I believe this question should be closed:

Test accuracy much higher than training accuracy

It's impossible to answer it without looking at the code, so the only reasonable suggestion I can give to the OP is to add code. However, once code is added, the question will become off-topic here and on-topic for Stack Overflow. Right now is not on-topic for SO, so I cannot propose it for migration. What reason for closure should I select?

PS I believe this case is quite different from a preceding, superficially similar question I already asked here

How to improve the following question

for the following reasons:

  • as Nick Cox noted, the other question is so complicated that even on SO it will probably be off-topic (modeling & coding issues are too intertwined). The current question, on the other hand, asks about a well-known, completely solved problem: handwritten digit classification from the MNIST data set.

  • the other user was a beginner at her/his first question, and the question was not bad for a first time. Here the user is not an expert for sure, but he's not at his first question either, and the question lacks a lot to be desired, so it should be closed IMO.

  • 2
    $\begingroup$ If you do not think there is enough information for even an expert to answer then I would be tempted to say it is Unclear what you are asking. $\endgroup$
    – mdewey
    Commented May 12, 2018 at 12:31
  • 2
    $\begingroup$ You might also find this Q&A interesting stats.meta.stackexchange.com/questions/4436/… as it raises a related question about whether the reason matters. $\endgroup$
    – mdewey
    Commented May 12, 2018 at 13:17
  • 1
    $\begingroup$ @mdewey, why not make that an official answer? $\endgroup$ Commented May 12, 2018 at 15:50
  • 4
    $\begingroup$ In reference to a premise in the question - a post containing code is not neccessarly off-topic. This is a common error made by a subset of close-voters. I feel our help center is pretty clear on the matter ("if it needs statistical expertise to understand or answer, ask it here; if it's about the implementation of an algorithm, routine data processing, or details of the language, then please refer to the collection of links to resources we maintain"). There's certainly room for disagreement about what counts as what for some of the conditions, but no doubt that posts may contain code. $\endgroup$
    – Glen_b
    Commented May 13, 2018 at 5:21
  • $\begingroup$ @Glen_b I think you misunderstood me. I never said that every post with code is off-topic. I believe that this specific question will become off-topic, once code is added, because the problem (handwritten digit classification on the MNIST dataset) is extremely well-studied, and we know that neural networks implemented in Tensorflow (the plots are clearly from TensorBoard) do well, the split in train/test set is predetermined, thus it's unlikely that the error is statistical. IMO the problem will be of implementation, thus off-topic here. Of course,I can't be sure. But I'm sure the 1/ $\endgroup$
    – DeltaIV
    Commented May 13, 2018 at 16:21
  • $\begingroup$ 2/ question is impossible to answer in its current form, and all existing answers concentrate on points that are totally irrelevant such as the train/test split, which for this dataset is given. $\endgroup$
    – DeltaIV
    Commented May 13, 2018 at 16:23

1 Answer 1


The answers in this Q&A Does unanimity of reasons to close matter? throw light on the related issue of whether the reasons for closure matter. The short answer is - they do not.

In the case of the question at issue here it seems from the comments that it falls into a common group of questions where the OP has found something unexpected and seeks enlightenment. This could be a failure of statistical understanding which is clearly on-topic or a problem with data or code. In the latter case it might be on-topic if statistical knowledge is required to answer.

  • $\begingroup$ I believe it's off-topic for the reasons in my comments above, anyway let's see what's happens when/if the OP adds code. $\endgroup$
    – DeltaIV
    Commented May 13, 2018 at 16:24
  • $\begingroup$ The idea that the reason given for closing doesn't matter at all seems too strong to me, & isn't my reading of the linked thread. I would say that some inconsistency is inevitable & it's OK if people get it wrong sometimes or give what they think is the best reason from several possibilities, even if that differs from others' votes. $\endgroup$ Commented May 14, 2018 at 12:58

You must log in to answer this question.

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