15
$\begingroup$

We have highly popular question, on extremely popular topic that has an outdated answers (most from 2010, newest ones from 2015): How to choose the number of hidden layers and nodes in a feedforward neural network? The most upvoted answers suggests to choose single hidden layer, while the recent achievements in deep learning show something opposite: the best results are achieved by networks with multiple hidden layers. Moreover, for some problems like image recognition, simple dense networks with single hidden layer are very far from state of the art. Since I'm not an expert in deep learning, I don't feel that I could give a detailed answer on that. On another hand, nobody did recently give an update. I myself offered a bounty for the update answer, but there was no response. What should we do in such cases? I guess this that we don't want to be a repository of outdated knowledge...

$\endgroup$
  • $\begingroup$ Also, why did you award the bounty to the accepted answer if you think it's so misleading? $\endgroup$ – amoeba says Reinstate Monica Jun 4 '18 at 7:06
  • $\begingroup$ @amoeba bounty time has passed, there was no new answers, the highest upvoted answer would give it anyway if I did not award it. $\endgroup$ – Tim Jun 4 '18 at 7:35
  • 5
    $\begingroup$ I am pretty sure that if there were no new answers during the bounty time, then the bounty would not be awarded to any answer automatically. The highest upvoted answer from the ones posted after the bounty was started automatically gets half of the bounty. The answers posted before the bounty was started play no role. $\endgroup$ – amoeba says Reinstate Monica Jun 4 '18 at 7:41
  • 2
    $\begingroup$ I think @amoeba is correct, at least on my reading of stats.stackexchange.com/help/bounty $\endgroup$ – mdewey Jun 4 '18 at 8:47
  • 1
    $\begingroup$ ...I don't think it changes anything in here, it is just about wasting the bounty, or giving it to outdated answer. $\endgroup$ – Tim Jun 4 '18 at 9:22
  • 2
    $\begingroup$ @Tim Kind of off-topic, don't know where else to post this -- often, having a question on the Bounty board for 7 days can stand as its own source of rep, aside from the rep awarded as part of the bounty itself, because being on the board raises the question's visibility. Ending a bounty at the earliest possible moment deprives the post of rep which it might have accrued as part of that process. $\endgroup$ – Reinstate Monica Jun 4 '18 at 15:33
16
$\begingroup$

While the canonical "StackExchange/StackOverflow" approach seems to be to allow downvotes, edits and new answers and hope the new good ones eventually float to the top, I think that in practice the best thing to do is make a new question (if possible, worded to be in as canonical form as possible), and enure it gets at least one decent answer. Then close the old one as a duplicate of the new one, adding a note that the answers there are now outdated and better answers are at the new one.

This should get to a desirable outcome faster.

Edit: I agree with amoeba that the standard approach should still be used if there's a reasonable chance of that working.

$\endgroup$
  • 7
    $\begingroup$ +1 Good advice. However, this should only be done when one is absolutely sure that the existing answers are all hopelessly unhelpful/outdated and when the chances that a new good answer would rise to the top are low. In this specific case I am not (yet?) convinced of the either. (1) There is a number of answers from 2015; a state-of-the-art answer from back then would certainly still be a good answer now in 2018. But I haven't read through all of them. (2) The OP is Rob Hyndman who is still active and could perhaps be convinced to change the accepted answer if one appears. $\endgroup$ – amoeba says Reinstate Monica Jun 4 '18 at 6:54
  • $\begingroup$ I think this is not going to (ironically) generalize to a lot of SE sites, so that's something that should be communicated to people if this becomes widespread practice here. I'm worried people will start marking-as-outdated questions or answers that become "incorrect" because the underlying technology has changed, rather than because our understanding of the facts has changed. (e.g. I wouldn't want a question about "how to print to stderr in Python" being marked as "outdated" because print >> sys.stderr, x should now be print(x, file=sys.stderr) in Python 3...) $\endgroup$ – Mehrdad Jun 14 '18 at 1:11
6
$\begingroup$

In this particular case, I would first talk to the most active and respected users in the tag (see the list of top users and the list of those who've been awarded tag badges). Of course not everybody there is still very active. I noticed that @DeltaIV, @Sycorax, and @JanKukacka have been posting good answers & comments in the NN tag lately, but I am not following this tag very closely.

Perhaps you could identify several such people and ping them to join this Meta discussion or a discussion in chat. If they agree that all current answers are suboptimal, encourage them to write a good (authoritative and comprehensive) answer. I can promise to put a large bounty on that thread once such an answer appears. Hopefully this would attract some upvotes too. Then we can reach out to @RobHyndman and encourage him to change the accepted answer. Profit.

$\endgroup$
  • $\begingroup$ Update: I have pinged three people I mentioned here. $\endgroup$ – amoeba says Reinstate Monica Jun 6 '18 at 10:33
  • 1
    $\begingroup$ I’ll take a look at the thread! This seems like a good idea. $\endgroup$ – Reinstate Monica Jun 6 '18 at 13:04
  • 1
    $\begingroup$ I'll have a look at the thread too. $\endgroup$ – DeltaIV Jun 6 '18 at 17:07
  • 1
    $\begingroup$ I have had a look at the thread, I was genuinely surprised that a question with highly upvoted answer got a bounty, but I did not realize the hidden message behind :) $\endgroup$ – Jan Kukacka Jun 6 '18 at 19:24
  • $\begingroup$ The accepted answer is definitely outdated: I will go through the other answers in the next days. Ps I note that @RobHyndman's question specifically concerns feedforward networks, thus RNNs are excluded. $\endgroup$ – DeltaIV Jun 6 '18 at 19:44
  • $\begingroup$ At what point will it be a good idea to simply post a new question, as per @Glen_b 's answer ? I am very interested in the current state of play regarding choosing hidden layers and nodes. I will rewrite the question if needs be.... $\endgroup$ – Robert Long Jun 15 '18 at 11:24

You must log in to answer this question.

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