GPLVM and lava floor distance for label-deficient semi-supervised learning: Case study

Label-deficient semi-supervised learning is a challenging setting in which there is an abundance of unlabeled data but a dearth of labeled data. We propose a method for applying Gaussian process latent variable models (GPLVM) in a label-deficient setting, a method in which the discriminative GPLVM objective function trains a back-constraining neural network followed by a transformation into a semi-supervised classifier using the lava floor distance. We provide a case study that visually details the necessities and drawbacks of all of the components of the method, including the behavior of the standard GPLVM as a method of dimension reduction, the effect of adding small amounts of labeled data via discriminative GPLVM, and the characteristics of back-constrained GPLVM using neural networks built with various nonlinearities as a way of handling large data sizes. The method achieves 81% accuracy on unseen testing data compared to a standard support vector classifier trained on the same labeled points which achieves 70% accuracy.