On the Evaluation of Generative Adversarial Networks By Discriminative Models

Generative Adversarial Networks (GANs) can accurately model complex\nmulti-dimensional data and generate realistic samples. However, due to their\nimplicit estimation of data distributions, their evaluation is a challenging\ntask. The majority of research efforts associated with tackling this issue were\nvalidated by qualitative visual evaluation. Such approaches do not generalize\nwell beyond the image domain. Since many of those evaluation metrics are\nproposed and bound to the vision domain, they are difficult to apply to other\ndomains. Quantitative measures are necessary to better guide the training and\ncomparison of different GANs models. In this work, we leverage Siamese neural\nnetworks to propose a domain-agnostic evaluation metric: (1) with a qualitative\nevaluation that is consistent with human evaluation, (2) that is robust\nrelative to common GAN issues such as mode dropping and invention, and (3) does\nnot require any pretrained classifier. The empirical results in this paper\ndemonstrate the superiority of this method compared to the popular Inception\nScore and are competitive with the FID score.\n