AI Algorithms and Methods are at the heart of Artificial Intelligence research and have dramatically impacted Data Science. The confluence of compute power, massive databases, and efficient algorithms to implement AI methods have impacted applications from engineering to the life sciences to social sciences. In spite of the success of AI across a wide range of applications, there remain many open problems and challenges. Interpretability, reliability, and uncertainty quantification in AI are requirements for many sensitive societal applications from biomedicine to decisions made by financial entities.

Topic Areas

Federated and Efficient Learning

There are two intertwined challenges to learning: learning algorithms must become more efficient in their use of resources, and in many applications, data transport to a central compute unit is often impractical or even prohibited by regulations. Both challenges are addressed by methods of federated and efficient learning.

Scalable Visual Computing

In the field of Scalable Visual Computing, we invent, improve and apply scalable Machine Learning and AI methods for generating, analyzing, and interacting with visual information.

Understanding Language

Language is typically viewed as the pinnacle of (human) intelligence. The seamlessness by which machines can be integrated with society depends on their understanding and mastery of language. Our research thus covers domain-specific large-scale language modeling, text manipulation algorithms, argumentation, and causal language, studied specifically in the context of conversational AI and connecting knowledge extraction and graphs with goal-driven dialogs, as well as in the context of mining the scientific literature.

Graph-based Artificial Intelligence

We significantly extend the research on graph analytics and the development of the distributed open-source system GRADOOP. We develop new indexing and graph summarization techniques and apply selectivity-enhancing mechanisms known from graph databases. Advanced representation learning in Knowledge Graphs will be integrated into the PyKEEN framework, and we explore knowledge-based methods for rich graph models.

Methods and Hardware for Neuro-Inspired Computing

We develop neuro-inspired hardware platforms for real-time AI applications, deployed as SpiNNcloud at TU Dresden. We also work on the integration with other computing hardware, like other neuro-inspired hardware systems or FPGA, and focus on unsupervised learning of spiking neural networks.

Knowledge Representation and Engineering

At ScaDS.AI Dresden/Leipzig, we carry out cutting-edge research in various aspects of knowledge representation and engineering such as ontology languages, with an emphasis on description logics and rule-based languages, ontology reasoning and algorithms, ontology-mediated access to large and incomplete data, both exact and approximate, knowledge graphs and Wikidata, non-monotonic reasoning, formal argumentation, explanation of the behavior of knowledge-based systems, and methods for dealing with heterogeneous and diverse knowledge.

Mathematical Foundations and Statistical Learning

We create, extend, and refine mathematical techniques in three major fields. We aim for a comprehensive, mathematically sound framework for learning transformation rules in abstract rewriting systems. Furthermore, we develop new methods for stochastic models addressing the often-observed problem of outliers in classification or stochastic time series, and we advance the field of learning theory in various respects.

Research Focus

Much of the modern work in AI Algorithms and Methods has not developed models that can be easily interpreted, cannot be manipulated via adversarial algorithms, and capture the uncertainty in their predictions. A major focus of our research will be to develop novel scalable algorithms and methods to develop AI that is interpretable, reliable, and captures uncertainty. Another challenge we address is that many algorithms that quantify uncertainty and have reliability guarantees for moderately sized data do not scale to massive data, the computational cost is too high. Using ideas from modern AI, numerical methods, databases, and data structures, we are developing innovative solutions that allow us to scale algorithms that work well for moderate size data to massive data.

We are also truly interested in the societal implications of AI and the influence of society on AI algorithms. Our research includes ethical and fairness aspects of AI Algorithms. A related problem that is an active area of research is the issue of preserving privacy. For many applications, we may want to preserve privacy, but often there is a cost to the accuracy of predictions when privacy is preserved, we are genuinely interested in this tradeoff and understanding when there is a tradeoff.

A recent application domain of AI has been to inverse issues and physics-based modeling. An inverse issue is to infer a data-generating model from a set of observations: for example, calculating an image in X-ray computed tomography, source reconstruction in acoustics, or calculating the density of the Earth from measurements of its gravity field. Inverse issues are an active area of research with applications ranging from climate change to monitoring manufacturing processes.