Technik- und Ingenieurwissenschaften
How can we develop systems that can use already acquired knowledge in new situations? My name is Pascal Klink and I have been working on this question for more than one and a half years as part of my PhD studies at TU Darmstadt. Before that, I completed a B.Sc. in Computer Science and M.Sc. in Autonomous Systems at TU Darmstadt, with a semester abroad at UBC Vancouver. During the course of the last years, research results in the domain of machine learning impressively demonstrated that we can develop virtual and physical systems that learn tasks autonomously. Nevertheless, there are hurdles that impede the widespread use of such systems. Learning a task is often time consuming and what is learned is not reusable. For example, a robot that has learned to assemble a workpiece from basic components must be "relearned" from scratch for a similar, non-identical workpiece. We humans function differently. When learning new skills, we build on what we know. As a result, we can solve more complex tasks as we gain experience. I would like to realize this property in AI systems, with a special focus on robotics. In this way, AI can be used by more companies in a future working world, as systems can be more easily adapted to their tasks.
Realizing aforementioned knowledge reuse and -transfer in AI systems raises a lot of challenging research questions in a variety of research domains. I particularly focus on the domains of Reinforcement Learning, Transfer Learning as well as Bayesian Inference.