Improved weight initialization for deep and narrow feedforward neural network