The CS department's Talk Series
![Talk](talk_Arndt.jpg)
Have you ever wondered what happens behind all these doors in our department?
Are you curious what is happening around you?
What are all these people busy with?
What research is emerging and which results are actually being produced in the CS department?
The CS department's Talk Series is a forum to researchers from different groups. It offers a great opportunity to learn about cutting-edge research, to talk to members of other research groups and to interconnect your research.
We especially want to invite all PhD students and research staff to the meetings, but interested master students are very welcome.
If you want to participate by giving a talk yourself, please contact us at promotionsprogramm@cs.uni-kl.de!
Schedule for Summer 2023
PhD Presentation Day: June 15 2023, 09:00 - 18:00
Thanks to a generous gift from FIT, we will have a workshop day at Blechhammer Hotel-Restaurant with the following presentations:Juraj Fulir (Inspection planning Virtual Inspection Planning Research Group, Image processing department, Fraunhofer ITWM): Inspecting industrial surfaces with synthetic data using computer graphics
Industrial surface inspection focuses on defect recognition in product surfaces for the manufacturing and maintenance industry. The task is complicated by requirements of high accuracy, throughput and robustness to offsets. Deep learning is a good choice for this task, but it is impeded by lack of training data as defects are rare occurrences. Synthetic data generated using computer graphics is a viable substitute to real data. However, some problems still need to be addressed to increase model robustness and reduce the biases embedded into the synthetic data generator.
Dinesh Krishna Natarajan (Smart Data and Knowledge Services, DFKI, Prof. Andreas Dengel): Hybrid AI: Towards efficient learning from physics and data
Hybrid Artificial Intelligence refers to a decision-making system that uses a combination of artificial intelligence and domain knowledge. In the field of computational science, the domain knowledge could comprise of numerical simulations of the underlying physical phenomena or governing physical laws that constrain the phenomena. The goal of this research is to develop techniques that efficiently combine modern deep learning methods with available domain knowledge. The level of hybridity of the combined method is specific to the use-case and a general methodology on choosing the hybridity will be an outcome of this research. The focus of this work is on applications in Computational Science (e.g.: fluid simulations) and Earth Observation involving physical phenomena (e.g.: flood mapping). In this presentation, a short overview of the use-cases, literature and state of the art, and the challenges in learning from simulation data will be discussed.
Daniel Theis (Embedded Systems, Prof. Schneider): Finding a Basis for Non-Sequential Endochronous Functions in Data-Flow Process Networks
In general data-flow process networks sequentiality and determinism can always be ensured by obeying the so-called Kahn restrictions. It is however also possible to define firing rules which still guarantee determinism by construction but are not sequential anymore. While the resulting class of functions can be further separated into hierarchic sub-classes depending on the degree of determinism, the endochronous functions occupy one of the higher levels which can make them still relevant for practical computations. To make this possible, it is essential to find a basis of elementary nodes so that each endochronous function can be represented by a composition of those nodes (similar to the sequential case). The current research is about quantifying conditions that need to hold for such a basis, so that finally a minimal set of nodes can be defined and proven to be complete.
Jakub Pawlak (AG Robotersysteme, Prof. Berns): Embodied Deep Learning
Deep Learning methods have evolved significantly across various data regimes in the field's history. From early approaches that utilized simple toy datasets to contemporary methods fueled by complex, multi-modal internet-scale data, the progression has been remarkable. This talk investigates the potential of leveraging robotics as a new frontier for data collection, specifically through the physical world. The combination of unsupervised learning with robotics may potentially increase the training data available by orders of magnitude. We delve into the possibilities of employing momentum-based unsupervised learning techniques for learning from interactions with the environment.
Alexander Günther (Software Engineering: Dependability, Prof. Liggesmeyer): Reliable Upper Bounds for Failure of Machine Learning Components: A model specific and model agnostic approach
In the last couple of years, machine learning was one of the most famous and celebrated areas in computer science. AI Methods achieved stunning and impressive results in a wide range of tasks and applications, even super-human performance. This raised the wish to also use this method in safety-critical areas, for example, medicine or autonomous driving. That raises the problem of verifying the safety and security of these models. But common techniques to verify software, like fault trees, Petri nets, equivalence class testing, state-based testing, and so on, are not applicable to machine learning implementations due to their black box nature. The goal of my thesis is to develop new methods and techniques to generate reliable statements about their safety. In the first year, I develop a statistical-based, upper-bound estimation for the failure on demand in the case of special binary classifiers. Furthermore, I am currently investigating the uncertainty quantification of the model risk. As a next step, I want to have a look at the deployment area inside and outside the system boundaries.
Deepak Kumar Pathak (Smart Data and Knowledge Services, Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI): Multimodal Machine Learning and its Challenges in Earth Observation
Multimodal machine learning has gained significant attention recently due to its ability to leverage multiple data sources to improve the model's performance. Remote sensing and earth observation data offer many problems where multimodal machine learning methods and approaches can be used, such as crop classification, disaster management, etc. This presentation reviews multimodal machine learning applications in crop yield prediction tasks, where multiple strategies are used to leverage information from complementary modalities, and discusses associated challenges. Finally, it discusses promising future research directions to explore unsupervised and self-supervised machine learning methods to reduce reliance on supervised data.
Nikolas Ebert (Mannheim University of Applied Sciences / Prof. Wasenmüller, Prof. Stricker): Transformer-based Detection of Microorganisms on High-Resolution Petri Dish Images
Many medical or pharmaceutical processes have strict guidelines regarding continuous hygiene monitoring. This often involves the labor-intensive task of manually counting microorganisms in Petri dishes by trained personnel. Automation attempts often struggle due to major challenges: significant scaling differences, low separation, low contrast, etc. To overcome these challenges, we introduce our PLG-ViT backbone in combination with our AttnPAFPN and a state-of-the-art object detector, resulting in a high-resolution detection pipeline that leverages our novel transformer variants. Our streamlined approach can be easily integrated into almost any multiscale object detection pipeline, e.g., Faster-RCNN and Tood. In extensive evaluations on various publicly available datasets, we demonstrate the superior accuracy of our network over the current state-of-the-art.
Miro Miranda Lorenz (DFKI SDS, Prof. Dengel): Domain-Informed Neural Networks for Earth Observation
Yield prediction is an essential task in the agricultural sector. Yet, it is still challenging due to multidimensional factors defining the yield, including environmental factors, management, the genotype, and their complex interactions. Considering changing and fluctuating climate conditions, reliable yield predictions are increasingly challenging. Here, machine learning plays an increasingly important role. Still, we observe weak performances and overconfident models. This colloquium provides an insight of how yield prediction can be addressed using stochastic and domain-informed models, highlights current challenges and future research directions.
Philipp Engler (AG Artificial Intelligence, Prof. Dengel): Approaches to Time Series Analysis using Data Augmentation Techniques and Unlabeled Data
Time Series data arises from sensors everywhere around us and in various industries. They can contain valuable information that may require manual inspection by an expert to be uncovered or contain structures that remain unrecognized by humans. Machine learning algorithms can help us uncover hidden information in time series data, which may be used for more precise medical diagnoses, increased efficiency of machines or various quality of life improvements in smart devices. For my PhD thesis, I aim to develop time series generation techniques to augment available data, as well as unsupervised learning techniques to leverage large amounts of unlabeled data in order to overcome the bottleneck of hand-annotating datasets, which often is one of the major obstacles for real-world application of deep neural networks.
Duway Nicolas Lesmes Leon (AG Artificial Intelligence, Prof. Dengel): Exploiting Generative models for life Science
Deep learning approaches have shown promising results in several tasks in life sciences. It has helped the experts in the field by optimizing processes, retrieving complex information that is not easily capturable, preprocessing bulk amounts of data, etc. However, such technologies require a considerable amount of data to achieve satisfactory results, and life science is a field well known for its limitations in data accessibility.
Generative models have shown potential in tasks such as synthetic data generation, data annotation, cross-domain translation, data enhancement, and feature extraction. The goal of this Ph.D. is to assess and exploit the potential of generative models (e.g., GANs) to mitigate the limitations in life sciences, particularly in microscopic imaging.
Francisco Mena (AG Artificial Intelligence, Prof. Dengel): Multi-view Fusion Learning in Remote Sensing Applications
In Remote Sensing (RS), the gathering of multiple observations coming from satellites and drones poses a multi-view learning (MVL) scenario for AI applications. This is a complex task when considering the significant differences in resolution, magnitude, and noise of RS data, since there exists several RS providers. The MVL has been faced with different fusion strategies in the literature, of which this talk will describe some. Since RS is a quite dynamic environment for AI applications where results depend on the data and the region applied, this talk will present some results obtained recently in two agricultural applications.
Schedule for Winter 2022/23
PhD Presentation Day: February 6 2023, 13:00 - 17:00
The following talks will be presented in-person in building 48, room 680.
Felix Givois (ITWM, Prof. Gauger): Quantum Computing for material characterization
During the last decades, the computation power needed for material characterization skyrocketed with the improvement of material tomography imaging. Unfortunately, the needs for accuracy seem to be limited by current classical hardware bottlenecks and especially by space and time complexity of current homogenization algorithms. In order to match the demand in characterization accuracy, the promises of quantum computing could be of a good use. As it theoretically reduces space and time complexity exponentially for problems such as prime number factorization, it could also overcome the limitations of classical hardware for homogenization algorithms. To this extent, the development of a quantum computing based material homogenization algorithm is investigated. The initial work has been to create a hybrid quantum Fourier transform based material homogenization algorithm. It has been done with the main aim of testing and benchmarking current state of the art quantum computers. The efforts are now focused on the porting of the entire homogenization algorithm on quantum computer. However, this porting raises new issues and as main one : how to implement non unitary operations on quantum computers ? My talk will outline the different researches conducted along the PhD as well as the different ideas followed in order to solve the difficulties encountered in the porting.
WeiChen Li (Machine Learning Group, Jun.-Prof. Dr. Sophie Fellenz): Using reinforcement learning playing text-based adventure games
Text-based games like Zork are ideal for testing language-based reinforcement learning (RL) agents. These games present players with a series of challenges that require a strong understanding of language to be successfully navigated. RL agents can learn to make informed decisions in these games by receiving rewards, which allows them to improve their language understanding and decision-making abilities over time through experience and feedback. In the future, we plan to focus on using reinforcement learning in more natural language processing applications.
Pakaj Deoli (AG Robotics, Prof. Karsten Berns): Effective environment perception for off-road autonomous navigation
An off-road environment is characterized with different objects, often ambiguous in nature. For autonomous navigation, precise perception of the environment is an important aspect. Methodologies involving Semantic, Instance or Panoptic segmentation provide the information needed on a broader level. When compared to urban environments, off-road environment (in itself) is a broad (and vague) term. A forest environment is different from a beach environment, which has different characteristics from a landfill area, etc and therefore the vehicle should have the capability to understand its appropriate environment and then take further actions. The goal of the research is to be able to teach this to a robot while also considering utmost safety as a priority. My inspiration for doing this also comes from the human brain wherein different parts of it are specialized in specific tasks and then bigger tasks in conjunction. Not only developing theoretical concepts but also implementing them on Unimog U5023 is also a big challenge for effective navigation through off-road environments.
Tahira Shehzadi (AG Augmented Vision, Prof. Stricker): Mask-Aware Semi-Supervised Object Detection in Floor Plans
Research has been growing on object detection using semi-supervised methods in past few years. We examine the intersection of these two areas for floor-plan objects to promote the research objective of detecting more accurate objects with less labeled data. The floor-plan objects include different furniture items with multiple types of the same class, and this high inter-class similarity impacts the performance of prior methods. In this paper, we present Mask R-CNN-based semi- supervised approach that provides pixel-to-pixel alignment to generate individual annotation masks for each class to mine the inter-class similarity. The semi-supervised approach has a student–teacher network that pulls information from the teacher network and feeds it to the student network. The teacher network uses unlabeled data to form pseudo-boxes, and the student network uses both label data with the pseudo boxes and labeled data as the ground truth for training. It learns representations of furniture items by combining labeled and label data. On the Mask R-CNN detector with ResNet- 101 backbone network, the proposed approach achieves a mAP of 98.8%, 99.7%, and 99.8% with only 1%, 5% and 10% labeled data, respectively. Our experiment affirms the efficiency of the proposed approach, as it outperforms the previous semi-supervised approaches using only 1% of the labels.
Michael Schulze (DFKI): Enabling Semantic Invoices and Knowledge-Graph-Based Services in Purchase-To-Pay Processes
Electronic invoices, which are part of purchase-to-pay processes, are increasingly adopted by enterprises and organizations. However, such data in the purchase-to-pay process domain can be very heterogenous, for example in terms of form but also in terms of vocabulary used by different stakeholders. Furthermore, in such inter-organizational processes, data is often fragmented in different data silos. Knowledges graphs can be a technology for harmonizing vocabulary used in such processes and in this way, they can be utilized for realizing assistants that are directly embedded into the workspace of a knowledge worker. This talk discusses on the one hand challenges and approaches for constructing knowledge graphs in the purchase-to-pay process domain, and it further discusses downstream applications to assist knowledge workers.
Monday January 30 2023, 17:00
Tobias Kattmann (AG Scientific Computing, Prof. Nico Gauger): Efficient Adjoint-Based Design Capability for Unsteady Conjugate Heat Transfer Problems
Shape optimization using gradients computed via the adjoint method has become a common feature and also, practically, a requirement among the major CFD and multi-physics solvers.
Accurate sensitivities are widespread available for single-zone steady-state problems, with unsteady or multi-physics adjoint solvers being less common.
Yet, gradient availability is a desirable feature for all primal simulation capabilities.
Many practical flows of industrial interest are unsteady in nature, and for the simulation of heating/cooling devices the coupling between a fluid and solid domain is essential to accurately capture the system's behavior.
Combining these two general observations of the demand for sensitivities and unsteady conjugate heat transfer applications, is the objective of the present work.
This talk presents the development and application of a computationally efficient unsteady discrete adjoint solver for conjugate heat transfer in the open-source solver SU2.
PhD Excursion to Schloss Dagstuhl – Leibniz Center for Informatics
The following talks will be presented in-person at a 2-day meeting at Schloss Dagstuhl – Leibniz Center for Informatics.
Eike Gassen (Robot Research Lab, Prof. Berns): A Service-Oriented Architecture on a Modular Agricultural Robot System
This work is based on the plan to design an expandable, (partially) autonomous agricultural robot architecture. According to the Federal Ministry of Food and Agriculture of Germany, the demographic development is causing an overall decline in the supply of labor in agriculture. This robot should make it possible to compensate the lack of human labor and also increase the cost efficiency of the farm. The system is intended to be available to the farmer as a year-round working tool through a modular design.
Michael Will (Scientific Visualization Lab, Prof. Christoph Garth): Distributed Parallel Feature Tracking Algorithms
Features play an important role in the visualization of simulation output data. While feature identification is a well-treated problem and the evolution of features in time-varying datasets has received some attention, the growing scale of computational science model output data necessitates new approaches in performing feature tracking on petascale and exascale architectures. My thesis will investigate and develop parallelization strategies for commonly used feature tracking methods and evaluate these using real-world use cases.
Saurabh Varshneya (Machine Learning Group, Prof. Kloft): Interpretable cross-modality interactions in multimodal data
With an enormous increase in the amount and complexity of data and problems available for a machine to tackle, learning from multimodal data is becoming an increasingly important area of research. Inspired by the success of deep learning, many neural-network-based methods are used to fuse information from multimodal data. Although these methods achieve excellent performance on the multimodal datasets, they are either black-box with limited interpretability or require complex architectures to fuse information. Deviating from a large body of recent works in multimodal learning, we adopt a generalization of the Multiple Kernel Learning (MKL) algorithm to fuse different modalities. Our algorithm opens a simple yet effective way to obtain the relative importance of individual modalities in the dataset. Furthermore, we extend our method's interpretability to compute the contribution of particular interactions among the modalities solve a machine learning problem.
Patrick Hansert (AG Database and Information Systems, Prof. Michel): Ameliorating data compression and query performance through cracked Parquet
In this talk, we propose to exploit synergy effects between partitioning and compression for Dremel-encoded nested data serving as the data storage for Spark-style processing jobs. The encoding proposed with Dremel has found widespread use in the form of open approaches like Apache Parquet, which can be used with a multitude of storage engines and processing frameworks, like Apache Spark. It stores the presence of objects in additional columns compressed using run-length encoding. Using partitioning, we can decrease the number of runs while at the same time using the partitions for data skipping. These effects can achieve a compression ratio of 1.37 while also reducing the query runtime by a factor of 1.22 in our test setup.
Hagen Heermann (Entwicklung Cyber-Physikalischer Systeme, Prof. Christoph Grimm): Runtime Verification of Hybrid Systems using Affine Arithmetic Decision Diagrams
My research conducted in the Cyber Physical Systems AG under Professor Grimms supervision focuses on runtime verification of hybrid systems. Hybrid systems are systems that continuously evolve over time and are controlled by a discrete controller that introduces discontinuous changes and discrete states. Examples of hybrid systems include medical equipment, manufacturing controllers, automotive controllers, and robots. The core problem for runtime verification of hybrid systems is the complexity introduced due to discrete and continuous behaviour and the uncertainties introduced by real systems. As the core idea of the runtime verification approach is the comparison of real measured trajectories of a system to the model of the system, the modeling of uncertainties plays a huge role due to the vastly different behaviour of hybrid systems with different initial states. Due to this, approaches, such as as Monte Carlo simulations, are not applicable. To tackle these challenges, my approach uses the data structure of affine arithmetic decision diagrams (AADD). AADD enables the representation of the continuous state space vector in different discrete states at the same point in time. This is achieved through a binary tree structure with internal nodes containing a linear constraint and the leaves containing affine forms.In comparison to typical approaches to runtime verification, my approach does not generate a monitor based on a temporal logic. Instead it is more closely related to reachability analysis tools such as CORA. This allows for a simpler generation of a monitors that can efficiently check if a measured trajectory adheres to the model. My approach will be demonstrated on two systems: a water tank and a Σ – Δ modulator.
Reinhard Leperlier (AG Embedded Intelligence, Prof. Lukowicz): Development and Evaluation of a Reinforcement Learning Model as an Enabler for Autonomous Blast Furnace
The Blast Furnace (BF) is equipped with thousands of sensors providing an enormous quantity of data enabling training of machine learning models.
Facing this large amount of time series measurements, the operator is not able to process manually those signals to decide in real time the best actions to perform for optimal thermal regulation of his BF.
The context of this research project is to study the application of machine learning/deep learning in order to reach the target of a self-regulating BF.
Attempts to build a simulator and to adapt state of the art Reinforcement Learning agents to the Blast Furnace context has led us to the conclusion that we need to focus on the definition and transformation of the data describing the current state of the Blast Furnace.
Brian Moser (Smart Data and Knowledge Services, DFKI, Prof. Dengel): Image Super-Resolution and its challenges
Image Super-Resolution is the process of generating High-Resolution images from Low-Resolution images. The application of Super-Resolution ranges from natural image reconstruction to highly advanced satellite imagery or medical imaging. Despite its long history, Super-Resolution remains a challenging task in Computer Vision. In this talk, I will present an introduction to deep learning applied to Super-Resolution and its current challenges, such as arbitrary scaling and high-frequency reconstruction.
Gajendra Doniparthi (Information Systems & Computational Biology, Prof. Deßloch): Multi-omics Data Management - Challenges and Solutions
With the advances in bio-science research, high throughput technologies, and big data processing, scientists now have access to large amounts of high-dimensional research data at a significantly reduced cost. The integrated analysis of these complex data sets provides opportunities to understand the biological systems and their relationships to create new and meaningful biological knowledge. However, the bio-science datasets are complex, and there are inherent challenges in handling the data. This talk outlines the research ideas we are working on to address those challenges and present our multi-omics data management solution.
Ahmed Tawfik Aboukhadra (AG Augmented Vision, DFKI): Two-Hands Object Reconstruction from RGB images using THOR-Net
Realistic reconstruction of two hands interacting with objects is a new and challenging problem that is essential for building personalized Virtual and Augmented Reality environments. Graph Convolutional networks (GCNs) allow for the preservation of the topologies of hands poses and shapes by modeling them as a graph. In this presentation, we describe our latest publication, "THOR-Net", which combines the power of GCNs, Transformer, and self-supervision to realistically reconstruct two hands and an object from a single RGB image.
Christoph Balada (AG Artificial Intelligence, Prof. Dengel): Machine Learning in the Smart Power Grid Domain and Graph-like Data Structures
Digitization already reached even conservative domains like distribution networks for electrical energy and down the road of data, many potential use cases arise. In the first part of my initial PhD talk, I want to give a brief overview of my first findings in the context of smart power grids and ML. The second part of my talk will give an overview of my future work and zooms out of this actual area of application, to shed light on the structure of the underlying data. Beyond the power grids, the data can be treated as graph data. However, current state-of-the-art approaches struggles with data like that of the power grids. Due to their multi-modal, timeseries nature, covering both node and edge features, graph neural networks may need to be rethought.
Ko Watanabe (Psybernetics Lab, Prof. Ishimaru / Prof. Dengel): Accelerating Knowledge Transfer by Sensing and Actuating Social-Cognitive States
Perceive, master, and transfer are key elements of the learning cycle. In previous studies, "perceive" and "master" has been further investigated, but there have been fewer studies on analyzing "transfer". In my PhD research, I am working on analyzing the transfer of knowledge from human to human. We are currently studying knowledge transfer using a variety of approaches. Visualization of knowledge acquisition procedures through visualization and analysis of user search information for the purpose of knowledge transfer from the past to the future. We are also conducting meeting analysis to facilitate real-time knowledge sharing. Our goal is to develop a system called TrackThink to visualize and propose search information for knowledge transfer utilizing past knowledge. We are also developing an online meeting system to improve knowledge transfer capability as a real-time approach. Using these two axes, we are working to develop a system that facilitates knowledge transfer.
Fabian Hartung (BASF Ludwigshafen / AG Machine Learning, Prof. Kloft): Time-series anomaly detection on chemical processes
Progress in anomaly detection on time series is currently slowed down to a large extent because there are only few benchmark data. Existing datasets often have the problem of triviality (anomaly always at the end or solvable by a one-liner of code) or bad/wrong labels of the data where it is anomalous. Starting with an evaluation of current data, my focus is on the generation and labeling of time series from chemical engineering and chemical processes in cooperation with BASF in Ludwigshafen and the research group of Prof. Marius Kloft.