Seminar Assignment Winter 2024/2025
The central registration for all computer science seminars will open on September 11th.
This system is used to distribute students among the available seminars offered by the CS department. To register for any of the seminars, you have to register here until October 16th, 23:59 CET. You can select which seminar you would like to take, and will then be automatically assigned to one of them on October 18th.
Please note the following:
The assignment will be performed by a constraint solver on October 18th, 2024. You will be added to the respective seminars automatically and be notified about this shortly thereafter. Please note that the assignment cannot be optimal for all students if you drop the assigned seminar, i.e., make only serious choices to avoid penalty to others.
This seminar (AI4AD) is concerned with advances in autonomous driving research from the perspective of AI. In particular, we will take a closer look at techniques and systems of AI for autonomous driving with focus on predicting the behavior of pedestrians and vehicles as well as collision-free navigation of self-driving cars in various traffic scenarios, and discuss their strengths and weaknesses. The selected approaches are based on deep learning, neuro-explicit and large language models.
The seminar type is classic in the sense that registered participants will present and discuss assigned topics. In addition, there will be two dedicated opponents for each presentation of an assigned topic.
The seminar takes place on Wednesdays, 4:15pm - 6pm, at DFKI Saarbrücken (SIC Bldg. D3.2), in room "Turing II" (DFKI NB +2.31). Please ask at the DFKI reception for direction to room "Turing II".
First seminar session (introduction and topic assignments) is on Wednesday 23.10.2024 at 4:15pm
For more information, please visit the seminar website: https://www.dfki.de/~klusch/AI4AD-seminar-ws24
Requirements: This seminar aims primarily at advanced bachelor and master students in Computer Science and DSAI. Solid knowledge in AI (ideally, taken introductory courses on AI, ML, genAI, or sufficient knowledge on these areas from other sources) is required. Participants should be very interested in the domain of autonomous driving research and development. Selected background papers for the seminar are referenced on the seminar website (topics).
Places: 10
Learning-augmented algorithms is a recent line of research that seeks to blend the strengths of machine learning with classical algorithms, aiming for both practical efficiency and robustness. In this framework, we are provided with predictions about missing data—such as future data in online algorithms or private data in truthful mechanisms—though the accuracy of these predictions is uncertain. The objective is to improve the performance guarantees when predictions are accurate and to incur only a constant loss when predictions are poor.
Some initial lectures will be taken by the instructors to explain the basics that will help students to select their paper/topic. The seminar is open for all interested students and postdocs. Students aiming to get credit points must give a regular presentation (of the chosen research paper) and write a short summary about it. The presentation needs to be discussed with us at least one week before your scheduled talk.
See the seminar website for more information (TBA) or contact us (nrathi@mpi-inf.mpg.de, gshahkar@mpi-inf.mpg.de) in case there are any questions!
Requirements: This is a theoretical seminar that will require mathematical maturity (in particular, the ability to understand and write formal mathematical proofs) and a good background in algorithms. A proper preparation of your talk will require non-trivial effort. The target audience of this seminar are master students, PhD students, as well as postdocs.
Places: 10
A multi-site seminar on "how to make computer systems sustainable".
As the world struggles with climate change, the computing sector's role as a source of greenhouse emissions (and other bad environmental impact) is steadily increasing. This seminar, co-located and co-hosted with the BOSS Team at Ruhr-Universität Bochum (RUB), explores the intersection of computer technology and sustainability - identifying the key challenges of and innovative solutions for a sustainable future with computers.
We begin by looking at the origins of carbon emissions and learning to understand where emissions are caused. We then focus on known and new methods for measuring and modeling emissions in computer systems. Finally, we show ways in which systems and software can be adapted to become more sustainable.
More details here: https://dcms.cs.uni-saarland.de/ascs_2425/
Requirements: - Open for advanced Bachelor and any Master students.
- Ideally, you have taken courses from the "systems & software" domain: e.g. Data Networks, Operating Systems, Distributed Systems, Software Engineering.
Places: 8
In this seminar, we will take a look at how psycholinguistic theories can be formalized as computational models and how these models can help specify and improve those theories and generate predictions. The first two sessions will give participants an overview of the modeling frameworks – the cognitive modeling framework Adaptive Control of Thought – Rational (ACT-R) and Bayesian probabilistic models, including the Rational Speech Act (RSA) framework. Each participant will present a paper on a seminar-relevant topic, such as modeling the role of working memory and processing speed in language comprehension and production, eye movements and sentence comprehension, listener adaptation and rational overspecification.
The goal of the seminar is to take stock of the range of approaches used in computational cognitive modeling and the recent applications of cognitive modeling to psycholinguistic and pragmatic phenomena. At the end of the course, participants may choose between a more traditional term paper and a project modeling a psycholinguistic phenomenon of their choice in one of the frameworks discussed in the seminar, after consultation with the instructors.
Places: 12
As mathematical research advances, researchers become more and more specialized, and the mathematics they produce becomes more and more complicated to verify.
The possibility to formalize and check proofs thanks to computer programs is thus more relevant than ever. What’s more, tremendous progress in recent years make it so that formalizing actual research level mathematics is possible, and formalizing student level mathematics is accessible to students.
In this seminar, students will practice with the LEAN proof assistant (https://leanprover-community.github.io/).
We meet weekly on Zoom, and discuss informally: each student gets a chance to speak, to explain the work they have done in previous weeks, and to plan ahead.
See the course's page to obtain the zoom link.
https://www.math.uni-sb.de/ag/bartholdi/cap/
IMPORTANT: If you have taken the seminar before, we are very happy to have you participate again, giving you advanced topics and further material depending on your experience.
Requirements: Students may obtain up to 8 credit points, by formalizing a theorem or new definitions, thus contributing to mathlib, the library which gathers all mathematics that has already been formalized in LEAN.
Topics suitable to both MSc and BSc students will be offered, no prior knowledge is required.
Places: 10
From finding a mate, to booking a holiday, our lives are increasingly mediated by online platforms. Digital traces left by these interactions provide opportunities to study societal phenomena while creating challenges around the responsible use of data. In this seminar, students will learn how computational methods and machine learning can be applied to study society through such data.
The first part of the seminar will familiarize students with existing work in computational social science. Each week, we will focus on a different topic such as “Digital Democracy” or “Gender Gaps” and explore methods to quantify these phenomena. The second part of the seminar will be about projects in which students are asked to quantify a societal phenomenon of their choice using computational methods. Here, students can both propose topics or choose from topics defined by the lecturers.
The overall course performance will be based on (i) overall course participation, (ii) assigned paper presentations, (iii) a project pitch and (iv) the written project report.
Apart from learning about interdisciplinary research and applications of machine learning, students will also learn research skills such as how to read and discuss papers, how to plan a project, how to present their work, how to write a scientific paper, and how to work in teams.
Students can take this course as a seminar.
Future details at https://cms.sic.saarland/das_2425/. Timing is Fridays, 10am-12am.
Requirements: Msc students only – the project-based element of the seminar will require some Python programming and data analysis experience.
An interest beyond the foundations of CS, and caring about societal problems is a must.
Places: 12
We study state-of-the-art articles on algorithmic decision procedures for specific theories such as arithmetic, bit vectors, or theories defined by first-order fragments.
Requirements: Successful participation in the lectures Automated Reasoning and Algorithmic Quantifier Elimination.
Places: 5
This seminar course delves into the crucial and evolving field of explainability in machine learning (ML). As ML models become increasingly complex and integral to various domains, understanding how these models make decisions is essential. This course will explore different methodologies for interpreting ML models, including rule-based, attribution-based, example-based, prototype-based, hidden semantics-based, and counterfactual-based approaches. Through a combination of paper readings, discussions, and presentations, students will gain a comprehensive understanding of the challenges and advancements in making ML models transparent and interpretable.
More details will be released at https://dcms.cs.uni-saarland.de/explainingml_2425/
Requirements: The student should take a course in machine learning or have sufficient knowledge from other courses; The student should speak English and understand that the seminar will be conducted entirely in English
Places: 8
Graphs have long proven to be a powerful data representation across a wide range of applications. Prominent examples include social and transportation networks, as well as small molecules and proteins. Recently, graph neural networks (GNNs) have emerged as a powerful tool for extending the success of deep learning to the graph domain.
In this seminar we aim to cover both the foundations of GNNs as well as more advanced topics such as their limitations and expressiveness, relation to transformers, extensions to geometric graphs and graph generative models.
During the seminar, participants will create talktorials, i.e., self-contained IPython notebooks that explain (teach) a select topic both from a theoretic point of view and in terms of a practical demonstration. The mandatory part of the seminar will conclude with final talks where participants present their talktorials to their fellow students.
Further details will be published on the course website: TBD.
Requirements: Succesful participation in a machine/deep learning course, preferably the machine learning core lecture or "Neural Networks: Theory and Implementation".
Practical deep learning experience, preferably with the PyTorch framework.
Places: 9
Spectre, Meltdown, and other microarchitectural attacks have been in the limelight in recent years. These attacks exploit subtle timing and behavioral differences of processors that are caused by microarchitectural optimizations such as caches and speculative execution to gain access to secret information.
The vulnerabilities exploited by microarchitectural attacks are not captured by today's hardware-software contracts, i.e. instruction-set architectures (ISAs). Traditionally, ISAs only capture the "functional" behavior of a system and thus have a blind spot when it comes to side channels. Recently, there has been a push to augment conventional ISAs with a formal specification of information leakage, resulting in more general hardware-software contracts. Such contracts enable writing secure code, e.g. implementing cryptographic algorithms, in a rigorous manner.
In this course we will study recent advances to prove the security of hardware designs or find vulnerabilities in them. We will focus on two main techniques: formal verification to and fuzzing.
For more details get in touch with Jan Reineke (reineke@cs.uni-saarland.de) or consult the seminar page: https://cms.sic.saarland/securehardware24
Requirements: Basic knowledge of computer architecture (e.g. due to System Architecture) is required.
Knowledge of security and formal methods is a plus, but not required.
Places: 8
While the Internet has started as a research effort, it has consistently evolved throughout the decades into the largest commercial network. Today, many research efforts focus on understanding the Internet's structural trends, optimizing its packet delivery, and laying the foundations for its future developments. In this seminar, you will receive a closer look at bleeding-edge research published at the top conferences in our domain. Discussing state-of-the-art approaches and recent findings from a broad range of network-related topics with your peers and instructors will provide you with a deep understanding of your assigned topic. Preparing the accompanying research survey and topic presentation will not only strengthen your academic reading and writing skills but also help you to present your future work in a more accessible and structured way.
Requirements: Participants should have successfully participated in "Data Networks" or an equivalent course.
Places: 12
In algorithmic research, many works focus on obtaining faster, simpler, or otherwise improved algorithms. But how do we know when to stop searching for better algorithms, how do we know that we have exhausted all possibilities for improvement and reached the optimum?
In this seminar, we will take a look into the wonderful world of impossibility results, i.e., results that tell us that obtaining a certain algorithmic objective is provably impossible. While, in general, impossibility results come in many different flavors, the impossibility results that we will study are, to a large extent, runtime lower bounds that rule out the existence of algorithms (for a fixed problem or problem class) that are faster than a certain runtime threshold. Moreover, our focus will be on impossibility results for distributed (and related) algorithms for graph problems. More specifically, the computational model that we will most frequently encounter in this seminar will be the LOCAL model of distributed computation. Note that this is a theory seminar, i.e., we will focus on lower bound proofs, algorithm analysis, and the like.
In this seminar, each student has to present one or two assigned paper(s). Each presentation is followed by a discussion lead by the presenter. Besides presentation and participation in the discussion, the grade will depend on written deliverables. More information will be available on the seminar webpage: https://cms.cispa.saarland/irla2425/
In particular, please check there whether the weekly slot fixed for the seminar works for you. It might be possible to change the dates later if everyone involved agrees, but we cannot guarantee that.
The required language for the presentations is English.
Requirements: There are no formal prerequisites for this seminar, but a general interest in graph theory and designing/analyzing algorithms as well as a basic understanding of probabilities and algorithmic analysis (e.g., O-notation) will be helpful.
Places: 10
In this seminar, we will discuss information theoretic approaches to machine learning. We will investigate, among others, the following questions:
What is interesting and meaningful structure? How can we identify this from data without overfitting? What is a good model when we don't have a decent prior, target, or even know what we're looking for? What is the ultimate model, and how can we approximate it in practice? We'll explore these in light of Algorithmic Information Theory, and its practical variant, the Minimum Description Length (MDL) principle. We will consider the relevance and application of these to a wide range of problems, from description and prediction to genearlization, from neural to symbolic, from associative to causal, and so on.
We will generally meet once a week. The first part of the course will feature regular lectures covering the basic topics of the course, and discussion sessions in which we will discuss scientific articles in light of the lectures. During the second part students will have to write an essay based on scientific articles assigned to them by the lecturer and give a presentation. The presentations will be scheduled in a block-seminar style, that is in one or two days near the end of the semester, the exact date is to be announced.
Requirements: Students should have a strong mathematical background and a basic working knowledge of machine learning, data analysis and/or statistics, e.g. by successfully having taken courses related to machine learning, data mining, and/or statistics, such as Topics in Algorithmic Data Analysis, Machine Learning, Elements of Machine Learning, Optimization for Machine Learning, Probabilistic Graphical Models, Information Retrieval and Data Mining, etc.
As this seminar works best with a small and active group, I will give preference to highly motivated students. Hence, please write a short (1 or 2 paragraphs) motivation for why you want to take part.
Places: 10
Inpainting has been introduced as a technique to restore missing or deteriorated regions by using information from within the image. There are various ways to fill these missing parts such as simple interpolation, diffusion-based methods, or by comparing image patches. Nowadays, specialised inpainting methods can remove objects seamlessly or can even reconstruct an image using only a tiny fraction of the original information. In this seminar, we cover both the foundations of the field as well as forefront research in the field of inpainting.
For more information: https://www.mia.uni-saarland.de/Teaching/ifa24.shtml
Requirements: The seminar is for advanced bachelor or master students in Visual Computing, Mathematics, or Computer Science. Basic mathematical knowledge (e.g. Mathematik für Informatiker I-III), some knowledge in image processing and computer vision as well as basic knowledge in neural networks is required.
Places: 12
It's time to connect to your inner kid in this fun and engaging seminar format (https://colinraffel.com/blog/role-playing-seminar.md.html): Several students read the same paper, but each student takes on a specific role, which defines the lens through which they contribute to the discussion. Students cycle through roles throughout the course of this seminar on overparameterization in deep learning.
Content:
Deep learning continues to impress us with breakthroughs across disciplines and is a major driving force behind a multitude of industry innovations like ChatGPT. Most of its successes are achieved by increasingly large neural networks that are trained on massive data sets and still achieve zero training loss. This recent trend to overparameterize neural networks defies classic concepts of statistical learning theory that suggest to avoid overfitting by reducing the number of trainable parameters. We will look into recent explanations of this puzzling phenomenon, discuss related insights, and challenge the modern belief that scaling up neural networks is always the best way to move forward. Are the simplest models always the best choice? And is counting parameters really the best way to measure model complexity? Please join the seminar if you enjoy thinking about this kind of questions.
More information can be found here:
https://cms.cispa.saarland/overparam_24/
Requirements: The course has no formal requirements but preference will be given to Master students in Computer Science and related fields with prior knowledge in deep learning and a convincing motivation.
Places: 14
The rise of artificial intelligence (AI) is transforming everyday lives, including education, and it requires us a deep understanding of and research insights into its applications and implications in the field. This seminar aims to equip students with the knowledge and skills needed to critically analyze AI in education and contribute to this evolving field. The seminar is jointly taught by Prof. Tomohiro Nagashima in the CS department (https://tomonag.org/) and Dr. Sarah Malone in the Education Science department (https://www.uni-saarland.de/lehrstuhl/bruenken/personen/dr-sarah-malone.html), and it targets students in both departments.
During the seminar, students will collaboratively design and develop a "Living AI-education Dashboard," a dynamic resource that summarizes and visualizes current research, trends, and data on AI in education. Through project-based learning, students will gain hands-on experience in data visualization, dashboard creation, dashboard design, and research methods (e.g., how to conduct systematic literature review). They will also develop interdisciplinary thinking by integrating concepts from both computer science and education science and through collaborations across the domains. The course is taught by an interdisciplinary team that encourages collaboration between departments and prepares students to tackle complex, real-world problems.
The seminar will be offered on Mondays 10-12:00 (location TBD). Feel free to contact nagashima@cs.uni-saarland.de for any questions.
Requirements: Students in the Computer Science degree programs are expected to have past experiences and skills in data visualization
Places: 10
Deep learning is the predominant machine learning paradigm in natural language processing (NLP). This approach not only gave huge performance improvements across a large variety of natural language processing tasks. In this seminar the topics will be based on recent papers from ICLR/ICML/Neurips/ACL/EMNLP or similar.
The seminar will be a block seminar in the spring break 2025
For more information see: https://www.lsv.uni-saarland.de/block-seminar-machine-learning-for-natural-language-processing-spring-2025/
Places: 8
Monte Carlo ray tracing is a popular technique to render realistic images. It is used for movies, architecture, video games, and product design. This seminar looks at a broad range of methods to make rendering via Monte Carlo ray tracing more efficient. We will look at rendering algorithms, from a basic path tracer to photon mapping, at sampling and material models, Markov chains and path guiding, and more. Each student will implement the assigned paper in a simplified setting, present the paper, and write a short summary of it.
Additional information can be found at https://graphics.cg.uni-saarland.de/courses/ray-2024/index.html
Feel free to contact hua@cg.uni-saarland.de for any questions.
Requirements: The student should have passed either Computer Graphics or Realistic Image Synthesis. The student should care about rendering.
Places: 10
Symbolic AI relies on explicit, human-readable symbols and rules to represent knowledge and solve problems. This approach contrasts with modern, data-driven methods like machine learning, which rely on statistical patterns rather than explicit rules. Recently, both approaches have been combined, leading to the development of neuro-symbolic methods.
For instance, heuristic search, as a very successful part of symbolic AI, must be applied individually to each instance of a domain/problem. In contrast, neuro-symbolic general policy learning focuses on neural network-based policies that can solve all instances within a specific domain, combining symbolic and data-driven AI benefits.
In this seminar, titled Neuro-symbolic General Policy Learning, we will explore the emerging trend of learning neuro-symbolic general policies where the underlying architecture is based on Graph Neural Networks or Transformers.
The seminar will cover reading and preparing scientific publications, giving a talk, writing a summary, and reviewing other write-ups.
Requirements: This seminar will build upon several lectures/concepts:
- Machine Learning
- Neural Networks
- Graph Neural Networks/Transformers
- AI Planning
While we do not expect you to have knowledge about all these areas, a solid background in some of them will help you to understand the scientific papers that we will discuss in the seminar. Please state in your motivation what prerequisites you bring in order to participate.
Places: 11
Machine learning has witnessed tremendous progress during the past decade, and data is the key to such success. However, in many cases, machine learning models are trained on sensitive data, e.g., biomedical records, and such data can be leaked from trained machine learning models. In this seminar, we will cover the newest research papers in this direction.
Requirements: Students are required to have basic knowledge of machine learning.
Places: 15
In this seminar, we take a hands-on approach to explore the workings of internet components and provide a deep understanding of the underlying protocols. Specifically, we examine how routers and switches can be configured and connected to a working IP network.
The seminar consists of a series of activities starting with setting up the hardware and introducing basic configuration for VLANs and IPv4/IPv6 networks and gradually moving to more advanced topics, including but not limited to, routing protocols, virtual private networks, security, network monitoring, and troubleshooting.
Requirements: Participants should have sufficient knowledge of computer networks and how the Internet works, e.g., by having successfully participated in "Data Networks", an equivalent course, or by demonstrating relevant prior practical experience.
Places: 12
Das Seminar „Legal Tech und eJustice” ist ein interdisziplinäres Seminar für Informatiker und Juristen.
Die Veranstaltung wird dieses Jahr hybrid angeboten.
Dozenten: Prof. Dr. Christoph Sorge, Dr. Stephanie Vogelgesang, Dr. Jochen Krüger
Zeit/Ort: Blockveranstaltung (Hybrid)
Weitere Informationen werden auf der Homepage des Lehrstuhls sowie in Moodle veröffentlicht.
Ablauf des Seminars
Studierende erhalten während der Vorbesprechung ein Thema und müssen eine Seminararbeit hierzu anfertigen. Innerhalb des gewählten Themas sind die Studierenden in der Schwerpunktsetzung grundsätzlich frei.
Vor der Abgabe der fertigen Seminararbeit ist die Einreichung eines Abstracts nebst einer vorläufigen Gliederung vorgeschrieben. Die Besprechung der Abstracts findet verpflichtend statt, um inhaltlichen Missverständnissen entgegenzuwirken und den Studierenden bereits ein Feedback durch die Betreuer zu ermöglichen.
Danach ist die Abgabe eines „Preprints“ der Seminararbeit erforderlich. Daraufhin findet ein Peer Review der abgegebenen Preprints durch die weiteren Teilnehmer des Seminars statt. Jedem Studierenden werden hierfür 3-4 Paper anderer Studierender zugeteilt, welche von ihm/ihr begutachtet werden. Schließlich finden – in der Regel kurz vor Ende der Vorlesungszeit – die Vorträge statt. Abschließend werden die fertigen Seminararbeiten abgegeben.
Allgemeines: Es gilt der Grundsatz „Tiefe vor Breite“. In der Regel werden breit formulierte Seminarthemen gestellt. Teil Ihrer Aufgabe ist es, selbständig die Literatur zu sichten, sich einen Überblick über das Thema zu verschaffen und sodann einen Schwerpunkt (oder ggf. auch zwei, falls es sich im Einzelfall anbietet) zu setzen. Sie sollen zeigen, dass Sie sich in ein Forschungsthema aus Ihrem Fachgebiet einarbeiten und Sachverhalte aus diesem Forschungsthema vollständig durchdringen können. Das zwingt Sie, viele ebenfalls interessante, angrenzende Aspekte wegzulassen. Ein wesentlicher Teil der Schwierigkeit der Aufgabe ist die Entscheidung, was Sie weglassen können (oder müssen) und welche Aspekte für das Verständnis Ihres Themas wesentlich sind.
In der Einleitung geben Sie eine Motivation und ordnen ihr Thema ein; ggf. bietet sich ein weiterer Abschnitt zu verwandten Arbeiten an, um den Stand der Forschung darzustellen. Dann sollten Sie aber in die Tiefe des von Ihnen gewählten Schwerpunkts gehen. Den Abschluss bildet ein kurzes Fazit.
Abstract: Einarbeitung in das Themengebiet, Literaturrecherche und Studium erster wissenschaftlicher Veröffentlichungen. Auf dieser Grundlage Konkretisierung des gewählten Themas bzw. Wahl des Schwerpunkts innerhalb des Themas sowie die Entscheidung über einen sinnvollen Aufbau. Das Abstract ist eine kurze und aussagekräftige Beschreibung des konkretisierten Themas auf mindestens einer Seite (aber nicht mehr als 1,5 Seiten). Eine erste Gliederung sollte mit dem Abstract eingereicht werden.
Preprint: „Vorversion“ der Seminararbeit, wobei der aktuelle Stand ersichtlich sein muss. Hierbei muss es sich um keine bereits vollumfänglich fertige Arbeit handeln, jedoch muss der Inhalt und das Ziel der Arbeit sowie die Auseinandersetzung mit der einschlägigen Literatur erkennbar sein. Literaturverzeichnis, Gliederung und Schwerpunkt sind notwendige Bestandteile.
(Peer) Review: Feedback, welches sich inhaltlich auf das gesamte Paper erstrecken muss (Quellen, Aufbau/Gliederung, inhaltliche Nachvollziehbarkeit, klare Schwerpunktsetzung sprachliche/grammatikalische Schwächen etc.). Dieses muss aus mind. 400 Zeichen bestehen und soll die Möglichkeit einer ersten Einschätzung der Arbeit geben. Unschädlich ist, dass Juristen Themen der Informatik oder umgekehrt bewerten.
Vortrag: 20 Minuten mit anschließender Diskussion (30 Minuten bei Vorträgen von Zweier-Teams). Die Vorträge sollten sich an ein Publikum aus dem eigenen Fach (also Informatik bzw. Jura) richten, aber ohne spezielle Kenntnisse im Thema des Seminars verständlich sein. Idealerweise sind die Vorträge so gestaltet, dass Fachfremde (also z.B. Juristen bei Informatikvorträgen) die grundlegende Problemstellung erfassen können; eine Verständlichkeit des ganzen Vortrags für Fachfremde ist aber nicht erwartet.
Die Gesamtnote ergibt sich aus dem Preprint, den verfassten Reviews zu den Ausarbeitungen der anderen Studierenden, dem Vortrag und der finalen Seminararbeit.
Die Nichtabgabe des Abstracts, des Preprints, der Reviews oder der finalen Seminararbeit sowie ein Nichterscheinen zum Vortragstermin bzw. den Vortragsterminen führt zum Nichtbestehen des Seminars.
Requirements: Es wird erwartet, dass die Teilnehmer in der Lage sind, Vorträgen in deutscher Sprache zu folgen und Ausarbeitungen in deutscher Sprache im Rahmen des Peer Reviews zu lesen (eigene Vorträge und Ausarbeitungen können aber in deutscher oder englischer Sprache angefertigt werden).
Places: 7
The pivotal role of software in our modern world imposes strong requirements on quality, correctness, and reliability of software systems. The ability to understand program code plays a key role for programmers to fulfill these requirements. Despite significant progress, research on program comprehension has had a fundamental limitation: program comprehension is a cognitive process that cannot be directly observed, which leaves considerable room for (mis)interpretation, uncertainty, and confounding factors. Thus, central questions such as “What makes a good programmer?” and “How should we program?” are surprisingly difficult to answer based on the state of the art.
Recently, researchers began to lift research on program comprehension to a new level. The key idea is to leverage recent methods from cognitive neuroscience to obtain insights into the cognitive processes involved in program comprehension. Opening the “black box” of human cognition will lead to a breakthrough in understanding the why and how of program comprehension and to a completely new perspective and methodology of measuring program comprehension, with direct implications for programming methodology, language design, and education.
One of these novel methods is eye tracking, a small device observing the focus of a programmer’s eyes as x and y coordinates on the screen over time. Using this method we can answer research questions on visual attention, comprehension strategy, and differences between programmers (such as expertise).
In this seminar, you will be able to experience major steps of an eye-tracking study. You will review eyetracking literature and pose your own research questions based on existing study. Next, you can exploratively answer your research question by analyzing an existing set of eye tracking data. Specifically, each participant has to perform a literature search and propose a set of research questions targeting code comprehension measured with eye tracking. In the next step, we will provide you with real world eye-tracking data from our prior experiments. Subsequently, the research questions, the results of the analysis, and the interpretation of the results have to be incorporated into a presentation and a written thesis.
To aid the literature search, the analysis, and the presentation, this seminar includes multiple preparatory sessions at the beginning of the semester. The student presentations will be held on-site on two days in March (10. and 11.03.2025). All other sessions will take place on-site at the university on Thursdays 12:15 PM - 2:00 PM. Participation in all sessions is mandatory.
The first meeting will take place on Thursday November 07, at 12:15 PM. Further information will be provided via e-mail after registration.
Requirements: Basic knowledge on software engineering and programming.
Places: 12
The Web Security Seminar will teach students to present, analyze, discuss, and summarize papers in different areas of Web security. The seminar combines a reading group with (almost) weekly meetings and a regular seminar, where students will write a seminar paper.
Each student will get a topic assigned, consisting of a lead and a follow-up paper. The student will present the follow-up paper in a 20-minute presentation followed by a 10-minute Q&A. Afterwards we will all discuss the lead paper as a reading group. All students must read the lead paper and, before each session, must submit a summary with strengths and weaknesses.
Finally, each student will write a seminar paper on the topic assigned to them, for which the two papers serve as the starting point.
Places: 10
Optimization lies at the heart of many machine learning algorithms. This seminar teaches an overview of modern mathematical optimization methods, for applications in machine learning and data science. In particular, we will discuss the theoretical basics of stochastic optimization, scalability of algorithms to large datasets, and challenges in distributed optimization, such as for instance in decentralized or federated machine learning. We will cover a set of foundational papers, but also a selection of recent publications.
Organization:
- There is no weekly meeting. The presentations will be clustered into 2-4 slots (dates be decided) roughly between end of November - January.
- A kick-off meeting will be held virtually during the second week of the semester.
Note that - besides the block format - there are some written deliverables due during the semester, as well as a (mandatory) meeting with the tutor to get feedback on your report and presentation slides.
Additional information can be found at https://cms.cispa.saarland/optml_seminar_24/.
Requirements: This seminar aims primarily at master students in Computer Science or related fields. Previous experience in machine learning, data analysis, or optimization is beneficial.
Places: 12
In recent years, large-scale machine learning models, including large language models (LLMs) and vision transformers, have significantly advanced the fields of natural language processing and computer vision. This seminar will provide an in-depth exploration of the challenges and strategies involved in training these massive models, which require enormous computational resources, data, and expertise.
Key topics include:
- Scaling Up Neural Architectures: Understanding the evolution from traditional neural networks to large-scale models, including transformer models, such as GPT, Llama, and Mistral.
- Data Requirements and Preprocessing: Techniques for managing vast amounts of training data, addressing biases, and ensuring quality.
- Training Infrastructure: Insights into the hardware and software requirements for training, including the use of distributed computing, GPUs, TPUs, and large-scale parallelization strategies.
- Optimization Techniques: Advanced methods for optimizing training, such as gradient accumulation, mixed precision training, and addressing issues like vanishing gradients and overfitting.
- Fine-Tuning and Transfer Learning: How to efficiently adapt large models for specific tasks with limited data and resources.
- Cost Management: Detailed analysis of the costs associated with training large-scale models, including computational, storage, and maintenance expenses, and strategies to optimize budgets.
- Business Model Development: Exploring how large-scale ML models can drive value creation, including use cases, market analysis, monetization strategies, and integrating AI into existing business models to generate sustainable revenue.
This seminar is aimed at computer science students majoring in data science, AI, and machine learning who want to scale their models to meet the growing demands of real-world applications. Each student will conduct a thorough study of a specific model by reviewing its documentation, research papers, and performance metrics. They will present their findings in the seminar presentation and write a short report. Participants will leave with a solid understanding of current best practices and future directions in training large-scale machine learning models.
Organization: The kick-off date will be announced soon. Information about the seminar, important dates and further specifications will be available in Moodle: https://moodle.uni-saarland.de/course/index.php?categoryid=448
Requirements: Computer Science master students with successful participation in the “Data Science” lecture. Familiarity with neural networks, basic knowledge of deep learning, and an understanding of machine learning workflows.
Places: 6