Can we improve meta-learning model in few-shot learning by aligning data distributions?