3 ms·
I mean cool but I don't see this as doing anything more than just memorizing answers? at least a cas performs transformations. also the time limit metric is art
by mlevental 7y ago
I mean cool but I don't see this as doing anything more than just memorizing answers? at least a cas performs transformations. also the time limit metric is artificially necessary for being able to perform the comparison in a statistically significantly way in a reasonable amount of time - an apples to apples comparison would be letting both systems run until completion and then compare accuracy but that would take too long because it is the case that algorithms that fit those transformations are often have combinatoric runtimes.
good idea to use the cas to generate though. I'm always trying to think up tasks that operate on procedurally generated data (e.g. inverse graphics) because then the labeled data is cheap.
- James_Henry 7y agoOne potential use for this, even with it just memorizing answers in a clever way, would be to speed up a CAS. For problems that are taking the CAS a minute or two, it is could be faster to have the model outlined in this paper propose solutions that the CAS simply has to check.
- eximius 7y agoI mean, at the very least I assume that they can generate a function, check that it wasn't in the training set, then solve it and verify. Then what would be your complaint?
- mlevental 7y agoI don't mean literally memorizing. "memorization" is tongue in cheek for overfit. I'm saying that the net won't actually generalize to classes it hasn't seen examples from (while reduction/transformation algorithms will).