Slany, EmanuelEmanuelSlany2025-05-132025-05-132025https://fis.uni-bamberg.de/handle/uniba/107376Dissertation, Otto-Friedrich-Universität Bamberg, 2025An inherent challenge of statistical machine learning models is the possibility of erroneous outcomes, which underscores the importance of corrigibility, particularly in sensitive domains. Advances in the field of explainable Artificial Intelligence have revealed that machine learning models might conduct correct decisions based on incorrect decision-making mechanisms. The possibility of detecting erroneous decision making does not prevent incorrect decision-making mechanisms, which is why the research field explanatory interactive machine learning has equipped researchers with the opportunity to revise explanations. Explanatory interactive machine learning procedures vary in their feedback injection mechanism: Some utilize model-specific internals, e.g., they penalize gradients in indecisive regions. Others are model-agnostic and induce additional training data, termed counterexamples, which only contain the relation of human explanation revisions and the target. CAIPI, the origin of the latter category, is an algorithmic framework to optimize machine learning models that iteratively combines components to predict, explain, and interact with instances. It generates counterexamples in iterations with correct prediction and erroneous decision-making mechanism. CAIPI, despite being claimed to be model-agnostic, has been proposed as a theoretical concept, which does not transfer well to practical application scenarios beyond image and text classification. Moreover, its formalization requires refinement, particularly when evaluating the impact of counterexamples on the optimization framework. This thesis aims to address CAIPI’s limitations through a formal analysis of how counterexamples influence the iterative optimization of machine learning models, together with algorithmic modifications of the original CAIPI framework.engExplanatory Interactive Machine LearningExplainable Artificial IntelligenceInteractive LearningMachine LearningCAIPI004doctoralthesisurn:nbn:de:bvb:473-irb-1073760