I'm doing some machine learning where I have lots of data and through optimization I'm trying to learn the weights for the model.

I'd like to check that my learning actually works correctly. For that I can create a set of known weights. My question is from these weights, how can I create synthetic data? On this data I can perform the training and check if the weights in the end are the same as the weights I created.

  • $\begingroup$ Your question is slightly confusing. Are you saying you have response data and know the form of the model, and are trying to fit model parameters? If you have a model, you should be able to generate "realistic" parameters and simulate its response. The process to do this would depend greatly on the model type. $\endgroup$ – Godric Seer Apr 11 '13 at 21:03
  • $\begingroup$ @GodricSeer What I'd like to verify is that my learning is able to find the optimum parameters. How can I test this? I'm not talking about testing the accuracy of the model, but rather that the learning itself indeed works. The only way I can think of is to compare the learned weights in the end to some known weights that I created. However, for my learning to arrive at my known weights, I need to create data that fits my known weights. How can I do that? I see that it may depend on the model type. $\endgroup$ – siamii Apr 11 '13 at 23:36
  • 1
    $\begingroup$ That makes more sense. It will take more information on the type of your model to be able to answer. In general, you will need to either find an analytical solution to the weight optimization problem for your model or construct a similar but simpler model that has the same analytical characteristics of your real model. $\endgroup$ – Godric Seer Apr 12 '13 at 0:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.