In this talk I will give an overview of our contributions to interactive machine learning in a broad sense, i.e., not only including interactions between the machine learning algorithm and a human but also between the machine learning algorithm and some environment. Interactive machine learning goes beyond one-off model training and deployment by repeatedly adapting the model to new data. Interactions between machine learning algorithms and humans can be guided by the human--as is the case in interactive data visualisations--or they can be guided by the algorithm--such as the algorithm dynamically adjusting the difficulty settings of a computer game. In both cases, the runtime of the algorithm is key and often on top of polynomial time computability a different notion of efficiency is useful such as the responsiveness or the frame rate. In the case of interactive data visualisations it is also important that the types of interactions available and the flexibility of the visualisation are suitable. In the case of dynamic difficulty adjustments, it is also important to ensure that there is a limit on the number of times the algorithm picks a setting which is too difficult or too easy. Other interactions with the world are often in the form of cyclic discovery processes in which the algorithm proposes a new object (design phase), the object is realised (make phase) and evaluated (test phase) in its context, and the new data obtained in this way is incorporated to improve the model (analyse phase). For that, the machine learning algorithm needs to deal with the cyclic nature of these analyse-design-make-test processes, search a huge space of potential designs, and handle complex data structures describing the designs. Example application areas are drug design, game level design, and the optimisation of organic solar cells.

Thomas Gartner is currently Professor of Data Science at the University of Nottingham. Before that, he has been the head of a research group on `Computational Aspects of Mining and Learning', jointly hosted by the University of Bonn and Fraunhofer IAIS in Germany. His main research interest is machine learning in general and learning with structured in and output variables in particular. He has been an editor of the Machine Learning journal for several years and served as a senior programme committee member for several international flagship conferences. In 2010 he received an award in the Emmy Noether programme of the DFG and he co-organised the leading European conference in his field in varying roles in 2012, 2016, and 2018.