Triage

[arvind]

High Priority

• [ ] Semantic dict with different bases/neuron groups (@arvind).
• [ ] Finalize running examples (@arvind, @colah).

Lower Priority

• [x] Finalize example picker UI -- sticky/hovering/embedded? (@shancarter).
• [x] Better colab stickers/captions (@shancarter).

Done

• [x] Neuron groups sankey diagram (@enjalot).
  • [x] Recompute data+visualizations to align with other diagram (@arvind).
• [x] Standalone example section w/model bias, picklehaube (@arvind, @colah).
• [x] More explicit sentence description of diagrams (e.g. SemanticDict) based on Emma Pierson's feedback.
• [x] Keyword blog draft (@colah).
• [x] Interactive teaser image (@arvind, @colah).
• [x] More technical detail in footnote/appendix about matrix factorization (@colah).
• [x] Explicitly discuss future work for each section (@colah).
• [x] Design space diagram + updated conclusion (@enjalot, @colah, @arvind, @shancarter, @ludwigschubert)
• [x] Create example-aware static diagrams (@arvind, @colah)
• [x] Label parts of first neuron activation equation (@arvind).
• [x] Fix NMF heatmap w/lightness mapping to magnitude of attribution and hues to the factors (@arvind).
  • [x] Try blurring gradients on attribution diagrams for smoother attribution (@colah, @arvind).
• [x] Improve labeling of channel attribution diagram selected state (@arvind).
• [x] Make activation grids bigger and fewer layers? (@arvind, @colah).

[colah]

I've done a little bit of gesturing at future work for each section, but it doesn't quite feel right to break the flow, and there isn't a huge amount to say. I think we should work 1-2 lines into our conclusion or the diagram space section andwe'll be good. :)

[colah]

I also added some footnotes giving a reasonable description of what we did for the two matrix factorization diagrams. In the final version, I'd also like to have notebooks demonstrating each.