Exploiting Invariance in Training Deep Neural Networks