ssProp: Energy-Efficient Training for Convolutional Neural Networks with Scheduled Sparse Back Propagation