Efficient Uncertainty Quantification and Reduction for Over-Parameterized Neural Networks