Our vision is to find solutions that enable us to build smart systems which provide a new quality of dealing flexibly with uncertainties and safely support learning, thus forming a new generation of smart devices. In order to achieve this goal, we address all of the challenges mentioned above by a unifying approach to tackle uncertainties and to provide learning systems that can be applied on-line even in uncertain environments to safety-critical applications like control systems in industrial automation, assistance systems for autonomous driving or in the field of medicine. All that requires much more than just parameter tuning.
On the one hand, we therefor develop methods for self-monitoring, self-protecting and self-guarding systems and introduced the Trust Management framework for a unified uncertainty treatment. It comprises representing uncertainties, may they be probabilistic or possibilistic in their nature, by scalar attributes – called trust signals - as well as architectures and algorithms. These methods exploit trust signals at the coarse-grained architectural level or incorporate them directly into the data processing algorithms at a fine-grained level in order to increase robustness and/or to reduce the impact of uncertainties.
On the other hand, we use machine learning for self-tuning, self-optimizing and self-adapting systems which do not compromise the reliability of the systems even if they face uncertainties or when shift and drift occur. This is done by extending learning algorithms by specific methods to control and to steer the learning process itself and its dynamics in the ongoing system operation.
We apply on-line learning systems which introduce design-time methods for deliberate control of the learning process. This is done on the one hand to tackle ignorance and ambiguity in an easy to handle way and to offer an a posteriori interpretability. On the other hand, this aims at justiciability for liability issues. Furthermore, we want to achieve this even in the extreme case where the learning has to be done from scratch while treating uncertainties, i.e. there is no prior knowledge about the precise parameters neither of the task nor of the environment the system has to deal with.
Our final vision is to obtain new technologies for real-world applications to support control systems, decision making, data analysis and other applications with and without real-time constraints. Here, in several aspects we are going beyond learning.