AI Feynman: A physics-inspired method for symbolic regression.
Udrescu, Silviu-Marian, and Max Tegmark. “AI Feynman: A physics-inspired method for symbolic regression.” Science Advances 6.16 (2020): eaay2631.
First of all, about the misleading title: This “AI Feynman” is really only 20% AI, at least in the neural networks sense of the word. The majority of the algorithm is really a carefully designed bag of tricks, and I don’t mean that in a bad way - it is a truly ingenious bag of tricks.
If you extrapolate the ideas here to an optimistic version of the future where this research has matured, science could look very different. Given the huge amounts of data being collected in the world every day, could we possibly just have discovery systems running on those huge datasets of the real world, and have it regularly point out analytical relationships for things we are interested in? Of course, we are in the early days yet, but I wonder: What laws of nature, or of human nature, are just sitting out there waiting to be discovered?
Caveat: The paper talks about how many natural laws follow these simple low-polynomial or simplifiable forms, and that is what allows this automated discovery to be feasible. However, I wonder if laws governing things like economics and social sciences will also honour this tradition? I hope so, but I suspect that beyond the fundamental equations of reality, it will be harder to find such neat equations in the human world that are accurate enough to be useful.
Anyway, I heard about this work when listening to Lex Fridman’s recent interview with Max Tegmark (this episode in particular, but the Lex Fridman podcast overall is highly recommended). In that interview, this idea was introduced not in the context of scientific discovery, but for simplifying the functions of neural networks into interpretable and robust formulae. To me, that goal is almost as exciting as discovering new scientific laws, because a future where neural networks are reduced to interpretable equations is a future where AI can be predictable and safe.
View in Google Slides.
Join the conversation on:
