On the Quantitative Analysis of Decoder-Based Generative Models

TL;DR

This paper introduces a method using Annealed Importance Sampling to accurately evaluate the log-likelihoods of decoder-based generative models, addressing the challenge of performance quantification.

Contribution

It proposes a novel evaluation technique for decoder-based models and validates its accuracy, enabling better analysis of model performance and data distribution coverage.

Findings

Annealed Importance Sampling provides reliable log-likelihood estimates.

Decoder-based models often overfit and miss important data modes.

Existing estimators can be inaccurate, affecting model assessment.

Abstract

The past several years have seen remarkable progress in generative models which produce convincing samples of images and other modalities. A shared component of many powerful generative models is a decoder network, a parametric deep neural net that defines a generative distribution. Examples include variational autoencoders, generative adversarial networks, and generative moment matching networks. Unfortunately, it can be difficult to quantify the performance of these models because of the intractability of log-likelihood estimation, and inspecting samples can be misleading. We propose to use Annealed Importance Sampling for evaluating log-likelihoods for decoder-based models and validate its accuracy using bidirectional Monte Carlo. The evaluation code is provided at https://github.com/tonywu95/eval_gen. Using this technique, we analyze the performance of decoder-based models, the effectiveness of existing log-likelihood estimators, the degree of overfitting, and the degree to which these models miss important modes of the data distribution.