Can questions of programming AI/NN frameworks be on-topic?

Question

By AI programming, I mean somebody asks how to solve programming issue which deals with AI logic specifically, e.g.

• Why does my NN not classify these tic tac toe pattern correctly?

The above example uses Keras, highly modular neural networks library written in Python.

---

Are these off-topic and why they cannot be here?

If not, would be this ever considered? Especially the code which is very specific to AI coding?

---

Related: What kind of experts are we trying to attract?

In fact, I think now that implementation questions would benefit from a dedicated site (my view has evolved since the Area 51 definition phase). I have replied and tried to reply to several SO questions related to ML tools, and I think some are out of place compared to other questions. For example, some TensorFlow questions are not really programming questions, and not really framework questions. I mean, there is background knowledge on graph construction and execution, as well as background knowledge about statistics and probabilities that are really necessary to make meaningful contributions.

This is not to say that all questions are out of place on SO. Some are really framework issues or (Python) programming issues, and they are good there.

Based on this opinion, I think the site should be interested in implementation experts, whether they work on ML or Expert Systems (or both?).

-- @EricPlaton

---

So we're talking about coding highly modular neural networks libraries which require advanced background knowledge and AI expertise, and it was suggested.

The same as other specific modular frameworks, where coding questions are on-topic on their dedicated websites, they're allowed on: Drupal.SE, Wordpress.SE, TeX.SE, Apex at Salesforce.SE, etc. For a standard programmer without specific expertise, these are a bit of out-of-place on Stack Overflow.

Answer 1

In general, I think that we should allow questions on neural network architecture in order to attract and retain experts who actually build state of the art systems.

But for that specific question, I'm torn and lean towards keeping it closed. It's a novice instead of an expert architecture question; a good answer to it looks more like an explanation of how the necessary number of training examples scales with rule complexity, and how to ensure that the model depth and breadth is sufficient to encode rules of a certain complexity, and maybe also how to 'cheat' on model size and training requirements with convolution layers.

But if we want a question to expound on that sort of 101 material, it should probably be a set of three specific, easily searchable questions rather than the question that actually exists.

Answer 2

As far as I'm concerned, programming questions should be on-topic IF they are highly specific to AI. That is, if somebody asks "How do I add an item to a collection in Java", that would be off-topic even if they were building an AI application. But if somebody is doing something very specialized like Answer Set Programming in Prolog, etc., then I think this community would be best suited to answer that and, as such, the question should be considered on-topic.

Answer 3

Our site was created to hold questions about the scientific and social - not technical or programmatical - aspects of artificial intelligence. That's evidenced by its original section in Area 51, and by several posts by Stack Exchange staff. For example, see this answer on the question about summarizing our scope (relevant part reproduced here):

The proposal that created this site was intentionally placed in the 'scientific' category. If you accept that we are not creating another programming site, I think we stumbled upon in interesting niche that describes the original premise of this site nicely:

Artificial Intelligence Stack Exchange is a site with a social and scientific focus on "Advanced Computing in Society."

Think about it. With autonomous cars, smart surveillance, and "the next big thing" capturing the headlines, this isn't a terrible idea for a subject. Draping it in the popular AI label gives it a better focus… and it completely disambiguate [sic] that this is not a technical implementation or programming site. We already have that.

The emphasis was in the original.

Especially when we actually have a pretty cool scope draft (as summarized there), I don't think we should add an extra place in the network for AI-related programming questions.

Answer 4

There are already plenty of questions on programming AI/NN frameworks on Data science and SO: I don't think we need one more place, SE is fragmented enough.