Explainable Attention for Few-shot Learning and Beyond