[Computer-go] Neural network -based symbolic integration outperforms Mathematica

[Juha Järvi]

There's a new paper from Facebook AI research:

https://arxiv.org/abs/1912.01412

8 GPUs and a few hours of training and the neural network can already solve symbolic integration tasks humans and Mathematica, Maple and Matlab cannot. Source: the author's tweet at

https://twitter.com/GuillaumeLample/status/1202178956063064064

Interesting if the same phenomenon appeared in mathematics as in Go: professionals studying results from an AI -based algorithm trying to understand new techniques it has discovered.

This seems like an interesting field similar to Go: huge set of possible inputs, lots of supposed creativity required to obtain correct result in complicated cases, hundreds of years of human research to reach the status quo, and a win condition (one of often many possible correct outputs in this case) that is easily checked.