Training Interpretable Convolutional Neural Networks by Differentiating Class-specific Filters.