Explanation Consistency Training: Facilitating Consistency-Based Semi-Supervised Learning with Interpretability

Unlabeled data exploitation and interpretability are usually both required in reality. They, however, conducted independently, very few works try to connect the two. For unlabeled exploitation, state-of-the-art semi-supervised learning (SSL) results have been achieved via encouraging consistency of model output on perturbation, that is, assumption. However, it remains hard for users understand how particular decisions made by SSL models. To this end, paper we first disclose assumption is closely related causality invariance, where invariance lies main reason why valid. We then propose ECT (Explanation Consistency Training) which encourages a consistent decision under perturbation. employs explanation as surrogate output, able bridge models alleviate high complexity causality. realize ECT-SM vision ECT-ATT NLP tasks. Experimental real-world sets validate highly competitive performance better proposed algorithms.