A Factorial Mixture Prior for Compositional Deep Generative Models

TL;DR

This paper introduces a factorial mixture prior for deep generative models that captures structured compositionality of high-dimensional data by modeling latent properties as a Cartesian product of quantized subspaces, enabling unsupervised and semi-supervised learning of properties.

Contribution

It proposes a novel factorial mixture prior that models latent properties as a Cartesian product of subspaces, enhancing compositionality in deep generative models.

Findings

Effective in capturing latent properties in images

Supports unsupervised and semi-supervised learning of properties

Improves interpretability of generative models

Abstract

We assume that a high-dimensional datum, like an image, is a compositional expression of a set of properties, with a complicated non-linear relationship between the datum and its properties. This paper proposes a factorial mixture prior for capturing latent properties, thereby adding structured compositionality to deep generative models. The prior treats a latent vector as belonging to Cartesian product of subspaces, each of which is quantized separately with a Gaussian mixture model. Some mixture components can be set to represent properties as observed random variables whenever labeled properties are present. Through a combination of stochastic variational inference and gradient descent, a method for learning how to infer discrete properties in an unsupervised or semi-supervised way is outlined and empirically evaluated.