Ternary reversible extreme learning machines: the incremental tri-training method for semi-supervised classification

Tri-training method proposed by Zhou et al., is an excellent method for semi-supervised classification; nevertheless, the heavy computational burden caused by the retraining strategy prevents the further application of tri-training method. To address this problem, this paper proposes the ternary reversible extreme learning machines (TRELM) which is an incremental tri-training method without relying on the retraining strategy. TRELM employs three reversible extreme learning machines (RELM) as its base learners and trains the RELM with extended (or detected) samples in each learning round. RELM is an incremental learning method with reversible derivation capability. RELM can overcome the difficulty for most incremental learning methods in removing the influence of previously learned mistaken samples. Experimental results indicate that TRELM significantly improves the learning speed of tri-training method. In addition, TRELM achieves comparable (or even better) classification performance to other effective semi-supervised learning methods. TRELM is an appropriate choice for semi-supervised classification tasks with large amounts of data sets or with strict demands for learning speed and classification accuracy.