few-shot learning aims to recognize novel queries with limited support
samples by learning from base knowledge. Recent progress in this setting
assumes that the base knowledge and novel query samples are distributed in the
same domains, which are usually infeasible for realistic applic