Seminar Applied Artificial Intelligence


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 presentation is carried out in the form of a block event. More details will be given in the obligatory introductory meeting.


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 accordingly.


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.


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 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 topics. Discussions with their supervisor will take place individually. We offer a mandatory 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 obligatory templates for their seminar paper as well as for the presentation slides.


List of topics with a short description and the corresponding supervisor:

  • Review of methods for cell tracking (Nabeel Khalid)Object tracking has many applications in various fields specially in biomedical domain. Cell tracking can be used to detect changes in cell movement patterns which can then be used to detect diseases, help in development of medicine etc. There are different approaches available for cell tracking using both traditional computer vision approaches and deep learning based approaches. This survey will focus on comparison of different approaches for cell tracking.
  • Anomaly Detection using GANs (Christian Nitsche)Anomaly detection occupies an important position among data science methods. In some cases, it is regarded as a standalone task as for example in fraud detection or operative analysis of network traffic. In other cases, it serves as preparation step for further processing, for example as outlier detection. That way, it is an important building block in the data flow of analysis. Also in the automotive domain it is used for a variety of tasks. The applications include classical detection of part or software failures, detection of deviations in change management, predictive maintenance, diagnosis, assistance in checking requirements for homologation as well as securing models. The domain brings its special set of challenges: most of the data are time series, which have their own difficulties, the amount of data is often high and most of all, the data has many dimensions. Depending on the application, extremely high precision is demanded, which sometimes even comes from legal requirements. Together, these challenges bring many methods to their limits. Recently, the use of GANs was proposed for anomaly detection and a range of GAN-based methods was developed, which promise a gain in accuracy. Starting from a list of methods in example applications, you will search more methods in the literature, evaluate how well they are suited for typical applications in the automotive space and compare them. There is a project associated with this seminar. We would like it if you took it, too. We think this might develop into a cool master thesis in our IAV-DFKI transfer lab "FLaP".
  • The Time Series Data Zoo (Peter Schichtel)Time Series Analysis in general is one of the most pressing problems in industry and science (think: health, hazard protection, IoT, physics, autonomous systems, energy, climate change, ....). While this field has witnessed a tremendous groth in the past decade, especially with the dawn of deep learning, it is far from clear what "a good" time series analysis method is. Solutions and algorithms are scattered over many domains and tasks. Different methods challange priority. Here we will provide you with an extensive list of time series data sets publically available. Your challange will be: Can you find more data sets with download links? Find the actual state-of-the-art publication for each data set. Can you sort the available information? Is there code associated with the publication? Can you present your overview in a comprehensive way? Both as resentation as well as written report? There is a project associated with this seminar. We would like it if you took it, too. We think this might develop into a cool master thesis in our IAV-DFKI transfer lab "FLaP"
  • UNet's in time series analysis (Peter Schichtel)Time Series Analysis in general is one of the most pressing problems in industry and science (think: health, hazard protection, IoT, physics, autonomous systems, energy, climate change, ....). On the other hand, Unets pose a powerful ansatz in image analysis. Can you collect the state-of-the-art of Unets in image analysis? We are particulary interested in time series analysis. Have people started using this methods in the time series community? There is a project associated with this seminar. We would like it if you took it, too. We think this might develop into a cool master thesis in our IAV-DFKI transfer lab "FLaP".
  • Stochastic Diffusion Models in time series analysis (Alireza Koochali)Time Series Analysis in general is one of the most pressing problems in industry and science (think: health, hazard protection, IoT, physics, autonomous systems, energy, climate change, ....). On the other hand, Diffusion poses a powerful ansatz in image analysis. Can you collect the state-of-the-art of Diffusion in image analysis? We are particulary interested in time series analysis. Have people started using this methods in the time series community? There is a project associated with this seminar. We would like it if you took it, too. We think this might develop into a cool master thesis in our IAV-DFKI transfer lab "FLaP".
  • Digital Twins in Agriculture and the Food Production Chain (Ansgar Bernardi)A Digital Twin unites sensor input, high-fidelity modelling, and physical automation, thus enabling a holistic approach to monitoring and automation. Originally formulated in the context of industrial automation and Industry 4.0, the concept promises rich contributions to the digitization of agriculture. However, the open world of agriculture and the complex food production value chains pose specific challenges. We are interested in reviewing the most recent literature in order to summarize the challenges and solution approaches, and investigate possible transfer and further research to apply such concepts within the "agriculture of the future" project SUSKULT.
  • Recent advances in Domain Adaptation (Jayanth Siddamsetty)In supervised learning settings, traditionally training and test domains are assumed to follow a similar distribution. However, in reality, the distributions are different. Domain adaptation is the problem of finding a model that performs well, also on the test domain. This seminar will help in understanding the recent advances in this solving this problem.
  • Unsupervised Representation Learning on Time Series (Philipp Engler)Time series data arises from sensors everywhere around us. Machine learning algorithms can help us make sense of this data, allowing them for instance to predict the action a human is performing or to detect faults in a machine. Machine learning models, such as neural networks, typically require large amounts of data 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. Unsupervised representation learning methods, allowing to pre-train models on unlabeled data, have improved significantly and gained increased attention lately. We are interested in surveying recent literature on unsupervised representation learning in the time series domain to obtain an overview of the current state-of-the-art.
  • On Privacy-Preserving Machine Learning - Understanding the Risks (Dominique Mercier)This seminar is about the privacy preserving deep learning and the impact of different attack and defense mechanisms. Deep learning has proven to achieve incredible results in different areas. However, data sensitivity plays a pivotal role when it comes to the real-world deployment of artificial intelligence in safety-critical domain. E.g. to utilize the full potential of deep neural networks is it crucial to understand the attack mechanism and evaluate their impact on the proposed model. In conjunction with the attacks it is important to understand the impact of possible defense mechanisms.
  • Review of Camera Only Object Tracking Methods for Autonomous Driving (Abdul Hannan Khan)The aim of this seminar is to review object tracking methods for autonomous driving based on camera only. Tracking objects in a traffic scene is essential for autonomous driving systems to avoid collisions. Additional modalities to camera do provide additional preception of the environment however with a heavy computational cost. Also, multimodal object tracking systems rely heavily on camera for semantic information of the scene as it provides denser information compared to other sensors. Therefore, improvements in camera only methods for object tracking reflect on overall improvements in object tracking.
  • Review of AI Based DNA Modification Prediction Algorithms (Ahtisham Fazeel)The goal of this seminar is to develop an understanding of the main concepts related to DNA modifications and to provide an overview of the available methods for predicting DNA modifications using artificial intelligence (AI). Most of these methods are based on either machine learning (ML) or deep learning (DL). The methods of ML focus on using DNA feature extraction methods and classifiers such as RF, SVM, etc., whereas, DL methods are based on unsupervised feature representation and CNNs, RNNs, or transformers. Overall, such an overview will help scientific community to understand the current trends in DNA modification prediction.
  • Recent Advances in XAI for Medical Image Analysis (Adriano Lucieri, pre-assigned)In recent years, Deep Learning set multiple new state-of-the-arts, outperforming classical statistical methods in many fields. However, their black-box nature makes it hard to track decisions or understand the classifiers' overall decision making behaviour. The goal of this seminar is to review works applying explainable AI to medical image analysis. Moreover, a broader view on the general state-of-the-art of XAI methods should be gathered, with the aim to relate it to the possible applicability on skin lesion analysis, in particular.
  • Geospatial Metadata (Julia Mayer)When working with many different large datasets we can’t get around a well-organized and established handling of metadata to ensure searchability, interchangeability and to keep an eye on various aspects of data quality. Here we will explore the standards, formats, common tools and problems when collecting data about datasets in a metadatabase, with a strong focus on geospatial data and
  • Methods of Spatial Decision Support (Julia Mayer)There are numerous examples of data-driven methods to support decisions in an urban context. Their nuances vary from data provision over meaningful visualizations to simulations. We want to create a collection with descriptions, pros, cons and best-practices.
  • Explainable Time Series for Earth Observation (Hiba Najjar)Forecasting tasks in Earth Observation rely on the time series of data collected by the satellites, and can be treated as either imagery or tabular data. To make use of recent advances in the field of deep learning (DL), there is an increasing use of neural networks for these remote sensing forecasting tasks. While the major downside of using complex architectures is there lack of transparency, the field of explainable artificial intelligence (XAI) proposes many methods which enables making the black box models more interpretable. This seminar is about investigating the XAI methods developed for the forecasting task, and identifying how can they be applied to remote sensing data, when the input is highly dimensional and/or when working with imagery data.
  • Explainable Multimodal Learning for Earth Observation (Hiba Najjar)Deep neural networks can ingest different modalities amongst language, vision, sensory and text in order to leverage their performance. Deep Multimodal learning call for complex architectures, which achieve outstanding results but are often limited by their black-box nature. Being able to understand the decision-making process of such networks increases their usability and social acceptance. To tackle this issue, here we will investigate explainable artificial intelligence (XAI) methods which are applicable to multimodal learning. We will focus on earth observation tasks, where the data is rich within its spatiotemporal modalities.
  • Graph Neural Networks in Recommender Systems (Mahta Bakhshizadeh)The main challenge in recommender systems (RSs) is to learn the efective user/item representations from their interactions and side information (if any). Recently, graph neural network (GNN) techniques have been widely utilized in RSs since most of the information in RSs essentially has graph structure and GNN has superiority in graph representation learning. This seminar topic aims to investigate the recent research eforts, challenges, and proposed approaches on GNN-based RSs.
  • Robustness of Machine Learning models to Missing Information (Francisco Mena)Machine learning (ML) models are techniques used to make predictions from input patterns to a task of interest, e.g. classify an image, based on having data to learn the task. However, the assumptions of these models usually imply that the input data is always correct, that it does not contain errors, and that it is complete, i.e. there is no missing information to make predictions. Based on this, an interesting research question has been explored in the literature: how to make ML models robust to errors and missing information in the data? This seminar will focus on exploring the proposals of the literature on this question.
  • Estimation of field scale soil moisture using remote sensing time series (Marcela Charfuelan)Soil moisture is a measurement of the water content of the soil, it can be measured with local (in situ) devices or derived from remote sensing data at various temporal and spatial resolutions. Estimation of soil moisture have applications in many fields, particular application examples are: weather prediction, precipitation estimation, flood risk modelling, irrigation assessment, crop yield estimation. Besides, according to the united nations, soil moisture, is one of the essential climate variables that critically contributes to the characterization of Earth's climate. Estimation of soil moisture using remote sensing is available nowadays at global scale, for some applications though, it is more useful soil moisture measurements at field level resolution (< 1Km). The objective of this seminar topic is to make a review of recent publications related to the estimation of soil moisture at field level, in particular using time series of remote sensing data at higher spatial and temporal resolution, such as Sentinel-2 and Sentinel-1 satellite data.
  • Temporal Knowledge Graphs (Marc Gänsler)In the last decade, Knowledge Graphs have developed into a major technology in the field of data representation. Knowledge Graphs enable the discovery of new correlations between data entries, and allow much more advanced queries compared to classical structures like relational databases. An interesting sub-type are Temporal Knowledge Graphs. Besides relations between entities, they also provide information regarding temporal aspects, e.g. at what time a certain relation was true. This enables new applications in the field of AI-driven knowledge discovery in large databases.
  • Harvesting employees' topics from unstructured sources (Heiko Maus)Although companies often have intranets with contact details of their employees as well as some means to identify their position, role, or tasks, searching for an adequate contact person is often futile if recent and everyday work topics and interests are required. Often there are no means to support such search tasks. However, there are sources which could help such as personal profiles, press releases, and wiki pages mentioning persons in meeting protocols, project descriptions, or fancy ideas. Looking at research, we find approaches for identifying skill sets from CVs, job descriptions, or social media profiles. In this seminar work, the student shall give an overview of different approaches for instance, based on knowledge graphs, word embeddings, building terminologies, and providing matches which could help for a company’s who-is-who.
  • Virtual Knowledge Graphs (Markus Schröder, pre-assigned)Data virtualization is a special data management technique to retrieve and manipulate data which exists virtually at query time. With this, technical details about data can be hidden from the user. In literature, virtualization is also applied in the context of knowledge graphs. Since data sources are usually not represented as knowledge graphs, virtualization can turn them into such data models. Because ontologies are used to represent virtual data, this method is also known as ontology-based data access (OBDA). The seminar paper should give an overview of various techniques in this area and work out their differences, advantages and disadvantages.
  • Driver Intention Prediction (Warda Khan)Is it possible to predict further in future? Recently published researche claims to predict driver’s intention with 98% accuracy up to 4.3 seconds into future. Using our in house system, currently, we have ca. 80% accuracy in around 5s into future. However, our system only considers limited scenarios, for example, changing lanes, turning left or right. We would like to expand this to other options as well i.e., driver is thinking to open the door, make an abrupt stop etc. A priori, it is not clear how the current state-of-the art concerning these technologies is. Thus, your task is to come up with a research/survey report on the two main questions: How many actions ofthe driver are possible to predict? How far into the future can the prediction be pushed according to the state-of.the-art? As the current solution is based on deep learning approaches we recommend to search in this area/community. A secondary line of questions is concerned with the input to the system. At the moment a camera only input is used. What does the state-of-the art say about that? Are observables like vehicle speed, steering angle, GPS info, etc. used there?
  • Image Super-Resolution with denoising diffusion probabilistic models (Brian Moser)A very recent trend of image synthesis is exploring denoising diffusion probabilistic models (DDPMs). The inference starts with pure Gaussian noise and iteratively refines the noisy output by using a convolutional neural network trained on denoising at various noise levels. This became very famous lately with the image generation application of Dall-E, where the DDPMs are conditioned on text input. In this review, we want to review recent advances and highlight possible future directions and current gaps in the field of DDPMs applied to Image Super-Resolution, starting with the paper "Image Super-Resolution via Iterative Refinement" (Saharia et. al 2021).
  • Recents Advances in Neural Architecture Search (Francisco Mena)Neural network (NN) models allow flexibility in their function design by connecting multiple modules in many different ways, making them state-of-the-art models for many problems. For example, many standard NN models are based on fully connected feed-forward networks, while others are based on recursive or convolutional operations, jumped connections, recursive layers, multi-scale, among others. When trying to solve a new problem with neural networks, one may need to search for the appropriate combination of components to build the best model for the data. This search for neural architecture could be very expensive in resources if all the possible combinations are tried (trained and tested). Therefore, alternative and ideally intelligent ways to search for the best model have been proposed in the literature, which will be studied in this seminar.

Topics marked as pre-assigned have been already assigned to students working with DFKI before.