Optimizing Memory Efficiency for Deep Convolutional Neural Networks on GPUs