5 comments

[ 4.8 ms ] story [ 14.6 ms ] thread
That is a pretty interesting work. Maybe a model derivation of this can be image (historic) restoration or new noise filters?
If this is your first time hearing about distill, they have so many more articles that are of equally astounding quality. I’ve been reading them for a while now and love what they do!
Why use differentiable models here, it seems like a big limitation to CA just to allow back-prop. Can't these be trained with genetic algorithms or evolutionary methods like SPSA?