Flexible Prior Distributions for Deep Generative Models

TL;DR

This paper proposes using flexible, data-driven prior distributions in deep generative models to enhance their power, interpretability, and control over generalization, challenging the traditional simple prior approach.

Contribution

It introduces a method to induce flexible prior distributions directly from data, improving generative modeling capabilities and latent structure understanding.

Findings

More powerful generative models achieved

Enhanced modeling of latent structure

Explicit control of generalization degree

Abstract

We consider the problem of training generative models with deep neural networks as generators, i.e. to map latent codes to data points. Whereas the dominant paradigm combines simple priors over codes with complex deterministic models, we argue that it might be advantageous to use more flexible code distributions. We demonstrate how these distributions can be induced directly from the data. The benefits include: more powerful generative models, better modeling of latent structure and explicit control of the degree of generalization.