Journal Club (disc: Lee ... et. al ... :-D )

Should artificial intelligence be interpretable to humans?

Matthew D. Schwartz

Nature Reviews Physics (2022)

https://doi.org/10.1038/s42254-022-00538-z [pdf]

Online Discussion

Starts here: https://gitter.im/episphere/DCEG-DataScience?at=6372ad1315a3885c70ef865b

Notes:

• https://projector.tensorflow.org

• https://cacm.acm.org/magazines/2022/4/259398-explainable-ai/abstract

• Use for explained interpretations - they still allow us to explore latent spaces directly and will occasionally lead to discovery of "interpretable observables".

Arguments (just in case ;-) )

Keeping a glossary may help - feel welcome to argue for different terminology

• Explainable interpretation --> interpretation predicated on a human subject

• Pandora effect --> once "explained" you may not take it back.

• Projecting --> NN transfer of data onto canonical latent space.

Hackaton

App dev - from zero to PWA

...

There are 3 parts to it - apologies to Lorena and others who were promised this a long time ago ...

1. The DOM - and its markup language representation, HTML

2. The Style, CSS - present in each element and operated under its own tag.

3. JavaScript - the browser's interpretable language, also under its own tag, now breaking free with ES6 and WinterCG and a compile point, WASM.