Identifying Generalization Properties in Neural Networks