. Das sind Convolutional Neural Networks (CNN) für die Klassifizierung von Bildern sowie die U-Net Architektur mit den typischen CNNs als Backbone für die Segmentierung von Bildern.
In diesem Projekt soll nun das Potenzial der ViT für materialwissenschaftliche Fragestellungen untersucht werden. Dazu werden vollständig annotierte Datensätze für die Bild-Klassifizierung sowie Segmentierung zur Verfügung gestellt. Zunächst soll eine Baseline Performance mit CNN-Ansätzen bestimmt werden. Diese wird dann mit der Performance der ViT verglichen. Durch ein Sub-Sampling der annotierten Trainingsdaten soll die Performance von CNN und ViT außerdem in Abhängigkeit der Trainingsdatenmenge untersucht werden. Am Ende soll eine Handlungsempfehlung zur Verwendung der jeweiligen Architekturen, in Bezug auf Performance, Trainingsdatenmenge und Aufwand der Implementierung stehen.
Topic 11: Machine learning approaches in the prediction of psychometric properties based on eye-tracking data (Vera Demberg, Margarita Ryzhova, Sarubi Thillainathan)
A large body of psychological research on reading shows that the way how individuals process texts is affected by their psychometric profile. Characteristics like working memory capacity and reading fluency influence the amount of time a reader needs to process particular information and the order in which information is processed. In turn, it has been demonstrated that certain eye movement measures could be used to identify the psychometric characteristics of individuals, e.g., to infer their level of reading fluency. Various methodologies have been explored in this task, using eye-tracking measures of different granularity in data analysis, encompassing tasks such as sentence or paragraph reading, and employing different analysis techniques. The choice of approach can significantly influence the accuracy and the amount of eye-tracking data necessary for successful prediction.
In this project, the objective is to develop a workflow for predicting readers’ psychometric characteristics based on their eye movements collected during a paragraph reading task. Given that eye-tracking yields numerous measures, the task also involves determining which measures are predictive of various psychometric characteristics. The provided dataset consists of eye movement measures of 41 individuals, each reading 8 texts in total. Additionally, we provide a dataset of readers’ psychometric characteristics that includes measures of working memory capacity and reading fluency. While building upon the methods previously employed for this task in the field, the aim is to explore a variety of machine learning techniques to achieve high prediction accuracy. Therefore, the goal is to identify the most effective approach.
Topic 12: Reliable Generation of Discharge Summaries (Vera Demberg, Mayank Jobanputra and Marian Marchal)
Detailed and comprehensible discharge summaries are crucial for a patient's well-being, but creating them demands a significant amount of time and effort from clinicians. To optimize hospital workflows and minimize administrative burden on clinicians, this project aims to automatically generate hospital course summaries and discharge instructions from electronic health records. Students will be given a dataset based on MIMIC-IV which includes data from >100K visits to the Emergency Department (ED), including chief complaints and diagnosis codes, at least one radiology report, as well as gold discharge summaries, containing a brief hospital course and discharge instructions. The goal is to generate these summaries, by building a highly reliable system that preserves patients' privacy. The students will obtain hands-on experience with using and fine-tuning state-of-the-art generative AI models and dealing with ongoing challenges in the field (reliability, privacy).
The task is based on an existing shared task – "Discharge me" from the BioNLP workshop (co-located with ACL 2024). More information can be found here https://stanford-aimi.github.io/discharge-me/ (Participation in the Shared Task is not anticipated).
Topic 13: Microscopy image analysis I (Franziska Lautenschläger, Galia Magela Montalvo Bereau)
In fluorescent images of cells we need to decide what is the front and the back of the cells. The images are cells stained with 4 different colors. 2 colors will define the front and back, with this, we can define 2 compartments, and the fluorescence intensity of the other 2 channels in each compartment is what we need. So far, I manually defined the front and the back, and determined the mean fluorescence intensity on each compartment. This result is particularly relevant to support a scientific result that is on the publication process, however the manual determination is a weakness, and the result will be more robust if an automatization process can be developed.
Topic 14: Use of AI to optimize the image segmentation of migrating cells (Franziska Lautenschläger, Lukas Schuster)
To study the dynamics of migrating cells, it is important to extract morphological properties and (eccentricity) of the cells. This can be accomplished by fluorescently labelling the cell shape during migration in a live cell tracking experiment. The automatised tool should help to process the time lapse movies and segment the cell shape based on the fluorescent signal.
Topic 15: Microscopy image analysis III (Franziska Lautenschläger, Enrique Colina)
Differentiating different types of fibers from images and the length. In our experiments, when treated with certain drugs, microtubules in the mitotic spindle are affected, thus compromising its structure. Sometimes the deformation is evident, however subtle changes in specific areas of the spindle can occur that are not easy to identify. We are interested in generating a tool that could compare normal spindles against our samples and detect and quantify differences in length and shape of the fibers forming these structures.
Topic 16: Microscopy image analysis IV (Franziska Lautenschläger, Erbara Gjana)
Evaluate the uptake of particles (number, size, deformation) by macrophages. In our experiments we have macrophages engulfing particles in µm range in diameter. The particles are different sizes and sometimes are deformed by the cell. We are interested in measuring the number, size (diameter, surface, volume) and possibly deformation of each particle inside the cells. I currently get this information manually from the images that I possess. As this process takes time, we would be interested in automizing the extracting of this information from the images. Currently we have analyzed images and we plan to have more in the future.
Topic 17: Estimating revisit times of satellite constellations in space (Ingmar Weber, Till Koebe)
Satellite images are an indispensable data source for monitoring environmental processes and societal change. However, the monitoring frequency of a certain place large depends on how often a satellite constellation passes over it - the revisit time. While many providers boast with high revisit times, there is no bundled information about revisit times of the different constellations out there. However, a number of satellite trackers exist. Can we estimate revisit times those satellite tracking data for any given location on earth?
Topic 18: The strength of weak ties: How does the physical distance to online friendships affect wellbeing? (Ingmar Weber, Till Koebe)Friendships, whether online or offline, are important for social learning and mental health. But does it make a difference whether your online friends are 5 or 5000km away? We exploit Meta's Social Connectedness Index and investigate for the US whether the level of aggregation affects the effect size of being socially connected on sexual health behaviour.
Topic 19: Developing a Tailored Recommender System for Elderly Homeowners (Patrick Schneider, Vikram Kamath)
In Germany, the worsening housing crisis, characterized by rapidly increasing rents, impacts elderly citizens significantly, jeopardizing their financial well-being and ability to find affordable housing. This initiative aims to mitigate the issue by developing a unique recommendation system specifically designed for elderly homeowners considering renting out their properties. The project, structured as a course assignment, will utilize publicly accessible immo_data to create this recommender system with a focus on the needs of the senior demographic. It will begin with a preliminary analysis phase, during which students will examine various rent trends within the dataset. Subsequently, they will explore and implement diverse algorithms suitable for crafting a recommendation engine tailored to this demographic.