my favorite thing about the twitter cropping discussion is that there's an incredibly low barrier. almost everyone is running their own tests. even if we don't have "explainability", we still have folk experimentation.

in case you missed it, here is one thread from 2019 cropping out women in AI, while putting yann "it's just the dataset" lecun in the foreground https://twitter.com/AnimaAnandkumar/status/1307470107401170947

i think today's discussion started with this thread about zoom's background removal feature, inadvertently demonstrating a similar bias on twitter https://twitter.com/colinmadland/status/1307130447671984129

and then this is the one that exploded, showing a clear preference for mitch over obama. and showing a few variations that would inspire people to try their own experiments. https://twitter.com/bascule/status/1307440596668182528

twitter has a paper on their cropping algorithm here https://arxiv.org/pdf/1801.05787.pdf and google has a saliency model on github https://github.com/PAIR-code/saliency but it's hard to imagine a better intervention to come out of this than "get everyone posting variations to interrogate the system"