Artificial neural network models consist of many layers. How can we present a deep learning model architecture in a way that shows key features, while avoiding being too complex or repetitive? There is no standard for plots - neither for research nor didactic projects. It remains a challenge at the intersection of deep learning and data visualization.

Arguably the most important diagram of a neural network, AlexNet, consists of a cropped image of an otherwise wonderful diagram. I think we should aim for better, with usefulness and aestheticism in mind.

I will cover various aspects of neural network visualization:

• Complex blocks (e.g. LSTMs by Chris Olah and convolutions by Explained Visually)
• Architecture (with classic examples: VGG, Inception, ResNet, U-Net)
• Cost functions (in more complex cases e.g. YOLO for image detection)

I will show an overview of tools for visualizing whole networks and particular blocks, from very simple ones (e.g. my ASCII diagrams of sequential models in Keras) to complex (TensorBoard diagrams are verbose). Using existing visualizations (tools and manually created diagrams) I will discuss the following trade-offs:

• data viz vs data art (useful vs beautiful)
• explicit vs implicit (should I show ReLU all the time? But what about tensor dimensions?)
• shallow vs hierarchical (e.g. as in this detailed diagram)
• static (works well in publications) vs interactive (provides more information)
• specific vs general (does it work for a reasonably broad family of neural networks?)
• top to bottom, bottom to top, or left to right (conceptual metaphors related to space in action)

The author created an open source Python package keras_sequential_ascii converting sequential models in Keras into ASCII diagrams. He works on a general D3.js visualization of neural network diagrams.

Slides: https://www.dropbox.com/s/a7xako61ihuh82k/20180707_network_viz_pydata_berlin.pdf?dl=0