Semi-supervised Learning with Support Isolation by Small-Paced Self-Training

In this paper, we address a special scenario of semi-supervised learning, where the label missing is caused by preceding filtering mechanism, i.e., an instance can enter subsequent process in which its revealed if and only it passes mechanism. The rejected instances are prohibited to labeling due economical or ethical reasons, making support labeled unlabeled distributions isolated from each other. case, learning approaches rely on certain coherence distribution would suffer consequent mismatch, hence result poor prediction performance. propose Small-Paced Self-Training framework, iteratively discovers subspaces with bounded Wasserstein distance. We theoretically prove that such framework may achieve provably low error pseudo labels during learning. Experiments both benchmark pneumonia diagnosis tasks show our method effective.