Inverting Variational Autoencoders for Improved Generative Accuracy

TL;DR

This paper introduces a semi-supervised deep generative model that leverages large unlabeled datasets with known structure to improve disentanglement and prediction accuracy, demonstrating benefits on spectroscopic and digit data.

Contribution

It presents a parameter-efficient model that exploits unfeatured data with known structure, enhancing generative and discriminative performance in semi-supervised learning.

Findings

Improved disentanglement of latent variables.

Enhanced discriminative prediction accuracy.

Effective utilization of large unlabeled datasets.

Abstract

Recent advances in semi-supervised learning with deep generative models have shown promise in generalizing from small labeled datasets ($\mathbf{x},\mathbf{y}$) to large unlabeled ones ($\mathbf{x}$). In the case where the codomain has known structure, a large unfeatured dataset ($\mathbf{y}$) is potentially available. We develop a parameter-efficient, deep semi-supervised generative model for the purpose of exploiting this untapped data source. Empirical results show improved performance in disentangling latent variable semantics as well as improved discriminative prediction on Martian spectroscopic and handwritten digit domains.