UC Irvine

Latent variable models provide a robust framework for modeling complex data distributions while accounting for known or desired inductive biases. The compact, low-dimensional representations learned by mixture models, variational autoencoders and other latent variable models also provide useful and interpretable bases for downstream prediction tasks. Our work introduces a novel framework for training latent variable models using prediction constraints, which aims to balance two important objectives: high-quality generative modeling of complex data and accurate prediction of semantic labels. Our framework acknowledges the inherent asymmetry of our discriminative objective, which is to learn how to predict labels from data, rather than to predict data from labels. We show that addressing this issue allows us to effectively leverage latent variable models for both supervised and semi-supervised learning while retaining interpretability and generative performance. We further introduce additional consistency constraints, derived naturally from the generative model, that requires predictions on reconstructed data to match those on the original data. We show that enforcing consistency is crucial when labels are very sparse. We apply our framework to a variety of latent variable models including: mixture models, topic models, hidden Markov models and variational autoencoders. Our experiments show state-of-the-art semi-supervised learning performance on diverse tasks using each of these models.