From: Marc LeBrun <mlb@well.com>
Don't recall this getting discussed here: https://arxiv.org/pdf/1912.01412.pdf
Amazing results. My only gripe is that it's not clear how much processing time was used by the neural net, but they gave, e.g., Mathematica 30 seconds.
Interestingly treats symbolic integration as a kind of linguistic translation.
Well, they treat it in the same way people treat linguistic translation: with a sequence-in, sequence out interface. Here they encode both as forward polish strings. If you think of a neural net as like a matrix multiplication, you can do a state transition including an input and/or output in a single step. It's actually quite a bit like the different transforms of an IFS fractal. A friend of mine was pushing in and reading out balanced parenthesis strings (known to be not-absolutely-trivial) using a hand-designed 2D IFS around Y2K and I thought it was pointless. https://simondlevy.academic.wlu.edu/files/publications/bics2004.pdf
Leverages the cost asymmetry "trap door" versus differentiation when training.
Do you mean the way they generate problem-solution pairs by starting with the solution and differentiating to get the problem? (I don't remember which sage said that intelligence amounts to inverting functions.) --Steve