12-12-2016 02:36 PM
Came to know from articles that e miner automatically selects the model that has the least average square error (or missclassification) on the validation data set. This is called stopped training method which definitely helps to ensure NN does not overfit. So it's mean that even if the model converges after say 56th iteration, it may select the final model with 39th iteration if validation dataset has minimum error (or missclassification) on this particular iteration. So my confusion is if I assign training and test data set (insetad of validation) then will the e miner select the last model (that means weight estimate after 56th iteration in my example) ? Since there is a direct involvement of the validation data on building the model, how correct the method is? I mean the model might be different if I select some other validation data set.
12-19-2016 04:23 PM
If you don't use a validation partition for early stopping, then yes, as you say the neural network model will likely overfit and not generalize well to new data. So you really want to use a validation partition when building your neural network.