They say that their results are interpretable due to being logical formulas but, later, show a toy problem in which their formula has 10000 clauses. By that def

They say that their results are interpretable due to being logical formulas but, later, show a toy problem in which their formula has 10000 clauses. By that definition deep-learning is probably interpretable. However I do think that it is an interesting concept that would probably be worth testing if your inputs and outputs have a natural mapping to booleans.