Feature attributions, e.g. through a heatmap for images, are the prevalent approach to explain black box machine learning models. In Deep Learning, usually gradient information is utilized to construct such attributions. A recent survey [1] unified existing methods and compared them theoretically. In this thesis, different methods should be compared and benchmarked and possibly extended to improve current methods.

Keywords: Explainable AI, Deep Learning, Feature Attributions

Literature

1. Deng, Huiqi, et al. “Unifying Fourteen Post-hoc Attribution Methods with Taylor Interactions.” https://ieeexplore.ieee.org/abstract/document/10414149