Measuring Mode Collapse: Comparative Study Between Architectures Gan

Abstract

Generative Adversarial Networks (GANs) are a type of generative model that have been widely used in various applications, but they often suffer from a common problem called mode collapse. This phenomenon occurs when the generator learns to produce only a small group of images instead of diverse ones. Mode collapse can occur for two reasons: firstly, when the discriminator becomes so effective that the generator can no longer learn, and secondly, when the generator finds a way to deceive the discriminator with a small number of samples, causing it to lack motivation to diversify its outputs. This study presents a comparison of multiple GAN-based models by evaluating them with various metrics that measure mode collapse. The behavior of models with similar parameters is analyzed, as well as the results obtained using the parameters originally published for each model.