My group studies the neural and cognitive mechanisms of episodic memory and spatial navigation. We employ neural network models, abstract cognitive models, and robotics for modeling and, in collaboration with experimental labs, apply advanced data analysis methods.

The optimization of adaptive systems workgroup is concerned with the design and analysis of adaptive information processing systems. We are interested in systems that improve over time through learning, self-adaptation, and evolution. Our systems improve autonomously based on data, in contrast to manual instruction or programming.

Currently wa are working on:

• fast training and model selection for support vector machines
• learning in deep networks
• development of new evolutionary search algorithms and their analysis
• open source implementation of a large number of machine learning algorithms

Our research in autonomous robotics is aimed at demonstrating that neural dynamic architectures of embodied cognition can generate object-oriented actions and simple forms of cognition. We organize the work around a scenario in which a partially autonomous robot system interacts with human operators with whom they share a natural environment. The robot system must acquire scene understanding to interpret user commands and autonomously perform actions such as orienting toward objects, retrieving them, possibly manipulating them and handing them over to the human operator. Based on analogies with how nervous systems generate motor behavior and simple forms of cognition, we use attractor dynamics and their instabilities at three levels to generate movement trajectories, to generate goal-directed sequences of behaviors, and to derive task-relevant perceptual representations that support goal-directed behavior.

The focus of my interdisciplinary research group lies on principles of self-organization in neural systems, ranging from artificial neural networks to the hippocampus. By bringing together machine learning and computational neuroscience we explore ways of extracting representations from data that are useful for goal-directed learning, such as reinforcement learning.

Our research operates at the intersection of Biology, Medicine and Pattern Recognition. In a multidisciplinary team involving
Biologists and Computer Scientists as well as Biochemsists, we work on questions in biomedical research that can be tackled using quantitative analysis of microscopic images. Our philosophy is that such biomedical reseach questions will be answered appropriately by algorithmic modeling, i.e., by translating them into dedicated, problem specific machine learning algorithms.

The Chair for Sys­tems Se­cu­ri­ty fo­cu­ses on dif­fe­rent to­pics in the area of ap­p­lied com­pu­ter se­cu­ri­ty. This re­se­arch field deals with the se­cu­ri­ty of de­ploy­ed sys­tems and the ana­ly­sis of re­al-world phe­no­me­na ob­ser­ved in the wild. Re­se­arch to­pics of the group in­clu­de for ex­amp­le soft­ware se­cu­ri­ty, bi­na­ry ana­ly­sis, vul­nerabi­li­ty as­sess­ment, pri­va­cy, se­cu­ri­ty of ma­chi­ne le­arning al­go­rith­ms, DNS se­cu­ri­ty, and si­mi­lar as­pects from the area of sys­tem and net­work se­cu­ri­ty.

We utilize machine learning for understanding time series signals, with applications centered (non-exclusively) on the transcription and analysis of speech and audio data. The underlying fundamental questions are how to achieve optimal cross-temporal and cross-modal information integration, when the reliability and information content varies over time and across sensory modalities. Algorithmically, this is achieved through hybrid, i.e. neural and probabilistic modeling, which allows us to combine the performance benefits of deep learning with the interpretability and reliability information of probabilistic models.

Intelligent methods are needed for process supervision and control of technological systems. This is the focus for the institute of Automation and Computer Control. Research is geared to new modeling, analysis and control methods with three focal points:

• Control theory
• Modeling and design of hybrid (mixed continuous-discrete) systems, and
• Process diagnosis and fault-tolerant control.

As part of our integral approach, we complement the theory with the design of methods that approach those limits. The circle comes to a close by experiments validating our results, such as:

• Universal Software Radio Peripheral (USRP)
• Raspberry Pi and Arduino

Our group works on novel techniques for representing and querying data and knowledge. As part of our research, we study the application of Machine Learning approaches to complete knowledge bases. For example, we leverage latent features from sub-symbolic representations of knowledge to uncover hidden patterns that can be used to predict missing assertions or properties in knowledge bases.

In a broad perspective, my main research ambition is to understand the fundamental computational principles of learning that characterize intelligence. More specifically, my research interests are focussed on the development, analysis, and application of deep learning models and methods. I am particularly interested in analysing and developing probabilistic models and inference methods, investigating biologically plausible deep learning, and understanding the stochastic processes involved in the training and optimisation of neural networks and probabilistic models.

Our research is about methods and theory for machine learning. Since recently, we have been working on statistical guarantees for deep learning. Our derivations are based on a wide spectrum of concepts from machine learning, statistics, and probability theory.

My research team considers and analyzase machine learning problems from the perspective of complexity theory. We try to determine the information or computational complexity of learning problems, and to design efficient learning algorithms. Furthermore, we investigate combinatorial optimization problems and explore the possibility of designing efficient approximation algorithms. We have a general interest in finding new methods for algorithm design and in inventing new data structures. We are furthermore interested in any results that shed more light on the famous P-NP problem.

Modern solid-state mechanics roughly covers the following areas: development of effective models for different materials, development and application of effective numerical methods to solve complex problems, development of accurate models for components and structures, and finally experimental verification of models and calculations, and attempts to determine the parameters of the models used. The Institute of Mechanics of Materials is dedicated to specific questions in all these areas. The focus here is on the modeling and simulation of different mechanical problems, regardless of whether the concrete issue now comes from the field of civil engineering, mechanical engineering, or even from other areas such as biomechanics. The chair conducts fundamental research on the theory and numerics of solid mechanics as well as cooperations with users from the academic and industrial sectors.

In our lab we apply machine learning (ML) and deep learning in particular to neuroscientific problems. We develop new algorithms to improve brain-computer interfaces and interpret data from non-invasive and invasive neural sources. Our aim is to utilize ML and AI in the clinical setting to raise the quality of life for the severely disabled. 

In our research we combine questions from theoretical and computational linguistics using a combination of methods from theoretical linguistics, experimental linguistics and computer linguistics (Linguistic Data Science). On these pages you will find information on current and most recent projects: questions, approaches and teams.
Additionally, you will find up to date interim results, publications, and announcements of upcoming presentations and publications.

Christian Straßer investigates logical aspects of defeasible reasoning. His research concerns theoretic foundations of non-monotonic logics and formal argumentation, as well as applications, including formal models of normative reasoning and agent-based models in the social epistemology.

The group works on the development of data-driven and physically-based modeling strategies and their application to analysis and interpretation of materials data. At the moment, one of key research topics is the development of physically-based and data-driven models to identify statistically sound correlations between materials chemistry, thermodynamic, microstructure and mechanical data of single crystal Ni- and Co-based super alloys. In comparison to the existing purely phenomenological modeling methods, such hybrid modeling strategies allow to identify the influence of individual physical effects from the considered contribution on selected material properties.

My group applies algorithms and methods from data sciences and machine learning to materials science problems. Our background is in Computational Materials Science. Specific applications currently include data fusion from simulation and experiment to improve interpretability, the development of interatomic potentials based on neural networks for alloys which are not well described by analytical formulations, and data mining in large defect simulations to improve the understanding of their collective behavior in order to inform coarse-grained models.

Our group works on natural language processing applications of language data in digital media. On the theoretical side, we build models for representing variable language online, in particular at the discourse level. On the application side, we develop methods for detecting and analysing negative communication practices in social media, for example hate speech and disinformation, using statistical and neural machine learning.

On the one hand we analyse data sets with biomedical / clinical research questions using clinical and / or lab data, possibly combined with quantitative high-throughput data from (Prote)Omics technologies. Depending on the research question, we usually search for biomarkers supporting diagnosis, therapy response, survival prediction etc. We also try to learn the group structure from the measured data (e.g. "healthy / diseased") and characterize specificity, sensitivity and ROC curves with respect to the predictability of assignment of newly presenting patients / samples to the learned groups. Thus we have methodically the usual practical problems concerning feature selection, avoidance of overfitting and (intra set or independent set)
validation.

On the other hand we try to characterize with deep learning the "dark matter of Proteomics", i.e. the usual phenomenon, that many recorded mass spectra - although looking  high-quality - cannot be identified with known amino acid sequences.