Artificial, learning systems are increasingly taking on a leading role in various areas of life. However, the dynamic properties of the real world confront conventional systems of this kind with substantial difficulties. To address these challenges, Dr. Kalweit has developed a novel class of deep neural networks specifically designed to process set-structured data.
Abstract: In deep learning, neural networks are used to learn from experience and generalize to new, unseen data. In this process, complex representations are automatically generated directly from the input data and the time-consuming manual creation of representations is avoided. However, classical architectures are limited in their ability to handle more complex data structures such as images, sequences, graphs, or sets - limited by the number of parameters required and the amount of training data needed. To circumvent this limitation and accelerate learning, structural assumptions of the learning algorithm can be made and explicitly encoded in the architecture of the network. This dissertation focuses on deep neural networks specifically designed to process sets as a data structure. They have been extensively studied in the field of autonomous driving for decision making as well as in personalized medicine for applications such as detection, prediction, and clustering. It is shown that the presented architectures outperform classical deep neural networks and other machine learning methods. This work thus brings the field of deep learning one step closer to a flexible and easy-to-use tool that can be operated by non-specialist users - without demanding and lengthy development.
For this work she received the Wolfgang-Gentner-Award, handed over by the rector of the University Prof. Dr. Krieglstein on the occassion of the ceremonial opening of the academic year.
About the prizewinner: Maria Kalweit performed her work in the Neurorobotics Lab under the supervision of Prof. Joschka Bödecker within BrainLinks-BrainTools (https://nr.informatik.uni-freiburg.de/welcome).