I got this feedback for my thesis paper. Can anyone explain this feedback with details?

  • 3
    Why don't you ask your advisor? Frankly, if I knew a bit more about your thesis, I'd expect that the feedback would be very clear to me. I can almost guess what it refers to with the meager information you provide. I'm astonished that it isn't clear to you.
    – Roland
    Jul 5 at 6:38
  • I just did my thesis alone and it was my undergrad thesis. And I submit the conference alone and gave this feedback from them. So that I need help from you who expert in this field. Jul 5 at 6:46

This sounds like machine learning, where the result of training a model (often neural networks trained with gradient descent) are known to depend on the random initialization of the network parameters. A quick google search will help you find much more details.

Just re-run the training multiple times with different random seeds (you should manually set them so you can be sure they are different) and report what happens as suggested.

  • Thanks a lot for your answer. "There should be multiple runs with means and standard deviations reported." - is that means to run with different random seeds? Jul 5 at 6:54
  • 1
    Yes. The seed changes the initialization.
    – cheersmate
    Jul 5 at 6:55
  • Thank you. Instead of run different random seeds if I run with k-fold cross-validation. Is this the same then? Jul 5 at 6:57
  • 1
    No, but you should ask this on another SE site like stats or ai.
    – cheersmate
    Jul 5 at 7:42

More context would be helpful. It is possible that your result is simulation-based (such as a bootstrap CI, simulated permutation test, or simulated P-value for a chi-squared test with sparse cell counts). Then the referee may wonder whether results are affected by your particular simulation.

In that case, you might (a) show the seed (and software) for the simulation, (b) estimate simulation error, (c) show results of two or three additional simulations, and/or (d) if feasible, use a larger number of iterations in the simulation to make simulation error smaller.

In any simulation, it is good practice to include at least (a) along with (b) or (c).

  • And the improvements shown in the results section could be a result of random initialization. There should be multiple runs with means and standard deviations reported. - the full review. In the title can't put full review for the word limit. Jul 5 at 5:54
  • There should be multiple runs with means and standard deviations reported. - This context is the main confusion for me. Jul 5 at 5:55
  • 1
    By context, I mean what are you trying to show, why are you using simulation, what do you suppose the referee means by 'random initialization', and how do you know how good your simulation is?
    – BruceET
    Jul 5 at 5:58

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