Publication

Journal	2019 6th Swiss Conference on Data Science (SDS)
Page(s)	19 - 24
Title of proceedings	2019 6th Swiss Conference on Data Science (SDS)
DOI	10.1109/sds.2019.00-13

In this paper, we present novel visualization strategies for inspecting, displaying, browsing, comparing, and visualizing deep neural networks (DNN) and their internal state during training. Despite their broad use across many fields of application, deep learning techniques are still often referred to as "black boxes". Trying to get a better understanding of these models and how they work is a thriving field of research. To this end, we contribute with a visualization mechanism designed explicitly to enable simple and efficient introspection for deep neural networks. The mechanism processes, computes, and displays neurons activation during the training of a deep neural network. We furthermore demonstrate the usefulness of this visualization technique through different use cases: class similarity detection, hints for network pruning and adversarial attack detection. We implemented this mechanism in an open source tool called DNNViz, which is integrated into DeepDIVA, a highly-functional PyTorch framework for reproducible experiments.