Reducing Linear Redundancy in Neural Networks