I’m rather confused about the precise function of the test dataset defined in the process of building an MLP with ROOT. The User’s Guide says this dataset “will avoid bias,” but I’m not exactly sure how. From using the “graph” option during training I know that the error on the test dataset is computed during training. However, I am not sure whether the test dataset is actually used during training; i.e., is some kind of “early stopping” mechanism implemented whereby the network will stop training when the test error starts to increase. Or is the test dataset error merely supposed to be consulted afterwards as an unbiased estimator of the error?
There is a fair amount of terminological confusion in the neural network literature about what exactly the terms “test dataset” and “validation dataset” mean, and I would like to make sure I understand how they are being used in this context.