Seminar: Learning Expressive First Order Rules – Introduction to Inductive Logic Programming

In order to overcome the opaqueness of machine learning models, different techniques can be applied. Among the prominent methods are Local Interpretable Model-agnostic Explanations (LIME) and Layer-wise Relevance Propagation (LRP). Both approaches create visualizations that serve as explanations for image classifications. Our research combines Inductive Logic Programming (ILP) with Deep Learning approaches to derive verbal explanations for the visual domain. In this talk, we first introduce the foundations of ILP, a logic-based interpretable machine learning approach. We explain how ILP can be used to generate global (model-level) and local (example-level) verbal explanations for classification tasks. In the second part of the talk, we present use cases taken from the research project TraMeExCo (Transparent Medical Expert Companion) and show how corrective expert feedback can be integrated into the process of explanation generation for medical image classification. We present results showing that the quality of explanations generated by an ILP algorithm can be adapted through constraint-based corrections. Furthermore, we illustrate that the presented approach can be used to identify noise and irrelevant features in the data set. We discuss new opportunities for model re-training and transparent machine learning that emerge by integrating ILP and Interactive Machine Learning into the process of image classification.