Which is more likely to result in overfitting a model? More training data, decreased nodes in neural network hidden layers, eliminating sparse features, or using a Gaussian or RBF kernel in SVM?