Overview about ScaDS.AI Dresden / Leipzig

• Wolfgang E. Nagel (TU Dresden / ZIH + ScaDS.AI)

ScaDS.AI (Center for Scalable Data Analytics and Artificial Intelligence) Dresden/Leipzig is a center for Data Science, Artificial Intelligence and Big Data with locations in Dresden and Leipzig. It is one of five new AI centers in Germany funded under the federal government’s AI strategy. Based on the history as a Big Data competence center, we aim to close the gap between the efficient use of mass data, advanced AI methods, and knowledge management. In addition to new Machine Learning and AI methods, research topics on trust, privacy, transparency, minority protection, and traceability of AI-driven decisions are key. In my talk, I will highlight some of the mayor research directions of the center and address key aspects how to cope with the big data challenges in the age of AI. Furthermore, I will discuss how high performance computing (HPC) can contribute to further advance data intensive computing, including aspects of AI.

Artificial Intelligence in Biomedicine

• Fabian Theis (Computational Health Center, Helmholtz Munich)

In modern high-throughput biomedicine, huge and complex data sets are being generated, in particular from the omics and imaging fields. The analysis and integration of these data sets is daunting but crucial not only for research but also for envisioned clinical use e.g. for precision medicine. The vision of a truly data-based medicine is thus enabled by data science techniques, in particular machine learning and artificial intelligence. In this talk I will first review general concepts and challenges in the application domain. Then I will show some examples from the Center, in particular how to use machine learning for the statistical integration of omics data. I will quickly review how to order samples according to linear and tree-like latent similarities using pseudotemporal ordering, with applications in cell cycle classification and patient trajectories in diabetic retinopathy. I will then focus on applications to single cell genomics, asking how to generalize predictions to phenomena absent from training data i.e. out-of-sample. I will finish with an example on clinically relevant phenotypes, and put this into context of ongoing large-scale international endeavors.

Practical Uncertainty in Machine Learning

• Philipp Hennig (University Tübingen / TUEAI)

Abstract: Like any good scientist, a decent machine learning method should be able to estimate its own error. Such quantified uncertainty has many uses beyond the basic error bar: It provides the principled mechanisms to guide exploration and active learning, motivate and critique design choices, and trade off the utility of information from multiple sources. Probability Theory provides the universal and rigorous framework to quantify and manipulate uncertainty. The application of this formalism — Bayesian inference — has a reputation to be complicated and expensive. This tutorial will try to dispel this myth. Starting from basic examples we will get to know the Gaussian case a practically-minded workhorse of Bayesian inference, which maps the abstract notions of probability theory onto basic linear algebra. We will then see that modern automatic differentiation tools allow us to transfer this rich language to virtually all of modern machine learning. In particular, we will see how quantified uncertainty can be constructed simply in deep learning, at low computational and implementation overhead.

The new generation of world models - why your kids watch the same movie on repeat

• Jonas Andrulis (Aleph Alpha Founder & CEO)

For some months now a new generation of machine learning models amazes us with capabilities far beyond what AI used to be capable of. These systems are trained self-supervised - thy learn the structure of our world just by looking at it, without the need for human-generated labels. Is the generalizability and logical reasoning of these models a hint towards true Intelligence - or are we being fooled by so-called stochastic parrots? Where will this development lead us and what are its limits?

Get Together + Break-out Sessions

Legal challenges of the further development of AI – How do we sustain Meaningful Human Control?

• Susanne Beck (LL.M. (LSE), Leibniz University Hanover)

In the lecture, some of the most important challenges of the further development of learning systems from a legal perspective will be presented. One focus will be on the problem of diffusion of responsibility, which can be traced back in particular to the difficult attribution of actual events to an actor. Among other things, the speaker will present a project in which she is involved (vALID, funded by the Federal Ministry of Research and Education, https://www.gesundheitsforschung-bmbf. /de/valid-klinische-entscheidungsfindung-durch-kunstliche-intelligenz-ethische-rechtliche-und-10430.php). This project is based on the idea that only with such a control can one still speak of e.B. "actor" or "responsibility" in the traditional legal sense. Here, therefore, it is investigated how the responsible use of AI-controlled clinical decision support systems can be made possible in practice, i.e. how meaningful human control over these tools can be obtained. For this purpose, interviews with the participants are carried out in mock-up experiments and thus an attempt is made to find out under which circumstances they still have the feeling of control over the decision made. Furthermore, it is necessary to transfer these empirical findings into legal concepts, which will also be reported on in the lecture.

Knowledge Representation and Reasoning: From Foundations to Products

• Ian Horrocks (Oxford University )

In this talk I will review the development of commercially successful Knowledge Representation and Reasoning (KRR) systems and their genesis in foundational research. I will trace the evolution of KRR systems from logical and algorithmic foundations, through academic prototypes and standardisation to robust and scalable systems that power applications in areas as diverse as search, healthcare, financial services and manufacturing. I will discuss the barriers and milestones encountered along the journey, and lessons learned about the exploitation of research.

Project Presentations of the Schaufler Lab @ TU Dresden

• Sandra Mooshamer (Schaufler Lab)
• Gina Glock (Schaufler Lab)
• Kirsten Vinzenz (Schaufler Lab )
• Lutz Hagen (TU Dresden, Schaufler Lab)

With the Schaufler Lab@TU Dresden, TU Dresden and THE SCHAUFLER FOUNDATION have established a lively forum for a forward-looking dialogue between science, art and society. In this project, young researchers and artists come together across disciplinary boundaries to research on current technologies, their origins, and their impacts on culture and society from perspectives of humanities and social sciences. Some doctoral students of the first project are associates from ScaDS.AI. Fellows of the Schaufler Kolleg@TU Dresden are researching together with the artist of the Schaufler Residency@TU Dresden on the leading topic of “Artificial intelligence as a factor and consequence of social and cultural change”. The speakers of the Lab, Lutz Hagen and Kirsten Vincenz, will give a talk on the concept and development of the Schaufler Lab@TU Dresden. Also, two fellows will talk about their dissertation projects: Gina Glock: “Work in Times of Disruptive Technologies: Effects of Artificial Intelligence (AI) on Work Autonomy in Germany” and Sandra Mooshammer: “Abilities and Journalistic Quality of the Communicator Automated Journalism And Their Influence on Human Journalists' Perception of the Technology”.

130 years of “the neuron”: A brief history of an influential discovery

• David Kappel (Ruhr Universität Bochum)

Since Santiago Ramón y Cajal identified the neuron as the elementary functional building block of the brain in the late 19th century, neuroscience has evolved fast to grow into one of the most productive scientific research areas today. At a similar pace, research on artificial intelligence and machine learning has grown to become one of the main driving forces of technological innovation. From its early days on, the neuron has inspired biologists and engineers equally and neuroscience has emerged as a prime example of how technology and natural science can profit from sharing knowledge. In this talk, I will give a brief historical review of the main discoveries in the research on biological neurons and their technological counterparts from the past 130 years. In addition, I will highlight some more recent discoveries in the field and provide a perspective on future developments.

Bringing Modern Hardware, IoT Data Management, and Data Science together.

• Steffen Zeuch (DFKI)

In this talk, Steffen Zeuch will first present his past research activities in the fields of modern hardware, stream processing, and IoT data management. After that, he will outline his research vision of a general-purpose data management system (NebulaStream) that is capable to cope with and exploit recent changes in data characteristics, workloads, hardware capabilities, as well as computing infrastructures. With NebulaStream, we envision a platform to enable researchers and practitioners to develop and test their algorithms and approaches in the context of future IoT environments. On top of this platform, researchers from different domains like Machine Learning, Signal Processing, Complex-Event Processing, or Spatial Processing could implement their approaches. One possible direction for future research is to combine different parts of data science, e.g., data analytics, data mining, or machine learning, in one system to maximize sharing potential, utilize new optimization potential, and provide users a unified view among their data.

The Archived Web Dataset

• Helge Holzmann (Internet Archive (archive.org))

The Internet Archive (IA) has been archiving broad portions of the global web for 25 years. This historical dataset offers unparalleled insight into how the web has evolved over time. Part of this collecting effort has included the ability to support large-scale computational research efforts analyzing this enormous dataset. Web archives give us the opportunity to process the web as if it was a dataset, which can be searched, analyzed and studied, temporally as well as retrospectively. Our engineering efforts address the very specific traits of the archived web for our interdisciplinary users and partners, by hiding all the complexity and abstract away technical details. This talk will outline different perspectives on computational research of archived web data, along with technical challenges, novel developments and opportunities as well as considerations to make when working with this unique dataset.

Automated Algorithm Selection: Using Machine Learning for Efficient Optimization

• Pascal Kerschke (TU Dresden, ScaDS.AI Dresden/Leipzig)

Optimization is an integral part of our lives: engineers fine-tune the shape of mechanical components, businesses aim at minimizing their costs, space agencies like NASA wait for the best possible conditions to launch their space shuttles, and in our private lives we use navigation systems to find the fastest route between two places. Many of these problems can be formulated as an optimization problem, which can then be "solved" using a suitable optimization algorithm. However, usually there is no algorithm that is superior to all its competitors. Consequently, even the choice of the "right" algorithm (for the optimization of a given problem instance) is a challenging task, and the choice of the "wrong" optimization algorithm can severely affect the overall performance. In my presentation, I will introduce the general idea of automated algorithm selection and thereby demonstrate how machine learning can be used to efficiently solve such tasks. Subsequently, I will outline the current state of research for an exemplary optimization task and finally discuss the respective benefits, open challenges, and research perspectives.

Towards Trustworthy AI through Neural-Symbolic Unification

• Tiansi Dong (University Bonn and ML2R)

Symbolic reasoning is explainable, rigor, but brittle to noisy inputs. Deep learning overcomes these weaknesses at the cost of the explanability and the rigor of symbolic approach. In this talk, I show the possibility for a precise unification of symbolic structure and vector embedding. The unified representation inherits elegant features from both parent sides, namely, explanability, rigor, robust, and trustworthy. This shapes a new style of AI.