This is a question from a coursera course:

Suppose we have a set of examples and Brian comes in and duplicates every example, then randomly reorders the examples. We now have twice as many examples, but no more information about the problem than we had before. If we do not remove the duplicate entries, which one of the following methods will not be affected by this change, in terms of the computer time (time in seconds, for example) it takes to come close to convergence?

a) full-batch learning
b) online-learning where for every iteration we randomly pick a training case
c) mini-batch learning where for every iteration we randomly pick 100 training cases

The answer is b. But I wonder why c is wrong. Isn't online-learning a special case of mini-batch where each iteration contains only a single training case?


1 Answer 1


Isn't online-learning a special case of mini-batch where each iteration contains only a single training case?

This is true, but somewhat irrelevant (since the question is specifically comparing full batch to batch size 1 to batch size 100).

(b) will be absolutely unaffected by the change (modulo memory usage and cache efficiency issues), since each step costs the same as it would have before and is identical. (Well, depending on the formulation, a regularization constant might be effectively halved as well.)

(c) is affected because when we choose a batch of size 100, we might choose some points twice, overweighting those points and removing the other useful information that may have been in their place. Thus we have a slightly worse estimate of the training data distribution, and so will probably be slightly less effective in learning the model.

  • $\begingroup$ Thanks so much. I couldn't think about a case where a a datum is included twice, which is not the case in the original testset. Thanks! $\endgroup$
    – DSKim
    Jul 31, 2014 at 17:17
  • $\begingroup$ does this mean that technically, running learning on the whole data set is basically the same as on the doubled data set (from an information theory perspective). It makes things less efficient per round sure, but it doesn't really hurt learning from a statistical perspective. Right? $\endgroup$ Apr 27, 2016 at 21:37
  • $\begingroup$ @CharlieParker Yes, the doubled dataset contains the same information. If you train to convergence you should get the same model, again modulo considerations about the scale of the regularization – it just may take longer with the doubled one. $\endgroup$
    – Danica
    Apr 27, 2016 at 21:42
  • $\begingroup$ @Dougal I was reading the question again yesterday, and thought that (c) might not be affected with "sampling with replacement" but would be affected with "sampling without replacement" because the distribution over the samples would be unchanged in the case of "with replacement" but changed in "without replacement". Am I right? $\endgroup$
    – DSKim
    Mar 7, 2017 at 18:40
  • $\begingroup$ @DSKim Yes, that seems right to me. $\endgroup$
    – Danica
    Mar 7, 2017 at 19:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.