Semiparametric Latent Variable Models for Guided Representation

Unsupervised discovery of latent representations, in addition to being useful for density modeling, visualisation and exploratory data analysis, is also increasingly important for learning features relevant to discriminative tasks. Autoencoders, in particular, have proven to be an effective way to learn latent codes that reflect meaningful variations in data. A continuing challenge, however, is guiding an autoencoder toward representations that are useful for particular discriminative tasks. A complementary challenge is to find codes that are explicitly invariant to irrelevant transformations of the data. To address these difficulties, we introduce the semiparametric latent variable model (SPLVM), which combines an autoencoder with a Gaussian process latent variable model. The SPLVM enables an autoencoder's unsupervised representation to both incorporate relevant label information and ignore irrelevant variations. We apply this model to several different datasets, examining how both labels and nuisance variables can provide cues for useful latent representations.