4 ms·
To me, it's an interesting question to ask how well NNs can perform without hand-holding, i.e. can they figure out all of the complex strategies that humans hav
by theptip 9y ago
To me, it's an interesting question to ask how well NNs can perform without hand-holding, i.e. can they figure out all of the complex strategies that humans have figured out, starting from a blank slate?
C.f. AlphaGo, which invented new strategies that proved interesting to human players.
Of course, if you're training your NN to solve a real-world problem, you wouldn't choose to train it in this way, you would do as you suggested and feed it as much relevant training data as possible, to give it a head-start.
(Your points RE: sizing and quantity of training data sound reasonable and are out of my expertise, so nothings to add there).
- PaulHoule 9y agoThere is a very simple strategy that solves all Sudoku puzzles. Pick one open square, try a number that is possible for that, backtrack if you get stuck. For better results, pick the open square which has the fewest number of possible choices. Human solvers might try something more efficient, but the above strategy is not at all bad for computer implementation. It's very much a solved problem with existing technology, see https://en.wikipedia.org/wiki/Satisfiability_modulo_theories https://en.wikipedia.org/wiki/Satisfiability_modulo_theories adding a neural net to it doesn't help in any way.
- andrewflnr 9y agoYeah, yeah, sudoku is easy for computers. That's not the point. The point is learning about neural networks. So yeah, "adding a neural net" does help with that.
- PaulHoule 9y agoIn which case the hard part is finding a suitable problem; you might be better off learning about SMT solvers because you don't have Marc Cuban telling everybody to do it.
- andrewflnr 9y agoWe don't really know if it's suitable until we try it.