Logo of the Physikalisch-Technische Bundesanstalt

"Interpretable Deep Learning"

Kolloquium der Abteilung 8

Deep neural networks (DNNs) are reaching or even exceeding the human level on an increasing number of complex tasks. However, due to their complex non-linear structure, these models are usually applied in a black box manner, i.e., no information is provided about what exactly makes them arrive at their predictions. This lack of transparency is a major drawback when applying DNNs to the sciences. In my talk I will present a general technique, Layer-wise Relevance Propagation (LRP), for interpreting DNNs by explaining their predictions. I will demonstrate the effectivity of LRP when applied to various datatypes (images, text, audio, video, EEG/fMRI signals) and neural architectures (ConvNets, LSTMs), and will summarize what we have learned so far by peering inside these black boxes.