A standardised diagram-key system for detailing artificial neural network structures in scientific papers.
The aim is to create a consistent, modular diagram key for neural networks, enabling fast interpretation and universal communication in scientific papers. Currently, a wide repertoire of depictions are used in papers, resulting in difficulty for readers.
This outlined key system is designed to represent clear, symbolic placeholders for mathematical functions, much like Feynman diagrams for quantum field theory or logic gates for mathematical logic.
The number of keys remains minimal to prevent bloat, but there are keys for unspecified operations that can later be defined.
Chris Olah's depiction of an LSTM as a clear representation of its function inspired me to develop this system.
I am very much open to community suggestions on modifications to this system, and I will keep an open GitHub to centralize the system. No accreditation is required if used in any publication; this is entirely open-source. I've been using and improving this for a few years now and I personally find it an efficient way for precise communication. I hope you find it useful and see its potential too! :)
BibTex Optional Accreditation:
@misc{Bird_2023, title={Bird’s Convention for Diagramatic Neural Networks}, url={https://github.com/GeorgeBird1/Diagramatic-Neural-Networks}, journal={GitHub}, author={Bird, George}, year={2023}, month={Apr}}
A PNG (DiagramaticSystemSheet.png) and drawio file (DiagramaticSystemSheet.drawio) of the above image are available for download.
A PNG (DiagramaticSystemExamples.png) and drawio file (DiagramaticSystemExamples.drawio) of the above image are available for download.
More examples will be added soon and a hand-drawn shorthand jpg example is also included.