Improving the Performance of Fully Connected Neural Networks by Out-of-Place Matrix Transpose