Seminar Applied Artificial Intelligence

Content

Selected topics from the field of socio-technical knowledge (see topics of the lecture Collaborative Intelligence). The seminar teaches the students how to write and present a scientific paper on a specific topic. Students are also introduced to doing a literature review of scientific papers. The final presentations will be carried out in the form of a block event, which will be in-person at the DFKI site in Kaiserslautern. More details will be given in the mandatory introductory meeting.

Requirements

This seminar is offered to both Bachelor and Master students. Registration via OLAT course is required for this seminar; the access code will be given in the introductory meeting. As we provide each student with a topic and a tutor, the number of seminar places is limited to the number of topics available this semester (see below).

Materials

You can find all course materials, news and information in the OLAT course. The access code for the OLAT course will be given in the introductory meeting.

Organisation

The seminar presentations of all students take place as a block event at the end of the semester. This will be organized in-person at the DFKI site in Kaiserslautern.

There will be a mandatory introductory online meeting via BigBlueButton (BBB), where the form and topics of the seminar will be presented, and the OLAT access code will be published:

During the lecture period, students will work on the assigned seminar topics with guidance from their supervisors. Discussions with their supervisor will take place individually. Meetings with the supervisors and organizers take place virtually. Based on requiremernt, we offer a two-hour course about scientific writing and working with LaTeX.

The paper is to be written in English and should be of length 10 pages (Bachelor: 8 pages) at the end. The presentations, which are also given in English, take place at the end of the semester and last about 25 minutes each (Bachelor: 20 minutes), including questions. Students should follow the provided templates for their seminar paper.

Topics

List of topics with descriptions and preferred level of student (Bachelor/Master/Any).

First round of topics:

  • [Any] Updates in Efficient Vision Transformers Vision Transformers (ViTs) have achieved remarkable success in computer vision, yet the quadratic complexity of their core self-attention mechanism remains a significant bottleneck. This seminar explores recent advancements tackling this challenge, focusing on novel, efficient attention variants. We will review various strategies designed to reduce computational cost and memory footprint while maintaining competitive performance. The student will gain insights into the latest techniques driving the development of more practical and scalable ViT architectures.
  • [Any] Multimodal LLMs in Document Analysis LLMs have recently become quite popular in document analysis to handle complex tasks such as classification, entity extraction, layout analysis, visual question answering, etc., in a zero-shot manner. Unlike standard language models that are related to a single task, Document LLMs use generative modeling to target all tasks with a unified objective. The area has seen significant popularity recently, with a growing number of models proposed in recent years. This seminar topic will focus on surveying the recent Document LLM models.
  • [Any] Biomolecule Design using Artificial Intelligence In recent years, Artificial Intelligence (AI) has become a powerful tool for designing new biomolecules—such as proteins, peptides, antibodies, compounds, and drugs. This seminar will introduce students to how AI is being used to create biomolecules with specific structures and functions—something that was once only possible through trial and error in the lab. We will explore different AI-based approaches, such as deep learning and generative models, and look at how these methods are helping scientists design biomolecules for medicine, industry, and research. The seminar will also explain the main types of techniques used and highlight key examples and breakthroughs in this exciting field.
  • [Any] Evaluation Strategies in AI-Driven Biomolecule Design How do we measure the success of AI-designed biomolecules like proteins, peptides, antibodies, compounds, and drugs? This seminar surveys the various evaluation strategies and metrics used across biomolecule design pipelines. Students will explore how different types of biomolecules are assessed—for example, using structure prediction accuracy, binding affinity, stability, or biological activity—and whether the same evaluation methods can be applied across different molecule types. Through this seminar, students will gain a clear understanding of the strengths and limitations of each strategy, and learn to critically compare how evaluation differs between domains like protein engineering and drug design.
  • [Any] A Survey of Protein-Peptide Interaction Predictors, and Datasets Protein–peptide interactions are central to many biological processes and therapeutic strategies. This seminar surveys the growing field of computational predictors developed to model these interactions. Students will explore a variety of prediction approaches—from classical machine learning models to modern deep learning frameworks—and how they differ in architecture, input features, and evaluation strategies. A key focus will be on the datasets used to train and benchmark these models: how they were collected (e.g., from databases like PDB, BioLiP, or curated experimental results), what types of protein–peptide complexes they contain, and how data quality and redundancy are addressed. The seminar will critically assess whether existing studies ensure fair comparisons across models—considering factors like consistent preprocessing, identical test sets, and standard evaluation metrics. Through this, students will gain insight into both the technical development and the benchmarking rigor required in this rapidly evolving field.
  • [Any] AI for Good: The Sugarcoat Mainstream media and academic discourse often promote the "AI for Good" narrative to convince the public of the positive impact this field can achieve, and encourage investors to pour their money in developing more AI systems and funding related research. Students are also impacted by this hype. Yet, in parallel to the large benefits of AI, it is also actively being used for harmful purposes. This seminar will examine major fields where AI is being exploited irresponsibly, and map the specific machine/deep learning tasks and research areas that enable such misuse.
  • [Any] Employing Long-Term Memory Techniques to enhance Large Language Model Conversations Large Language Models (LLMs) or LLM-based tools are often used in a chat-based manner enabling users to engage in multi-turn conversations. Recent LLMs support long context windows meaning they can handle long prompts. In the chat context, this capability allows to include multiple previous messages as context for a present user query to improve the contextual relevance of the LLM answer. However, there comes a point where the context window size is fully utilized making it impossible to include the entire previous conversation. Furthermore, passing the full conversation is not always beneficial since topics may shift over time and earlier interactions might become irrelevant. Additionally, in long prompts, a phenomenon known as ""lost-in-the-middle"" can occur, where content in the middle of the prompt gets overlooked by the LLM. Given these challenges of recalling relevant contents of past conversations, various approaches have been developed that establish external long-term memory storages and retrieve relevant historical interaction data during interaction time. For this seminar topic, a comprehensive survey and classification of methods enhancing LLM conversations through the use of long-term memory should be conducted. Here, it is particularly interesting how the different approaches process, enrich and store conversation data in the long-term memory.
  • [Master] How to Fool a Neural Network: Adversarial Attacks Adversarial attacks involve subtly altering inputs to deceive machine learning models, often without detection by humans. These attacks expose vulnerabilities in AI systems, particularly in image classification and natural language processing. Key methods include gradient-based techniques like FGSM and PGD, optimization-based approaches such as the Carlini & Wagner attack, and decision-based methods like the Boundary Attack. Your task is to explore how these attacks are constructed, what assumptions they rely on, and how their effectiveness is measured. A narrative starting point for your research is {https://medium.com/sciforce/adversarial-attacks-explained-and-how-to-defend-ml-models-against-them-d76f7d013b18}.
  • [Master] Autoregressive models for image generative modeling Autoregressive (AR) models are the current state-of-the-art (SOTA) technology for language generative models, thanks to its next-token prediction paradigm. Even though their concepts are used in some visual generative models, Diffusion models have been the most representative approach in this field. In this seminar, you will focus your review on "xAR: Autoregressive (AR) Visual Generation with Next-V Prediction"; a brand-new publication that surpasses diffusion SOTA performance with AR models. The goal of this seminar is to learn the foundation of AR modeling, and what modifications/considerations must be considered to adapt them to visual generation and understand its behavior with visual data.
  • [Master] Recent Advances in Self-Supervised Learning for Time Series Time series data is essentially omni-present. Any kind of measurement over time results in time series. Machine learning models, such as neural networks, can help us to analyze this kind of data. However, they require large datasets for training. While labeled data can be expensive to acquire due to the need of human supervision or expensive measurements, unlabeled data is often available in larger quantities. Self-supervised representation learning methods allow to pre-train models on unlabeled data. There have been significant improvements at such methods in recent year. The task in this seminar is to find and compare recent work on self-supervised learning methods for time series analysis.
  • [Master] Evaluation of LLM Robustness Robustness is the ability of AI models to behave in a predictible way in the event of an unknown or malicious input. Evaluating Robustness is important as it can be a signal for high variance in real world settings and also identify other attack vulnerabilities. Sufficient Robustness is also a regulatory requirement. This seminar will focus on how Robustness can be evaluated in LLMs. A good reference for this topic is the NIST AML document: https://csrc.nist.gov/pubs/ai/100/2/e2023/final
  • [Master] Continual Learning in Tabular Data: Methods and Challenges This seminar explores the emerging field of Continual Learning (CL) in the context of tabular data, a domain central to many real-world applications like finance, healthcare, and cybersecurity. We discuss the unique challenges posed by feature heterogeneity, data drift, and catastrophic forgetting, as well as recent advances in rehearsal strategies, adaptive models, and task-agnostic learning.
  • [Master] Explaining Deep Neural Networks for Medical Imaging Deep neural networks have shown great potential in medical imaging; however, their "black-box" nature poses challenges for clinical acceptance and trust. Concept-based explanation methods interpret the decisions made by these models using clinically relevant, human-understandable concepts, such as certain tissue types or disease indicators. These explanations also help identify biases within the model and detect potential novel biomarkers, improving transparency in critical diagnostic processes. This seminar will provide a comprehensive review of recent advancements in concept-based explainability within medical imaging, focusing on methods such as Testing with Concept Activation Vectors (TCAV) and other interpretable frameworks. Students will summarize state-of-the-art methods, highlight their limitations, and identify gaps in existing methodologies to propose improvements for interpretability.
  • [Master] Adversarial Attacks to Counterfactual Explanations Counterfactual explanations help us understand AI decisions by showing how small, meaningful changes to an input can alter predictions. Interestingly, adversarial attacks—designed to fool models—follow a similar principle of minimal changes but imperceptible to human eye. This seminar explores how adversarial techniques can be adapted to generate realistic and actionable counterfactual explanations. Student will analyze state-of-the-art approaches and theoretical frameworks to understand the intrinsic link between adversarial robustness and explainability and critically discuss challenges in existing methodologies to design more trustworthy and transparent AI solutions.
  • [Master] Disentanglement Techniques for Improving Explainable AI Explainable AI (XAI) methods are crucial for gaining trust and ensuring transparency in the decisions made by machine learning models. While various XAI approaches, such as concept-based and counterfactual explanations, offer valuable insights, they often entangle important model features. This seminar will provide a comprehensive review of recent research on disentanglement techniques for explainable AI, emphasizing methods that highlight distinct factors influencing the model’s decisions. Student will summarize the state-of-the-art methods, the loss functions designed to achieve disentanglement, and the evaluation metrics used to assess the quality of these disentangled explanations.
  • [Master] AI-based EEG Analysis for Distinguishing Information Processing and Retrieval Electroencephalography (EEG) is a non-invasive way to measure brain activity and has been widely used to study how we think, learn, and remember. In particular, it can help us understand the difference between when someone is actively processing new information and when they are retrieving something from memory. Your task is to explore how artificial intelligence (AI) can be used to tell these two mental states apart using EEG data. You will start by learning about the two main ways of analyzing EEG: event-related potentials (ERPs), which look at the brain's response to specific events, and spectral analysis, which focuses on brainwave patterns from specific frequency bands (e.g., alpha, theta). This will help you understand the problem space and choose the best AI methods depending on the task. Next, you will review how AI methods—from traditional machine learning to deep learning—are used to analyze EEG and classify cognitive states. Your literature review should cover the entire pipeline: from data preprocessing and feature extraction to model selection and evaluation. Finally, reflect on the current challenges in this field, such as small datasets, differences between individuals, and how to make AI models more interpretable. Suggest possible improvements or research directions based on the gaps you identify.
  • [Master] Understanding and Measuring Catastrophic Forgetting Catastrophic forgetting remains a significant challenge in continual learning, where neural networks tend to forget previously acquired knowledge upon learning new tasks. This seminar aims to gain a deeper look into the underlying causes of catastrophic forgetting and explore various evaluation metrics used to assess its impact. By critically analyzing recent research articles published between 2022 and 2025, participants will gain a comprehensive understanding of the mechanisms contributing to catastrophic forgetting and the methodologies employed to mitigate it
  • [Master] Enhancing Auto Machine Learning (AutoML) with LLM-based Intelligent Agents Large Language Model (LLM)-based agents are emerging as powerful tools for solving complex tasks autonomously, including those in machine learning. In this literature review, you will explore how LLM-driven intelligent agents can be applied to enhance AutoML systems. The focus will be on understanding current research trends, key frameworks, and how these agents can support or even automate parts of the AutoML pipeline—such as feature engineering, model selection, and hyperparameter tuning. You will also identify challenges and future research directions in this rapidly evolving field.
  • [Master] Temporal Self-Refinement for Video Instance Segmentation Object segmentation is a fundamental challenge in computer vision, often reliant on extensive annotation procedures. Video data introduces new opportunities for improving segmentation quality without needing more labeled data. By leveraging temporal consistency of objects, segmentation masks can be self-refined without needing additional human input. Techniques in this area often rely on mask propagation or self-supervised learning, where models learn and extrapolate from their own predictions instead of requiring human-made ground truth labels. By exploring these ideas, you will investigate how segmentation methods can move beyond traditional, fully supervised training and instead towards more autonomous systems that learn directly from raw video data with only minimal human supervision.
  • [Master] Exploring Gaussian Processes in Earth Observation Application This seminar will explore the application of Gaussian Processes (GPs) in Earth Observation (EO), focusing on their strengths in modeling spatial-temporal correlations, handling uncertainty, and integrating heterogeneous data sources. GPs provide a flexible, non-parametric Bayesian framework well-suited to EO tasks such as environmental variable estimation, data fusion, and spatio-temporal interpolation. The review will highlight advances in scalable GP models, informed kernels, and multi-output frameworks, emphasizing their potential in fusing satellite and in-situ data and guiding adaptive sampling. Key research directions include improving scalability, incorporating physical constraints, and developing expressive, task-specific kernels.
  • [Master] The Role of Knowledge Graphs in Industrial Green AI Green of AI and Green by AI are topics highly discussed and examined by companies across industries (for example in the finance sector). This seminar thesis will review approaches with the question of how knowledge graphs can contribute to Green AI in an industrial setting.
  • [Any] Visualizing Uncertainty and Missing Data in Spatial Analysis The visualization of spatial data plays a crucial role in decision-making processes across various disciplines, from urban planning to environmental monitoring. However, spatial datasets often contain uncertainties or missing values due to data collection limitations, processing errors, or privacy constraints. This seminar explores methods for representing uncertainty and incomplete data in spatial visualizations, addressing questions such as: How can cartographic techniques highlight unreliable data? What are effective strategies for communicating missing geospatial information to end-users?
  • [Any] Robustness evaluation of models for tabular data Robustness is the ability of AI models to respond reliably to unexpected or malicious inputs. It is critical for identifying vulnerabilities and ensuring consistent performance in real-world scenarios. This seminar explores methods for evaluating robustness in models built for tabular data.

Second round of topics:

  • [Any] Urban Data Bias Investigating Missing Information and Unequal Spatial Distributions: Urban data is increasingly used to inform planning, resource allocation, and policy-making. However, such datasets often reflect underlying biases—not only through missing or incomplete information, but also via structurally unequal spatial distributions of infrastructure, services, and accessibility. This seminar explores the multifaceted nature of bias in urban geodata. It addresses questions such as: Where and why do data gaps occur? How do these gaps correlate with demographic and socioeconomic factors?
  • [Any] Diffusion Models for Time Series Anomaly Detection Diffusion models, originally developed for image generation, are now being adapted to detect anomalies in time series data. By learning to reverse the noise process applied to normal data, diffusion models can reliably model normal patterns. This seminar explores how these models work, why they are effective, and how recent research improves their ability for time series anomaly detection tasks in practical scenarios.
  • [Master] Distilling Specialized Knowledge into Foundation Models This seminar explores how multiple task-specific models, each trained to excel at a particular function within the same domain, can transfer their knowledge into a single foundation model. By distilling specialized expertise such as segmentation, classification, or detection into a unified model, we can build systems that retain high performance across tasks while improving scalability, efficiency, and deployment. The report will examine key distillation techniques, the challenges of integrating diverse outputs, and relevant applications in areas like medical imaging and computer vision.
  • [Master] AI-based Microscopic Cell Analysis for Disease Detection  Microscopic cell analysis plays a vital role in pathological investigations, enabling the precise identification and classification of abnormalities, which is essential for diagnosing various diseases. Early and accurate detection of cellular-level diseases is critical for ensuring effective diagnosis and timely treatment. Automated detection using artificial intelligence (AI) offers a promising alternative to traditional laboratory-based methods, potentially saving both time and cost. This seminar aims to emphasize the key features essential for accurate disease detection and to discuss the major challenges and gaps that still exist in current detection methodologies.
  • [Master] Sensor-invariant Modeling in Earth Observation Sensor-invariant modeling refers to the development of models that produce consistent outputs regardless of the sensor used to capture the input data. Technically, this involves learning representations that are robust to variations in sensor characteristics such as spatial resolution, spectral bands, radiometric calibration, and noise profiles. Recent advances leverage domain adaptation, or contrastive learning to align features across sensor domains. As Earth observation data volume grows and sensors evolve, sensor-invariant modeling becomes key to maximizing data utility and supporting informed environmental decision-making. This seminar has the purpose of studying sensor-invariant modeling in the Earth observation domain, as well as relate it to permutation/fusion invariant and generalization for not-trained sensors (i.e. zero-shot learning).
  • [Master] Risk assessment of AI Systems Ignoring Risks of AI can have dire consequences on health, safety or fundamental rights of the society. In this seminar, we will focus on what should a risk assessment achieve and how can we assess risks of different AI systems. A good starting point for this topic is the MIT AI Risk Repository: https://airisk.mit.edu/

Contact

Next Social Web Mining