Seminar Applied Artificial Intelligence

Content

A wide variety of topics from the field of artificial intelligence, as defined by members of our research group. 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 presentation is carried out in the form of a block event. More details will be given in the obligatory introductory meeting.

Requirements

This seminar is offered to both Bachelor and Master students. Registration via OLAT 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 according to the number of topics available.

Materials

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

Organisation

The seminars of all students take place as a block event at the end of the semester.

There will be a mandatory introductory online meeting via BigBlueButton (BBB), where the course organization and topics of the seminar will be presented, and the OLAT access code will be published. Deadline for registration on OLAT is 28.10.2025.

• Introductory meeting: 27.10.2025 (Monday), 13:00 - 14:00
• Meeting link: https://bbb.rlp.net/b/nat-g1u-mfm-mdv
• Participant Access code: 391026

During the lecture period, students will work on the topics. Discussions with their supervisor will take place individually. We offer a mandatory one-hour course about scientific writing and working with LaTeX.

The paper is to be written in English and should be of minimum 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, including questions. The final presentations will take place as in-person event at the DFKI building. Students should follow the provided mandatory LaTeX template for their seminar paper and an optional template for the presentation slides.

Topics

The format of the list is [Student Level Preference] Seminar Topic (Supervisor Name). Student level preference could be Bachelor/Master/Any. Bachelor students are eligible for topics marked as Bachelor or Any. Click the topics to reveal a short topic description.

After the introductory meeting, the Topic Assignment will take place via OLAT based on a topic preference survey.

NOTE: Topics marked as {PRE-ASSIGNED} are not available for assignment.

• [Master] Neural Image Compression {PRE-ASSIGNED}

  Neural image compression is an emerging field that leverages deep learning techniques to efficiently encode and reconstruct images with minimal perceptual loss. Unlike traditional codecs such as JPEG or HEVC, which rely on handcrafted transforms and quantization schemes, neural compression models learn compact and adaptive representations directly from data. This approach enables better rate–distortion performance, perceptual quality, and adaptability to diverse content types. Current research explores end-to-end learned codecs, attention-based entropy models, generative priors for realistic reconstruction, and hardware-efficient implementations for deployment. Expectations: Conduct a survey based on the starting papers COIN and COOL-CHIC.
• [Master] Explaining Deep Neural Networks for Medical Imaging

  Deep neural networks (DNNs) have shown great potential in medical imaging; however, their ""black-box"" nature poses significant challenges for clinical acceptance and trust. Concept-based explanation methods aim to interpret model decisions by linking them to clinically relevant, human-understandable concepts, such as specific tissue types or disease indicators. These explanations not only enhance transparency in critical diagnostic processes but also help identify biases within models and detect potential novel biomarkers. This seminar will provide a comprehensive review of recent advancements in concept-based explainability methods within the medical imaging domain, focusing on techniques such as Testing with Concept Activation Vectors (TCAV) and related interpretable frameworks. Related Papers: Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV), Coherent concept-based explanations in medical image and its application to skin lesion diagnosis. Expectations: Students will summarize state-of-the-art methods, critically assess their limitations, and identify gaps in existing methodologies to propose directions for improving model interpretability in clinical applications.
• [Master] Explainable AI Methods for Video Understanding

  Deep neural networks (DNNs) have achieved state-of-the-art performance in various video understanding tasks, including action recognition, anomaly detection, and video question answering. Despite these advances, their complex and non-transparent decision mechanisms hinder interpretability and trust, particularly when deployed in safety-critical applications such as autonomous driving, medical video analysis, or behavioral monitoring. Explainable AI (XAI) methods for video models aim to uncover how spatiotemporal features influence model predictions. Recent works have extended image-based explanation techniques—such as saliency maps and relevance propagation—to the temporal domain, while newer methods employ generative models to provide counterfactual or semantic-level explanations. This seminar will present a structured overview of existing XAI approaches for video data, emphasizing approaches such as saliency mapping, temporal relevance propagation, and counterfactual video generation methods. Related Papers: Tcam: Temporal class activation maps for object localization in weakly-labeled unconstrained videos, Exploring explainability in video action recognition. Expectations: Students will review and summarize key research papers, critically compare their methodological differences, evaluate quantitative and qualitative metrics for explanation quality, and discuss challenges such as temporal coherence, multi-frame attribution, and computational complexity. The seminar aims to equip students with a solid understanding of current trends and open problems in explainable video AI research.
• [Master] Open World Continual Learning

  Open World Continual Learning (OWCL) is an emerging paradigm that combines ideas from Continual Learning (CL) and Open Set Recognition. While CL focuses on learning new tasks or labels over time without forgetting past ones, most existing methods operate in a "closed world", i.e, they assume a fixed set of known labels. However, in real-world scenarios, when a sample with new label is presented at test time, the model must recognize it as "unknown". OWCL enables models to not only learn continuously but also detect and adapt to new, unseen classes as they appear. Expectations: Review and summarize recent works that integrate open set recognition in continual learning frameworks. Identify their key limitations and suggest future directions.
• [Master] Recent Advances in Generating Time Series Data

  Time series data is omnipresent. We find them in various fields, such as medicine, finance and in our everyday life. Smart devices can track all kind of measurements over time, which contain valuable information that machine leaning methods can help to extract. One of the big limiting factors for machine learning applications in industry is data sparsity. Generative machine learning models, such as diffusion models can synthetize data and help to increase diversity of the available data. We are interested in recent techniques for generating time series data. Techniques for class-conditional generation are of highest interest, but generation for forecasting and imputation are also relevant. Expectations: Review, summarize and compare recent research papers on the series synthesis with machine learning models, especially diffusion models.
• [Master] Too Much Information: The Cognitive Limits of Understanding Explanations {PRE-ASSIGNED}

  This work explores the balance between informativeness and simplicity in human-centered explanations. Drawing from psychology, cognitive science, and explainable AI (XAI), it reviews empirical and theoretical studies to identify when an explanation becomes too detailed to be effective. The goal is to derive practical insights and rules of thumb for designing explanations that are sufficiently rich to foster understanding yet parsimonious enough to avoid cognitive overload. Expectations: Identify resources on cognitive load and similar from psychology and cognitive science. Additionally, identify common rules of thumb for good explanation sizes in XAI literature
• [Master] Bias Mitigation in Medical Image Processing {PRE-ASSIGNED}

  Your task is to discover, design, and evaluate a reproducible pipeline that audits, quantifies, and reduces bias in a medical image classification task (e.g., chest X-ray abnormality or dermoscopy lesion). Identify sources of bias (demographics, imaging site/scanner, prevalence shift), apply mitigation techniques (reweighting, group DRO, domain adaptation/harmonization, adversarial debiasing), and produce a fairness report alongside clinically meaningful metrics. Expectations: Conduct a survey with at least 30 reference papers.
• [Master] Evolution of Dataset Distillation

  Dataset distillation is the process of creating a small, synthetic dataset that can train models to achieve performance comparable to training on the full real dataset. Originally, the goal of dataset distillation was to compress large datasets into a few representative samples to save computation and storage while retaining model accuracy. However, over time, techniques have evolved significantly. Newer approaches use generative models, meta-learning, and diffusion-based synthesis to distill datasets. The objectives have also shifted: modern distillation methods often aim to improve generalization, privacy preservation, domain adaptation, or continual learning rather than pure compression. Start here: https://arxiv.org/pdf/2502.05673. Expectations: For this task, you should review the progression of dataset distillation methods, identify and summarize the latest key works and their contributions, and discuss what current research gaps remain, for example, limitations in scalability etc.
• [Master] State-of-the-Art in Video Understanding and Classification

  Video understanding, with action and event recognition at its core, is a critical area of computer vision focused on enabling machines to interpret temporal information in video data. This seminar provides a focused exploration of the current State-of-the-Art in Video Classification, specifically examining the transition from traditional 3D CNNs to powerful Transformer and multimodal architectures. The student is tasked with conducting a mini-survey by reading and critically synthesizing the findings of recent, influential papers from top conferences. The main goal is to compare the core architectural innovations, training paradigms (like Self-Supervised Learning), and efficiency techniques that define top performance, ultimately producing a concise written survey and an informed presentation on the field's technological frontier and future challenges. Expectations: Conduct a survey on recent approaches in video understanding and compare methods along multiple self-chosen dimensions
• [Master] Training-Free Medical Image Editing {PRE-ASSIGNED}

  Training-free image editing refers to methods that modify images using pretrained generative models without any additional training or fine-tuning. Unlike approaches that rely on model retraining—which can be computationally expensive, require large datasets, and risk overfitting—training-free methods operate entirely at inference time. They leverage the rich prior knowledge already encoded in powerful generative models and guide them to produce controlled edits while keeping model weights frozen. Expectations: Conduct a survey on recent approaches in training-free image editing from text and/or other labels, including concept sliders, with the secondary focus of medical images, in particular skin lesion.
• [Master] Deep Learning for Unsupervised Cell Segmentation

  Unsupervised cell segmentation aims to delineate cellular and subcellular structures without dense human annotations, a crucial step for scalable and reproducible analysis across microscopy modalities. This seminar will provide a focused survey of the current state of the art in unsupervised and label-efficient cell segmentation. The student will review recent influential work on reconstruction-driven variational autoencoders and diffusion models, with attention to how these approaches leverage unlabeled microscopy data. The primary objective is to compare core architectural choices and training paradigms, and to assess practical strategies that enable robust instance separation under low signal-to-noise ratios, scale variability, and domain shift. Expectations: Review recent unsupervised cell segmentation methods, highlight their weaknesses, and suggest practical ways to improve them.
• [Master] Autoencoders for Cell Analysis

  Autoencoders are increasingly used to extract compact, noise robust representations from high dimensional cellular assays, spanning single cell omics and high content microscopy. This literature survey will systematically review autoencoders and synthesize evidence on their utility for core cell analysis tasks. The primary goal is to compare core architectural ideas and modeling choices for cell analysis and to report practical guidance on scalability, interpretability, and generalization. Expectations: Review recent methods, highlight their weaknesses, and suggest practical ways to improve them.
• [Bachelor] Computerized Adaptive Testing for knowledge estimation

  Knowledge estimation is essential to learning and teaching. It shows whether content has been learned and gives an indication of what to learn next. Although the concept of adaptive testing is not new, generative AI has the potential to boost its usefulness. Expectations: Explain the bases of adapting testing (e.g., Item-Response Theory), survey the literature involving Computarized Adaptive Testing (CAT) and related topics. Focus specifically on how Generative AI can be implemented to enhance CAT (e.g., for Automatic Item Generation or personalized feedback), considering its strengths and weaknesses regarding psychometric validity and content quality.
• [Master] Automatic Progress Map Generation Based on Learning Materials

  One of the major challenges in learning concerns choosing what to learn and in what order. The progress map is a useful tool for structuring knowledge to guide students in their learning path. A combination of knowledge graphs and machine learning could facilitate the automatic creation of these complex structures from raw learning materials. Expectations: Explain the concept and purpose of progress maps in education, review literature on knowledge representation methods (especially knowledge graphs and embeddings), and analyze existing machine learning and NLP approaches for generating or inferring learning paths automatically. Discuss how Generative AI might assist in content extraction or validation, and evaluate the methods used to assess the quality of the automatically generated map.
• [Any] Exploring Food Security through Earth Observation Applications

  The student will delve into the world of earth observation data and applications, machine learning tasks, and analysis techniques to identify opportunities for improving food security outcomes. Remote sensing technology, leveraging satellite and aerial imagery, offers a powerful tool for monitoring agricultural productivity, crop health, soil moisture, and weather patterns that can impact food availability. Expectations: Develop a report outlining the findings, conclusions, and recommendations for policymakers, researchers, or practitioners seeking to improve food security monitoring and prediction. Answering the following questions: What are the various applications of remote sensing data in enhancing food security monitoring? Which types of data are most suitable for analyzing food security-related metrics? How can machine learning techniques be applied to remote sensing data to support food security outcomes?
• [Master] AI-Powered Requirement Engineeing

  AI-Powered Requirements Engineering leverages artificial intelligence to automate, enhance and optimize the software requirements engineering process. This approach addresses challenges such as ambiguous or incomplete requirements, traceability and conflict detection. Unlike traditional RE methods that rely heavily on manual analysis and stakeholder interviews, AI-powered RE uses intelligent methods to extract entities, classify requirements and detect inconsistencies automatically. Current research explores multi-agent systems for collaborative requirements handling, large language models for requirement understanding, knowledge graph integration for requirement engineering. Expectations: Review, summarize and compare recent research papers on the ai-powered requirement engineering.
• [Any] Evaluation of Document-based GraphRAG

  Compared to vector database-based Retrieval-Augmented Generation (RAG), which typically involves chunking documents and retrieving chunks similar to a given question, GraphRAG retrieves relevant information from pre-generated knowledge graphs that represent the structure of documents and the relationships between concepts. This enables more context-aware and semantically rich retrieval, often resulting in improved accuracy and interpretability - particularly in multi-hop reasoning tasks or scenarios where relational context is crucial. In this seminar, you are expected to create an overview paper that evaluates such GraphRAG approaches, including the datasets used for evaluation, which phases of the indexing or retrieval process are assessed, the baseline methods employed, and the metrics used for comparison. Expectations: Review, summarize, and critically assess existing evaluation approaches for GraphRAG. The objective is to classify individual approaches into broader high-level strategies and provide a structured overview of these categories.
• [Any] Explainability in Biomolecule Design

  Explainable AI has supported increasing transparency and trustworthiness in many application domains of AI as well as helping with tasks such as model-development and debugging due to a more profound understand of the inner workings of the analyzed models. In the context of biomolecule design explainability has often focused on Feature Attribution Methods as well as sequence data. The goal of this seminar is to characterize employed explanation methods and the biomolecules they have been applied to. Expectations: Review and summarize explanation approaches by extending an existing, but outdated, survey on interpretability in biomolecule design. In particular explanation methods should be categorized by the input data and molecule type the employed AI model is working on and the corresponding explainability approach. Achieving a deep understanding of all investigated explainability approaches is beyond this seminar.
• [Bachelor] Recent Advances in AI-based Antibody Design

  Antibodies area class of proteins employed by the immunsystem to help identify malignent entities in an organism by functioning as mediators. The design of novel antibodies to counter a specific disease is a promising direction especially for high mortality rate diseases such as Cancer Immunotherapy. However, classical design requires many lab-intensive work-cycles to design even a single novel-antibody which often show sub-optimal performance. Deep Neural Networks have shown good results in designing de-Novo Antibodies with high success-rates thereby offering a promising direction towards a time and cost-efficient development of these biomolecules. The goal of this seminar is to review recent advances in DNN's for structure-based Antibody design with a specific focus on investigating how AI can support the adaption of existing Antibodies as well as the identification of suitable targets for Antibodies to bind to. Note that no prior knowledge in Biology is necessary, we will solely focus on the AI aspect of designing Antibodies. Expectations: Starting with existing reviews on Antibody development identify recent research directions with a specific focus on Eptiope discovery (binding target identification) and adaptation of Antibodies to other organsims.Based on this the student should gain a deep understanding of a few select and relevant papers and systematically compare and critically assess the employed approaches.
• [Any] Development and research of tools for teaching fairness

  Fairness is one of the crucial topics in various domains. Such as in the medical field, humans and AI should always be treated fairly to make a safe decision. Also, in education, students should be treated fairly and not discriminated against in learning. Your task is to discover various research topics, especially in the HCI (human-computer interaction) field, that address the study of "fairness" or the education of "fairness". Expectations: Conduct a survey with at least 30 reference papers. Make sure the highlight how these papers address in educating fairness and also, how are each works different. Also, make sure to show clear vision of what kind of works should be done in addition for this field. Sample papers:https://dl.acm.org/doi/10.1145/3757414, https://dl.acm.org/doi/10.1145/3613904.3642621, https://dl.acm.org/doi/10.1145/3706598.3714402
• [Master] Explainable Link Prediction for Recommendation

  With link prediction, new knowledge in knowledge graphs can be acquired by predicting new links between nodes. In practice, this can be used for recommendation engines. However, in domains such as finance, explainability and interpretability can be a key requirement. To account for application scenarios in such domains, this thesis will review the research field of explainable link prediction. Expectations: Conduct a rigorous literature review on the topic of explainable link prediction. This includes constructing a concept map (table) where all literature is compared with each other along self-chosen criteria.
• [Any] NLP Datasets for Healthcare

  The idea is to conduct a literature survery for exploring public datasets in NLP domain focusing healthcare applications. The datasets should be mainly related to classification tasks, and accessible for public use.
• [Any] Recent Methods in NLP-based Healthcare

  The student is asked to explore recent AI based methods using NLP to advance healthcare domain domain. The focus will be on the classification tasks.
• [Any] Multimodal Datasets for Healthcare

  The idea is to conduct a literature survery for exploring public datasets using multimodalities for advancing healthcare domain. These modalities can be combiniton of Image-Text, Audio-Video etc. modalities.
• [Any] Recent Multimodal AI Methods in Healthcare

  The task is to find recent AI methods focusing on multiple modalities such as Image-Text, Audio-Video etc, used to support healthcare systems.
• [Master] Explainable AI in Financial Risk Prediction

  Explainable models such as logistic regression or decision trees play an important role in the context of ML-based analysis of corporate financial data, especially in the context of financial risk prediction. In contrast to black box models, XAI models can reveal which features and financial signals can lead to critical developments in the corporate domain. Expectations: Review and summarize recent research papers.
• [Bachelor] Survey on Document Large Language Models (LLMs)

  Multimodal large language models are rapidly becoming an integral part of modern end-to-end document analysis pipelines. Recently a huge number of benchmarks, papers, and models have been proposed in this direction and a comprehensive survey is needed to summarize the key contributions and future directions in this area. Expectations: The student is required to conduct a survey on latest document LLM models for various document tasks and provide a comprehensive survey on which models are the current state-of-the-art along with their novel aspects when adapting to the document domain. Review of at least 30 papers is expected.
• [Any] A review of privacy attacks on large language models (LLMs)

  Privacy is a major concern in existing landscape especially after several EU laws have been passed especially when it comes to textual data in legal and financial domains. The goal of this topic is to survey existing works that have applied privacy attacks on large language models in text domain to assess their danger in existing landscape. Expectations: Review of at least 30 papers is expected. Find and review existing works with privacy attacks on LLMs and highlight their key contributions.
• [Any] When Boundaries shape results - the modifiable area unit problem in Urban Analysis

  Urban data are often aggregated within administrative boundaries such as districts or neighborhoods. Yet these boundaries may have effects on analyses: different spatial delineations can lead to different statistical patterns and interpretations. This seminar paper explores the Modifiable Areal Unit Problem (MAUP) in urban contexts, examines its impact on spatial decision-making and urban research, and discusses methodological strategies to detect and mitigate aggregation bias.

Contact