There's been some comment discussion as to whether a couple of questions e.g. this one and this one have been on topic.

In my opinion:

  1. We should take care not to readily dismiss technical questions as being 'programming related'.

  2. It's worth asking whether (even if the question mentions a specific technique) it could be answered with reference to open issues in AI.

For example, quite a number of questions (most of which have, in my opinion rightly, been left open without issue) are concerned with how to choose features for learning. In one respect, this is the single biggest issue facing AI: the current vogue for DL approaches is precisely because of the progress they claim in this area.

In particular: the data science community has not solved this problem - they are in general consumers of relatively stable research, rather than at the cutting edge, as is the case for AI.

Hence, we maybe shouldn't dismiss these things as implementation if they can usefully be treated conceptually.

Perhaps we can use "Is this a solved problem (in research terms)" as a heuristic to help us here. There's certainly precident for this: it is precisely the distinction between the 'Mathematics' and 'Math Overflow' SE sites.


2 Answers 2


I would say that it's a judgment call on a case by case basis. I don't think there's a simple rule you can implement that can capture all of the nuance involved here. My feeling is, unless you say with pretty close to absolute certainty that a question which includes code would get a better answer somewhere else, it's better to err on the side of leaving it alone.

That a question might contain math is, to me, nearly completely irrelevant to whether a question belongs here or not. Irrelevant in that it's orthogonal to the issue of whether something is "conceptual" or "implementation". After all, math is the language of science.

  • 2
    $\begingroup$ FWIW, I actually agree with you on 'no simple rule'. I completely agree with you on the maths issue. $\endgroup$ Aug 31, 2016 at 21:09
  • $\begingroup$ "I would say that it's a judgment call on a case by case basis." Hmm, this could hint that there is no clear distinction between concepts and implementation or otherwhise we wouldn't need a case by case basis. $\endgroup$ Sep 6, 2016 at 19:23

When I think of "implementation", things like math and code come to mind, while the larger components of AI construction don't fall under that category. Selecting features to build an AI for a certain purpose would therefore be on-topic, though they could easily be too broad. Your first example approaches "how do I solve this important problem with AI?", which possibly requires a deep knowledge of that field.

Questions tangentially related to programming, but not actually about the coding of the AI itself, are also OK. Your second example asks how to represent part of an AI's state for debugging visualization. It's a pretty neat question in my opinion, landing squarely in the science part of artificial intelligence.

I would be a little wary of allowing questions about the fine details (i.e. the mathematical/statistical mechanics) of yet-to-be-solved research problems, as those are likely to be much better served at one of the math-heavy sites. Conceptual questions about what kinds of things they work on are interesting and well-suited to our site.

Executive summary: if a question has mathematical formulae or computer code as critical elements, the best home for it is possibly a different site. This answer contains a lot of weasel words to emphasize that's it's not at all a rulebook that applies everywhere. Such an answer would be a tome.

  • 1
    $\begingroup$ I think you're over-generalizing when you say "if a question has mathematical formulae or computer code as critical elements, the best home for it is very possibly a different site.". Computer code is not computer code, if you know what I mean. That is, a question about "how do I add an element to a collection in Java" is quite different from a question about an AI specific technique, like, for example, something to do with Answer Set Programming using Prolog. The latter is probably not going to get much of an answer on datascience.se or stats.se, or possibly even on so.com. $\endgroup$
    – mindcrime Mod
    Aug 31, 2016 at 21:03
  • 1
    $\begingroup$ @mindcrime Right, this answer is general advice. As you say, judgment calls are frequently necessary. $\endgroup$
    – Ben N
    Aug 31, 2016 at 21:13
  • 4
    $\begingroup$ "yet-to-be-solved research problems [...] are likely to be much better served at one of the math sites". Maths per se does not hold the answer to the research problems facing AI - I claim that's conflating AI with the mechanisms of currently popular regression methods such as DL. Research problems in AI will be solved by devising better architectures, which is surely within the remit of this site. $\endgroup$ Aug 31, 2016 at 21:14
  • $\begingroup$ @NietzscheanAI Ah, my misunderstanding, I thought "research problems" referred to the math and stats of those techniques. Answer clarified. $\endgroup$
    – Ben N
    Aug 31, 2016 at 21:17
  • $\begingroup$ @BenN Thanks for the clarification. $\endgroup$ Aug 31, 2016 at 21:18

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .