KI-Kolloquium der Universität zu Lübeck

Forschungsarbeiten zum Thema Künstliche Intelligenz bilden einen Schwerpunkt der Sektion Informatik/Technik in Lübeck.

Wir stellen Ihnen hier das Programm unseres KI-Kolloquiums vor.

Ort: Seminarraum Institut für Theoretischen Informatik, Geb. 64, Rm 2021, 17.00 st bis 18.00

Abstract for the presentation on Feb. 21st

Deep Machines That Know When They Do not Know

Prof. Dr. Kirstian Kersting, TU Darmstadt

Our minds make inferences that appear to go far beyond standard machine learning. Whereas people can learn richer representations and use them for a wider range of learning tasks, machine learning algorithms have been mainly employed in a stand-alone context, constructing a single function from a table of training examples. In this talk, I shall touch upon a view on machine learning, called probabilistic programming, that can help capturing these human learning aspects by combining high-level programming languages and probabilistic machine learning — the high-level language helps reducing the cost of modelling and probabilities help quantifying when a machine does not know something. Since probabilistic inference remains intractable, existing approaches leverage deep learning for inference. Instead of “going down the full neural road,” I shall argue to use sum-product networks, a deep but tractable architecture for probability distributions. This can speed up inference in probabilistic programs, as I shall illustrate for unsupervised science understanding, and even pave the way towards automating density estimation, making machine learning accessible to a broader audience of non-experts.

This talk is based on joint works with many people such as Carsten Binnig, Zoubin Ghahramani, Andreas Koch, Alejandro Molina, Sriraam Natarajan, Robert Peharz, Constantin Rothkopf, Thomas Schneider, Patrick Schramwoski, Xiaoting Shao, Karl Stelzner, Martin Trapp, Isabel Valera, Antonio Vergari, and Fabrizio Ventola.

Abstract for the presentation on Feb. 3rd

Probabilistic Artificial Intelligence for Autonomous Systems

Prof. Dr. Elmar Rückert

Artificial intelligence is regarded as one of the most groundbreaking developments of recent times. However, in controlling autonomous systems we are still far from achieving the human motor intelligence of a newborn or young child. In this talk I will discuss why current algorithms for autonomous systems and robot learning methods have not yet reached the required autonomy and performance needed to enter daily life. I will present current developments of biologically inspired decision models for autonomous systems and will discuss probabilistic prediction models that can be implemented in massively parallelizable neural networks. These neural networks are trained by a combination of supervised and unsupervised neuroinspired learning rules and enable complex decisions based on learned internal prediction models. The efficient learning rules allow the model to react to new environmental conditions within seconds and to process high dimensional tactile and visual data. These model properties are essential for adaptive, reliable, explainable and robust artificial systems.

Abstract for the presentation on Feb. 13th

Privacy Leakage of Machine Learning Models

Prof. Dr. Esfandiar Mohammadi

Training a machine learning model requires big amounts of training data. Yet, this training data can contain sensitive information about individuals. This talk explains how traces of the training data can be detected in an artificial neural network. Afterwards, the talk discusses how to protect individual training data points in neural networks and the challenges involved in protecting high-dimensional training data.

Abstract for the presentation on Dec. 2nd

Knowledge Graph Embeddings

Özgür Özcep, Institut für Informationssysteme, Universität Lübeck

The idea of embedding knowledge graphs into (low-dimensional) continuous spaces has been successfully instantiated within various algorithms and it has been applied to various problems in machine learning. Recent approaches either extend these ideas or come up with new ideas for embedding full-fledged ontologies defined in expressive logics. After an overview on classical and recent logic-oriented approaches to knowledge-graph embeddings I am going to discuss some ideas on embedding ontologies over languages that provide operators for negation, disjunction, and (restricted forms of) quantification.

Abstract for the presentation on Nov. 29th

Soft-Computing-Ansatz zur Parameteroptimierung von Convolutional Networks im Automotive-Kontext

Fabian Bormann, IAV Automotive Engineering (www.iav.com)

Die Entwicklung einer Architektur für Convolutional Networks erfolgt meist durch die Anpassung einer bestehenden Architektur gestützt von Erfahrungswerten und fachlicher Expertise. Trotz, oder auch aufgrund von, theoretischer Erfahrung durchläuft ein Architekturdesign oft mehrere manuelle Anpassungen und Evaluationsschritte. Die Wahl der Hyperparameter wird dabei meist ausschließlich mit den Ergebnissen einzelner Testläufe belegt. Entwickler achten beim Architekturdesign auf eine möglichst hohe Genauigkeit des Algorithmus bspw. hinsichtlich eines Testdatensets. In der Automobil-Industrie gibt es jedoch weitere Anforderungen, wie zum Beispiel die zeitliche Performance auf schwächeren Fahrzeugarchitekturen. Hierbei ist unter anderem die Größe der Architektur und die verwendeten Aktivierungsfunktionen ausschlaggebend. Ein möglicher Lösungsansatz zur Unterstützung bei der Architekturentwicklung könnte dabei ein Soft-Computing-Ansatz sein, der auf genetischen Algorithmen basiert und automatisiert anhand einer Fitness-Funktion bestehende Architekturen optimiert bzw. neue Architekturen hinsichtlich eines Problems entwirft. In diesem Vortrag soll dabei das Vorgehen vorgestellt und zur Diskussion gestellt werden.

Abstract for the presentation on Nov. 11th

Applications of Machine Learning in Biomedicine: From Basic Research to Clinical Translation

Lars Kaderali, Institute of Bioinformatics, University Medicine Greifswald

Artificial Intelligence (AI) and Machine Learning have the potential to fundamentally transform medicine. Applications range from basic biomedical research over improved diagnosis to therapeutic interventions. In some application domains, most notably in image-based diagnostics, algorithms are already outperforming human doctors, and are rapidly improving in performance also in other areas.

In my presentation, I will give several examples from our work at the University Hospital in Greifswald, showing how we employ machine learning and pattern recognition algorithms to improve medical data analysis and develop predictive models. I will show applications ranging from the development of predictive models from molecular data over the application of machine learning in the field of epidemiology to applications directly affecting treatment decisions at the patient’s bedsid