What is learning and what lies beyond? People learn a lifetime long, every day and thereby solve critical tasks, even in uncertain situations and ambiguous cases. Thus, we have come to expect similar capabilities of machines and technical systems in order to support us in our everyday life, may it be e.g. in car driver assistance systems, industrial plants and smart factories or medical data analysis.
As it is well understood in the fields of machine learning and computational intelligence to make a system learn from a set of data, it is still more challenging to learn from an ongoing stream of data in an incremental fashion. This becomes even more challenging if the learning system has to adapt to changing concepts due to shifts or drifts in some aspects of the task the system performs or of the environment the system works in.
But the greatest challenge for a learning system today is still the treatment of uncertainties during learning. Uncertainties arise in input data or learning data from different sources like measurement noise, outliers, insufficient specification, data sparsity, unforeseen interactions and so on. In on-line learning systems, uncertainties are even inherent in the learning system itself and vary dynamically over time. These uncertainties can drastically affect the behavior of the system and the learning process, and thus the application it is embedded in. Hence, it is very hard for a system to cope with uncertainties and always operate in a reliable and safe way - and to ensure this predictably right at design-time.