This is really interesting!

I love the approach and am curious how they might best scale it up to bigger problems. Looks like they're working on hybrid models and I'm curious how those could leverage the advantages of neural models without losing the interpretability advantages here.