Adding Reason to the Blackbox: Building Surrogate Models Based on Adaptive Abstractions
Interpretability allows to validate the correct behaviour of such models and may prove key for the widespread acceptance of machine learning . In this work a novel approach to add interpretability to the field of image classification is developed using surrogate modeling. This includes an image segmentation approach to transform image data into a suitable graph based representation. The human comprehensible representation is then used by a white box model to mimic the decisions of an arbitrary black box. Abstracting away the complexity of image data enables the white box model to reason about the image content in human understandable form. To assess the quality of the provided explanations an entropy based interpretability metric was developed. The developed system was evaluated using the Iris dataset [5, 29] and a self created image data set. The employed graph based representation proved powerful and flexible. Object occurrence, spatial relations and co-occurrence were spotted by the white box and could be visualised. The developed interpretability metric is stable and comparable. This thesis demonstrates that surrogate modeling can be applied to image classification. A powerful representation is necessary to deal with the complexity of image data. Graph based approaches are a suitable solution. Further an interpretability metric was developed allowing to assess the quality of the provided explanation. As this metric is promising it will be subject to future research.
Darmstadt, TU, Master Thesis, 2021