Mark Rothermel, Computer Science Bachelor Thesis, Aug 2019, TU Darmstadt.
Sum-Product Networks (SPNs) (Poon et al., 2012) are graphical models used in Deep Machine Learning for probabilistic modeling. SPNs lack explainability. Influence Functions (IFs) (Koh et al., 2017) are a mathematical concept which can be used to investigate the effect of a training example on the model's parameters and, in turn, the model's performance on a reference (test) example, revealing the local learning behavior of the model, which is helpful for understanding the model and the reasons for its predictions.
This work mainly uses the SPN library SPFlow, Tensorflow, and the implementation of IFs from Koh et al.'s repository influence-release.
For a full tutorial see .
For the full-text bachelor thesis see .