Deep Learning for Symbolic Mathematics. Guillaume Lample, François Charton

[alex.shkotin]

FYI

This paper about ANN trained to integrate and differentiate symbolically is absolutely technical. But their conclusion on p. 12 awakes the idea that maybe they have a model of intuition:-)

"We show that a simple transformer model

trained on these datasets can perform extremely well both at computing function integrals, and

solving differential equations, outperforming state-of-the-art mathematical frameworks like Matlab or

Mathematica that rely on a large number of algorithms and heuristics, and a complex implementation

(Risch, 1970)." 

https://arxiv.org/abs/1912.01412

Alex