Project 4: How do neural networks decide?

A chief use case for neural networks is classification: Given an input the neural network predicts a class. A recent example is the automatic recognition of Covid-19 infections based on x-ray images of the lungs. Especially in a medical context it is essential that we can trust that the decisions of a neural network are valid. So, a pertinent question becomes: What are the features that lead to the prediction "ill" or "healthy"? On what basis does the neural network decide? Here, simple sensitivity analyses suffer from high level of noise that is typical in applications and are thus unsuitable. More modern methods (eg., saliency maps, integrated gradients or SmoothGrad), however, are quite sensitive with respect to user chosen parameters (eg., the baseline) and thus have their disadvantages, too.

First supervisor:

Prof. Dr. Henning Bruhn-Fujimoto, Institut für Optimierung und OR, Universität Ulm

 

Tandem partner:

Prof. Dr. Reinhold von Schwerin, Technische Hochschule Ulm

Prof. Dr. Karsten Urban, Institut für Numerische Mathematik, Universität Ulm

 

Consutling experts:

Prof. Dr. Jan Beyersmann, Institut für Statistik, Universität Ulm

Prof. Dr. Matthias Klier, Institut für Business Analytics, Universität Ulm

Prof. Dr. Michael Munz, Technische Hochschule