# Should a “Neural Network” Stackexchange site be created?

Currently, many questions related to artificial neural networks and deep learning are hosted in the CrossValidated site. However, it seems to me that both the quality and quantity of answers to these questions are far inferior (currently) to the ones related to statistics or general machine learning questions. IMHO, the reason for this could be that we are dealing with two different crowds: statiticians deal with statistics and general machine learning and engineers (and sometimes computer scientists and mathematicians) deal with neural networks and deep learning. Of course, nowadays, we all do everything as labor is (very) finite.

Would it not be the time to create a specific "Neural Network" site on StachExchange? How would one go about doing that?

• If you add this site do we create even more redundancy than we already have? A lot of times we get questions that should be migrated to other sites but the poor user is not sure which site is best. StackOverflow shoots a lot of questions our way, some fit us and some do not. Similarly we send them questions that we think are pure programming problems but some do have a statistical question and the one asking the question may be more interested in. – Michael R. Chernick Dec 31 '16 at 18:40
• I am a statistician who only has a passing knowledge of neural nets and I almost always avoid those questions. It seems that it would be nice not to be confronted with such questions. But how do we draw the line? I have done enough work on statistical classification problems to feel that I am somewhat of an expert at it. But classification methodology is used in machine learning and neural nets are also used for classification. It is an interesting question to raise and we should evaluare the advantages and drawbacks. – Michael R. Chernick Dec 31 '16 at 18:46
• What would be the benefit to further fragmenting users with ML and statistics knowledge? It seems like creating this website would create lots of duplication, both in questions and answers but also for users who just want to find an answer. OP's unstated premise is that making a dedicated website would improve question and answer quality, but I don't see any reason that should be the case: questions about NNs are on-topic here today. If the questions and answers are low-quality, that means we need more specialists, not that we should send the ones we have elsewhere. – Sycorax Jan 1 '17 at 15:10
• @Sycorax Maybe we should remove some redundant sites given your comments and Glen_b;'s answer. – Michael R. Chernick Jan 2 '17 at 2:23
• @Michael If there's some topics you wish to avoid seeing, you can set your account to ignore their tags. While I don't exclude any tags here$^\dagger$, I do exclude a whole lot on stackoverflow via ignored tags. It's quite helpful if that's what you want to do... $\qquad$ $\qquad$ $\dagger$ (I wanted to see everything here even before I was a moderator -- at least to see what was being talked about, even on topics I didn't know well) – Glen_b Jan 2 '17 at 2:43
• @Glen_b I am not sure where you got the idea that I avoid questions because of their tags. When I have time to go through a string of new questions I will often open all of them but if the terminology is foreign to me and I really know nothing about the subject I would feel that I have nothing to contribute and move on. There has got to be some form of time management. Also I would hate to have the responsibility of a moderator. It is a tough job and to do it well you have to spend a lot of time cleaning up questions and trying to decide what to do with poorly written questions. – Michael R. Chernick Jan 2 '17 at 5:22
• When is the right time to close? I can vote to close but be only one vote while moderators can take the responsibility of making the decision on there own. I really think Bill Huber is the best at it. He maintains good humor and patience and is always polite and shows respect to every OP. He tries hard to help fix questions that have promise and is a good judge of when it becomes hopeless. Yet even while doing that he sometimes will take to providing very elaborate answers that almost everyone find them to be masterpieces. he is able to move through the questions very efficiently. – Michael R. Chernick Jan 2 '17 at 5:32
• He spends so much time on CV that you would think he has nothing else to do but I know he has a consulting business that he also has to attend to and a family life and hobbies. I couldn't come close to doing that. He has touched my life in a very positive way and I am sure most members of CV feel that way too. – Michael R. Chernick Jan 2 '17 at 5:35
• @MichaelChernick I got the idea that you avoid neural net questions from your own words earlier in this thread where you said: "I am a statistician who only has a passing knowledge of neural nets and I almost always avoid those questions". You went on to say "it would be nice not to be confronted with such questions". I attempted to help you avoid being confronted with them. I was suggesting that you could more easily avoid them by using the features already built in to the site, which includes the ability to (after a fashion) ignore questions with given tags.. – Glen_b Jan 2 '17 at 6:54
• @Michael The issue of the right time to close can be asked as a question of its own. It's important that our high-reputation users contribute to the closure of questions and other moderation issues (as far as it is possible to do, which is actually quite a lot) -- the mods can't be here all the time. – Glen_b Jan 2 '17 at 7:01
• @Glen_b I think you misinterpret me a lot. You don't have to be so defensive. Moderator do what they can and have time for. I am a little in awe of how well they do their job. I help out when I can. Most of what I commented on was how much I am impressed with Bill Huber. All i would expect from anyone who would comment on that would be agreement; – Michael R. Chernick Jan 2 '17 at 13:56
• I don't think I was being defensive at all in my previous comment. I often take the opportunity to encourage the people who may be reading a discussion to help with moderation, but the intent was just to encourage all readers of our exchange to help moderate - while one vote (or even just the ability to flag) may not seem like a lot, there are a lot more people with one vote (or with the ability to flag) than there are moderators. I should have made it clearer that the @Michael was for the 1st sentence & the rest was a general comment. – Glen_b Jan 3 '17 at 3:26
• Good discussion. Redundancy is a problem and it does increase to burden to all involved. I was coming from an "ever increasing specialization" angle. I'm a MD/PhD student and we constantly joke that in the future we'll have right-hand doctors and left-hand ones. Knowledge is increasing faster than Moore's Law and, for better or for worse, specialization is inevitable and "resistance is futile". From what I could gather, the main problem ANN tries to solve is an optimization one, from a mathematical perspective with very little stats involved (though we also maximize LL to estimate parameters.) – Leonardo Cordeiro Jan 4 '17 at 21:11
• As such, the questions, and the expertise required to answer them, may be a little different. At least for now... Even though we all read the "no free lunch" theorem, ANNs, in all its flavours, seem to be the "winning" method in several ML competitions. Perhaps, with time, the number of experts in ANNs in CV will increase (or will migrate to it) and this point will become moot... Difficult to say, though. – Leonardo Cordeiro Jan 4 '17 at 21:14
• I know it would be very difficult to fit into the Q&A format, but when I run across an article like this I want to have somewhere to post it. This one is truly amazing.. Neural Translation of Musical Style imanmalik.com/cs/2017/06/05/neural-style.html – SDsolar Jun 10 '17 at 9:59

Should a “Neural Network” Stackexchange site be created?

This question has been answered by the response on Area51:

Would it not be the time to create a specific "Neural Network" site on StachExchange?

You are allowed to, but be prepared to explain how you will succeed where others have failed.

Should my idea be part of an existing site, or its own site?

"In general, if a site makes sense as part of a bigger site, it's better to have one big site than a bunch of little niche sites. Site X should be subsumed by site Y if:

• Almost all X questions are on-topic for site Y
• If Y already exists, it already has a tag for X, and nobody is complaining
• You're not creating such a big group that you don't have enough experts to answer all possible questions
• There's a high probability that users of site Y would enjoy seeing the occasional question about X"

How would one go about doing that?

See: Changes to the Area 51 Process v3.0 and the FAQ.

• This answer gives a nice perspective. Continouing on this line allows to give a less definitive answer about a seperate NN site; maybe we might better say that it is currently not yet the 'time' for such a sub-site (and leave open the possibilty that a separate site is possible in the future when there is a big enough independent community that can 'maintain' and 'fill' it). – Sextus Empiricus May 30 '18 at 18:14
• Robert addresses restarts in "Can this closed proposal be reopened or do I have to make a new proposal?": "... when you submit a proposal for a site, it is generally assumed you have access to a large enough audience ... [if] you now have a community to see this proposal through ... you will need to resubmit the proposal. But without a subtantial increase in actvity with a larger community organizing their efforts, the proposal will likely be closed again.". – Rob May 30 '18 at 21:47

Besides the questions on our site, there's already Computer science, Artificial Intelligence and Data Science (all of which have questions on neural networks - and that's not even counting the neural networks questions on StackOverflow and on Theoretical Computer Science ...).

I'm not sure that adding a fifth (or seventh ...) site fielding questions on the topic will necessarily be viable.

[Though I am not particularly knowledgeable about neural nets myself, I think it's important to - at least as far as feasible - keep statistics and machine learning (including neural-net) topics together. While emphasis and terminology differ, we're often solving a lot of the same problems. There's benefit in being exposed to what each group is doing (and has done in the past) and how we each think about problems.]

• Redundancy is sub-optimal indeed. AFAIU, ANN is "simply" trying to reverse engineer a complex function of which we have no idea of the correct functional form based on (lots of) empirical data. I know little about ANN myself, compared to my experience in econometrics and ML (in general), but I reckon that the problems one encounters when dealing with it are distinct to the ones we face in "standard" ML. I also concluded, perhaps too early and erroneously, that there is very little stats in ANN since it is, basically, an optimization problem. That is why I thought about the question. – Leonardo Cordeiro Jan 4 '17 at 20:57

Neural networks is simply one method. It is statistic in a sense that the neural network can be thought as a complicated function that is trained to fit the data in a best possible way.

As such it is not that academic approach as an engineering thing; things that work are used and people build intuition with experience on the matter. Thus it is hard to say anything about problem X, in case Y, if you have no experience. There is this thing called playing with data, that is supposed to give some intuition on the methods that could be applied

NN is the ultimate on that sense, that it is almost impossible to say anything too concrete about them because the number of layers / neurons per layer / NN-architecture all vary case by case.

Because of that, the NN gives an illusion that it would be something completely different, but it is not. It just happens to be so versatile that it has many applications. But in the end, it is just a method for regression. I recommend this article about common misconceptions: http://www.turingfinance.com/misconceptions-about-neural-networks/

• I agree with your second point; I also see ANN as an engineering thing. I came to this conclusion by looking at the research record and affiliation of all the major experts in ANN (most of them trained in Canada). A little of history here: recode.net/2015/7/15/11614684/…. I don't quite agree with the first point. I may be immeasurably wrong, but my impression is that there's very little statistical theory in ANNs. A lot of math applications (engineering) but little stats. – Leonardo Cordeiro Jan 4 '17 at 21:23
• Actually, the experts I mentioned above are known for their deep learning work, yet another subdivision of ANNs. – Leonardo Cordeiro Jan 4 '17 at 21:26