Learning Purified Feature Representations from Task-irrelevant Labels | IEEE Conference Publication | IEEE Xplore