Skeletons
It's like teaching image recognition software with lots of pictures of cats and dogs, but then it's able to recognize elephants," said Shirley Ho, first author of the paper and a group leader at the Flatiron Institute. "Nobody knows how it does this, and it's a great mystery to be solved.
I must be missing something here, because the answer would seem obvious: the neural net is inferring the existence of skeletons. In the analogy provided, you are giving it lots of pictures of cats and dogs, and it "notices" that some bits are rigid and sized according to specific ratios, that others are "bendy" points, also located according to certain ratios, and it therefore infers underlying structure and rules governing movement. In the cats'n'dogs case, if you included data on their patterns of movement in specific environments, introducing the concept of behaviours toward goals (e.g. hunting by stealth in environments with abundant cover; scavenging in open regions where carrion/waste may be deposited), my wild-ass guess is that you could get the neural net to infer the approximate form factor for a "new" animal based on previously unimagined environments. From mice to cows. (It is intriguing to speculate what deep parallels there may be with how evolution actually works.)
Returning to the universes, it would seem that the neural net is inferring the existence of underlying rules despite not knowing their precise formulae. This is surely exactly what you wanted.
This leads me to ask: at what point can you inspect the neural net's results and ask "Why?" When does it become able to tell its human interlocutors "I have discovered the inverse square law for gravity"?
Because the neural net's next answer—some time later—may reveal rules/formulae that we didn't already know ...
Biting the hand that feeds IT
