Net2Vis: Transforming Deep Convolutional Networks into Publication-Ready Visualizations

To properly convey neural network architectures in publications, appropriate visualization techniques are of great importance. While most current deep learning papers contain such visualizations, these are usually handcrafted, which results in a lack of a common visual grammar, as well as a significant time investment. Since these visualizations are often crafted just before publication, they are also prone to contain errors, might deviate from the actual architecture, and are sometimes ambiguous to interpret. Current automatic network visualization toolkits focus on debugging the network itself, and are therefore not ideal for generating publication-ready visualization, as they cater a different level of detail. Therefore, we present an approach to automate this process by translating network architectures specified in Python, into publication-ready network visualizations that can directly be embedded into any publication. To improve the readability of these visualizations, and in order to make them comparable, the generated visualizations obey to a visual grammar, which we have derived based on the analysis of existing network visualizations. Besides carefully crafted visual encodings, our grammar also incorporates abstraction through layer accumulation, as it is often done to reduce the complexity of the network architecture to be communicated. Thus, our approach not only reduces the time needed to generate publication-ready network visualizations, but also enables a unified and unambiguous visualization design.