Evolutionary Generative Adversarial Networks

Abstract

Generative adversarial networks (GAN) have been effective for learning generative models for real-world data. However, existing GANs (GAN and its variants) tend to suffer from training problems such as instability and mode collapse. In this paper, we propose a novel GAN framework called evolutionary generative adversarial networks (E-GAN) for stable GAN training and improved generative performance. Unlike existing GANs, which employ a pre-defined adversarial objective function alternately training a generator and a discriminator, we utilize different adversarial training objectives as mutation operations and evolve a population of generators to adapt to the environment (i.e., the discriminator). We also utilize an evaluation mechanism to measure the quality and diversity of generated samples, such that only well-performing generator(s) are preserved and used for further training. In this way, E-GAN overcomes the limitations of an individual adversarial training objective and always preserves the best offspring, contributing to progress in and the success of GANs. Experiments on several datasets demonstrate that E-GAN achieves convincing generative performance and reduces the training problems inherent in existing GANs.

Coevolution of Generative Adversarial Networks

“…Recently, a model was proposed to use evolutionary algorithms in GANs [26]. Their approach used a simple model to evolve GANs, using a mutation operator that can switch only the loss function of the individuals.…”

Spatial evolutionary generative adversarial networks

“…Some remarkable examples analyze training a cascade of GANs [23]; sequentially training and adding new generators with boosting techniques [21]; training in parallel multiple generators and discriminators [8]; and training an array of discriminators specialized in a different low-dimensional projection of the data [14]. 1 Mustangs source codehttps://github.com/mustang-gan/mustang Recent work by Yao and co-authors proposed E-GAN, whose main idea is to evolve a population of three independent loss functions defined according to three distance metrics (JSD, LS, and a metric based on JSD and KL) [22]. One at a time, independently, the loss functions are used to train a generator from some starting condition, over a batch.…”

“…In this study, we add mutation diversity to the genome diversity provided by Lipizzaner. Thus, we use the mutations used by E-GAN to generate the offspring of generators [22]. E-GAN applies three different mutations corresponding with three different minimization objectives w.r.t.…”

“…The new generator is produced by using a loss function (mutation) to optimize one of the three objectives functions introduced above. i.e., E-GAN applies the three mutations to the generator (ancestor) and it selects the individual that provides the best fitness when is evaluated against the discriminator [22]. In contrast, Mustangs picks at random with same the probability ( 1 3 ) one of the mutations (loss functions), as it is shown in Figure 2, and then the gradient descent method is applied accordingly (see Algorithm 2).…”

“…Each of these objective functions improves training but none entirely solves all of its challenges. An evolutionary computation project investigated an evolutionary generative adversarial network (E-GAN), a different approach [22]. E-GAN, batch after batch, is able to guide its trajectory with gradient information from a population of three different objectives, which defines the gradient-based mutation to be applied.…”