Training Interpretable Convolutional Neural Networks towards Class-specific Filters