Selecting the Best in GANs Family: a Post Selection Inference Framework
This work addresses the need for a statistically rigorous method to evaluate and select among GAN models for researchers, though it is incremental as it builds on existing MMD and PSI techniques.
The paper tackles the problem of selecting the best-performing GAN variant by proposing a post-selection inference framework using an incomplete U-statistics estimate of maximum mean discrepancy (MMD_inc) to measure distribution discrepancies, and applies it to compare 7 GAN variants.
"Which Generative Adversarial Networks (GANs) generates the most plausible images?" has been a frequently asked question among researchers. To address this problem, we first propose an \emph{incomplete} U-statistics estimate of maximum mean discrepancy $\mathrm{MMD}_{inc}$ to measure the distribution discrepancy between generated and real images. $\mathrm{MMD}_{inc}$ enjoys the advantages of asymptotic normality, computation efficiency, and model agnosticity. We then propose a GANs analysis framework to select and test the "best" member in GANs family using the Post Selection Inference (PSI) with $\mathrm{MMD}_{inc}$. In the experiments, we adopt the proposed framework on 7 GANs variants and compare their $\mathrm{MMD}_{inc}$ scores.