What we know about NN diagrams used in scholarly publications
This is based on Guy Marshall’s PhD thesis, and contains some high level comments.
- There are lots of tool to help make them, and people often use multiple different diagramming tools.
- Most diagrams use different visual encoding techniques, and they are challenging to group nicely.
- Diagrams are sometimes used for readers to “filter” a paper to decide it’s relevance, and sometimes used to allow the reader to instantiate an example.
- Using the guidelines results in a small positive change to existing diagrams.
- Compliance with the guidelines is correlated with a higher citation count.
- The guidelines may encode good practices which go beyond “only” being good design or good NN practices (according to experts).
A. Use conventional graphical objects where possible These are aesthetically preferred, and less likely to cause confusion
B. Only use one type of arrow for information flow This is less likely to cause confusion. Reserve different types of arrow for fundamentally different uses. Experiments suggest this is less useful than other guidelines
C. Use precision with care Using (for example) 4 of a thing will make some readers think there
are 4 of the thing and others 𝑛of the thing
D. Include the input and output of the whole system This helps make the overall purpose of the system clear
E. Consider using a single consistent example throughout This helps some readers to understand by instantiating the example and then generalising
F. Use visual encodings meaningfully When using a visual encoding principle, such as grouping by proximity or alignment, there should be a reason for it
G. Make navigation easy Ensure it is easy to navigate a path through the diagram. Labels for
layers, arrows, and linear alignment help to make navigation straightforward
H. Do not use colour for aesthetics If you use colour, it should indicate grouping, otherwise it can cause confusion Experiments suggest this is less useful than other guidelines
I. Use available conventions For example, if representing a CNN, it seems good to use the conventional 3D CNN format, and include all the filter widths numerically Experiments suggest this is less useful than other guidelines
J. Consider what people might expect to see For example, if representing a CNN, put pooling in as a step. If you don’t use pooling and that is important, consider noting that in a caption or label, as otherwise it may be assumed present
K. Be specific For example, “BERT” is better than “embedding”. This aids interpretability by avoiding obvious gaps
L. Consider that some readers may use the diagram without text For these readers, a relatively self-contained diagram is particularly helpful