why for neural networks it is advised to set a test dataset apart to check if it overfitted while for statistical models fitted through mcmc this is never done? if a model has too many parameters shouldn't it overfit with both methods?
$\begingroup$
$\endgroup$
2
-
1$\begingroup$ I presume what you mean by "MCMC methods" are Bayesian methods, in which case the "guard" against overfitting stems from the prior, not from the MCMC sampling. $\endgroup$– DurdenCommented Mar 17, 2023 at 15:55
-
$\begingroup$ @Durden but no one ever set apart a test dataset for models fitted with monte carlo samplers, why? $\endgroup$– AlucardCommented Mar 17, 2023 at 18:12
Add a comment
|