Johannes BrustApr 14, 2023
Mark Riedl

Food for thought. A proposal on how to draw neural networks better

https://medium.com/@amaster_37400/please-stop-drawing-neural-networks-wrong-ffd02b67ad77

I sort of agree with the danger of leaving the bias inputs out. The diagrams are pretty. I feel like some of the notation could be a bit subtle to the untrained eye.

Please Stop Drawing Neural Networks Wrong - Aaron Master - Medium

If you’re one of the millions of people who has tried to learn neural networks, odds are you’ve seen something like this: By which we mean confusing, incomplete, and probably wrong. The diagram…

Medium
Apr 13, 2023 at 3:17pmWeb
Show threadMatt Bauman Apr 13, 2023

@Riedl I still feel like we can do better than individual arrows to nodes — the confusion of arrows in a dense neural network are just noise. If I were redesigning the standard diagram, I'd say we treat each layer as a black box with N inputs and M outputs. Just use big fat lines with the number of values to represent each vector instead of a gazillion little ones. Something like:

=2=>◻=5=>◻=3=>

Two inputs to a layer that outputs five values to a second layer that results in three outputs.

Show threadMark RiedlApr 13, 2023
@mbauman Most DL researchers just diagram the computation graph (actually a simplification of the full computation graph). Depends on audience. It's a bit of a leap from all-the-connections to computation graph at first.