Teaching-to-Learn and Learning-to-Teach for Few Labeled Classification

摘要

Co-training is a semi-supervised learning paradigm that trains some classifiers and let them label some unlabelled instances for each other during the learning process. One challenge of the co-training style algorithm is to train an initial weakly useful predictor when the number of labeled instances is very limited. In this paper, we use Teaching-to-learn and Learning-to-teach strategy, which each the ground truth is regarded as a "teacher". So that the entire labeled process is guided by the teachers and selected from simple instances to more difficult ones. In the teaching-to-learn step, teachers select the highest evidence instances based on distance and local density, which can automatically, spotted and excluded outliers from analysis. In the learning-to-teach step, teachers reversely learn from the co-training feedback to properly select the highest evidence instances for the next iteration. Thus, we propose a framework based co-training style algorithms to efficiently improve the performance of them. Moreover, we can tackle with real-world applications in which may have very few labeled training instances, eventually only have one labeled instance for each label. Finally, we give specific algorithm for efficient solving in different tasks and outperforming in UCI datasets, also achieve better performance in content-based image retrieval.