Semi-Supervised Learning with Very Few Labeled Training Examples

Abstract

In semi-supervised learning, a number of labeled exam-ples are usually required for training an initial weakly useful predictor which is in turn used for exploiting the unlabeled examples. However, in many real-world applications there may exist very few labeled train-ing examples, which makes the weakly useful pre-dictor difficult to generate, and therefore these semi-supervised learning methods cannot be applied. This paper proposes a method working under a two-view set-ting. By taking advantages of the correlations between the views using canonical component analysis, the pro-posed method can perform semi-supervised learning with only one labeled training example. Experiments and an application to content-based image retrieval val-idate the effectiveness of the proposed method.