Short Description of Research Topics
Main focus of research in the Cogsys Group are algorithmic approaches for interpretable machine learning. Since more than 15 years we are developing white-box machine learning approaches for relational domains, mainly inductive programming and structure generalisation. Furthermore, we are researching cognitive aspects of machine learning as well as the effects of cognitive companions on learning and trust. Currently, we are working on explanation generation in the context of cognitive companions based on interactive learning.
see: Projects -- Advised Theses -- Publications
Inductive Programming
The construction of generalized (recursive) programs from small sets of positive input/output examples is investigated. More specific, our research addresses
- incremental approaches to inductive logic programming (ILP)
- explanation generation from ILP learned rules
- combining ILP with neural network classifiers for image data to explain classifications based on relational characteristics
- theoretic and algorithmic foundations of the induction of functional programs,
- inductive program synthesis as an approach to cognitive modeling of learning from problem solving experience, and
- application of inductive synthesis techniques to enduser programming support (e.g. inducing XSL transformations with recursive template applications from small example XML documents).
Learning Structural Prototypes
Development and application of structural generalisation methods (least general generalisation, anti-unification)
- for incident mining – to provide standard solutions or to retrieve previous solutions for incident reports as support for help desk engineers, and
- as a cognitive model of adaptation effects in aesthetic judgements.
Analogical Problem Solving and Generalization
Several psychological as well as formal aspects of analogy making are explored:
- anti-unification as approach to automated analogical reasoning and generalization
- empirical investigation of conditions for preference of derivational (replaying an old solution) vs. transformational (mapping entities of base and target problem) strategies in human problem solvers
- empirical demonstration of re-representation of problem structures during analogy making
Applications of Planning and Learning
Current research topics are:
- individualized pain classification from facial expressions
- distributed problem solving in rescue and emergency scenarios
- classifier learning for medical and technical diagnosis
- cognitive assistance for people with mental deficits