News for hackers. Stuff that matters.
Re:Neural programmer better than Quicksort by YeGoblynQueenne on Friday July 10, 2020 @12:40PM (#23793338)
>> Like, if you give the algorithm that input for free what's the point.

The point, as also explained in the paper, is that even with such information and with a handful of basic operations from which to compose a program, the space of all possible programs is huge and finding a prorgam that does what it should in any reasonable amount of time is extremely difficult.

If you are familiar with the literature of program induction (which includes a lot more than neural program induction and goes back a few decades) you will notice that the hypothesis spaces for program learning are truly vast and it is impossible to make any kind of progress without some kind of "dirty trick" to reduce their size. The common approach is to impose a strong inductive bias on the representation language, so as to guide the search to find a correct program. The approach in the article is no different. And of course "inductive bias" here always means domain knowledge.

But, amicably, if you think there's any better way to do this, many people, myself included, would be very curious to hear what you have to say. In fact, if you could really innovate in this field you'd be guaranteed an illustrious research career. Learning programs from examples is hard.

There are no comments on this comment yet.