Diagrams take up a lot of space. In scientific publications, space is at a premium… so why do scientists use diagrams at all?

In this paper, I explore a possible relationship between how people are thinking and how they are communicating. This topic has been quite widely explored (e.g. Extended Mind Theory), but what is less done is examining a collection of diagrams to try to see patterns in how the diagram creators might be thinking.

One of the interesting aspects of this paper from a software perspective are the different abstraction levels which can be prioritised in diagrams:

• Function
• Data
• Mathematics
• Code

In enterprise software architectures, we might also have e.g. non-functional or commercial considerations. As noted in this post, this is (or should be!) about the user’s needs!

Exec summary

We might be able to get insights into how people are thinking, from the diagrams they create. (And maybe the prioritisations in the diagrams we create might also help shape how people interpret our systems.)

Happy New Year! Continuing the theme, I’ve another post about diagrams, more qualitative than the pre-Christmas quantitative citation count treat :).

For decades, Yuri Engelhardt and Clive Richards have been researching and teaching about graphics, especially graphical representations of data and information visualisations. I took the latest of their work, VisDNA, and applied it to some of the latest Neural Network diagrams of Computer Vision and Natural Language Processing systems.

In the paper and in the video below, I discuss a few examples and apply with VisDNA framework, suggesting some extensions to this “grammar” and using it to describe some of the ways scientists are communicating about complex systems using diagrams.

Exec summary

There are frameworks we can use to describe, reflect and critique on diagramming practices. See thediagramguy.com if you’re interested in finding out more

This one was an unexpected and pretty cool result… It turns out that, in scholarly AI publications, including 2-3 diagrams is correlated with higher citation counts after 3 years. The meta- thing here is that it is a bit odd that diagramming practices were found to be at all related to citation count.

There is another big block of work, which is the majority of my PhD thesis, which is about guidelines for NN architecture diagrams. It turns out that there is a correlation between compliance with >10/12 guidelines, and citation count. (This, of course, is only a small part of the evidence supporting the claim that the guidelines are useful – the majority being empirical user studies.)

There is quite a lot of information… I’ve distilled it in the video below, and also made all the code and data publicly available.

Exec summary

Diagrams are measurably important in science, and may capture aspects of good practice.

There was a really cool workshop at ICLR 2021, with the same title as this blog post, so I had to get involved! I wrote this paper about diagrammatic summaries for neural architectures.

Exec summary

The paper discusses some of the options for the scientific community for diagramming. The tl;dr; is that diagrams are used a lot but no-one thinks about it much. Sound familiar from business?! I’ve made a youtube video about the paper:

The main takeaway of my view on this is that we should start from where we are with our diagramming, and be a bit “agile” about how we move things. Scientific publishing is a strange beast, with large organisations and slow processes, but with really disruptive thinking going on by the “users”, especially in ML research. A very brief summary might be:

If anyone thinks slightly more about their diagramming, that is a good result.