Pseudo Orthogonal Bases Give the Optimal Solution to Active Learning in Neural Networks