Ultra-low Precision Multiplication-free Training for Deep Neural Networks