I'm actually splitting my data into training and test sets. Some people use "validation" datasets for "tuning" the classifier parameters (specially in neural networks). My question is if can I just select a balanced subset from my training set as validation set.