Is it better to have more training data which are close to the decision boundary, or more data which are "typical" of their class? • /r/MachineLearning