A note on the evaluation of generative models

TL;DR

This paper critically examines common evaluation metrics for image generative models, highlighting their independence and limitations, and emphasizes the importance of application-specific evaluation methods.

Contribution

It clarifies the independence of evaluation criteria and advises against using Parzen window estimates, promoting direct, application-oriented evaluation of generative models.

Findings

Average log-likelihood, Parzen estimates, and visual fidelity are largely independent.

Good performance in one metric does not imply good performance in others.

Parzen window estimates should generally be avoided.

Abstract

Probabilistic generative models can be used for compression, denoising, inpainting, texture synthesis, semi-supervised learning, unsupervised feature learning, and other tasks. Given this wide range of applications, it is not surprising that a lot of heterogeneity exists in the way these models are formulated, trained, and evaluated. As a consequence, direct comparison between models is often difficult. This article reviews mostly known but often underappreciated properties relating to the evaluation and interpretation of generative models with a focus on image models. In particular, we show that three of the currently most commonly used criteria---average log-likelihood, Parzen window estimates, and visual fidelity of samples---are largely independent of each other when the data is high-dimensional. Good performance with respect to one criterion therefore need not imply good performance with respect to the other criteria. Our results show that extrapolation from one criterion to another is not warranted and generative models need to be evaluated directly with respect to the application(s) they were intended for. In addition, we provide examples demonstrating that Parzen window estimates should generally be avoided.