Robust Mixture-of-Expert Training for Convolutional Neural Networks