Abstract by Mike Brodie
MC-GAN: Multiple Choice Generative Adversarial Networks
Since its introduction in 2014, generative adversarial network (GAN) paradigms have become an increasingly popular means to produce state-of-the-art synthetic outputs. Despite impressive results in various domains, GANs require protracted training times and remain difficult to train successfully. In addition to disappearing gradients, GANs frequently suffer from mode collapse, where models only learn to generate a small subset of samples with low levels of diversity. While various extensions have since improved model stability, output diversity remains a significant obstacle to GAN training. We introduce a novel Multiple Choice GAN (MC-GAN) ensemble training approach that encourages greater domain specialization in individual models. We present comprehensive experimental results that demonstrate that MC-GAN ensembles yield improved Inception scores compared with baseline approaches. Additionally, we show that MC-GAN inherently avoids mode collapse and produces models specialized in largely non-overlapping subsets of the input domain.