When training a neural network, the general process could be something like this:

While error < min_error

  1. Forward pass
  2. Compute error and cost funcion
  3. Back propagation
  4. Update weights

But when we get out of the loop because the error is small enaugh, if we save the neural network as it is, we are saving something that we have not tested yet. Wouldn't it be better to keep the previous weight values, as those are the ones for which we know the error?

  • $\begingroup$ if you are doing the loop this way, then quite possibly yes $\endgroup$
    – user253751
    Feb 24, 2023 at 19:08
  • 1
    $\begingroup$ If you're not going to save the last set of weights, then there's no need to compute them in the first place. $\endgroup$ Feb 24, 2023 at 21:39


You must log in to answer this question.

Browse other questions tagged .