Advances on interpretability of deep Neural Nets at ICIAM 2019

An introduction to different methods for Interpretability can be found here. During the ICIAM Theoretical advances of deep learning mini-symposia, there were some talks on interpretability, perhaps the most interesting ones were by Wojciech Samek, Fraunhofer Heinrich Hertz Institute, and by Stephan Waeldchen, Technische Universität Berlin. The first talk debated how LRP can be understood as a deep Taylor decomposition of the prediction. Some more information and tutorials on these can be found on their webpage. One of the methods to study the interpretability of a net is sensitivity analysis. For this, the changes of the gradient are used to decompose the neural net, however, the gradient is unreliable.