1) get more training data 2) using weight decay to limit capacity of the model 3) using early-stopping 4) using cross-validation 5) average many different models 6) drop-out