Why Neural Networks Need Random Weight Initialization