TGPGAN: towards expression-based generative adversarial networks

Before the advent of Generative Adversarial Networks (GANs), Evolutionary Computation approaches made up the majority of the state of the art for image generation. Adversarial models employing Deep Convolutional Neural Networks better fitted for GPU computing have been at this point more efficient. Nevertheless, motivated by recent successes in GPU-accelerated Genetic Programming (GP) and given the disposition of expression-based solutions towards image evolution, we believe in the prospect of using symbolic expressions as generators for GANs, instead of neural networks. In this paper, we propose a novel GAN model called TGPGAN, where the traditional convolutional generator network is replaced with a GP approach. The generator iteratively evolves a population of expressions that are then passed to the discriminator module along with real images for backpropagation. Our experimental results show that it is possible to achieve comparable results to a typical Deep Convolutional GAN while benefiting from the flexibility enabled by an expression-based genotype. Moreover, this work serves as a proof of concept for the evolution of symbolic expressions within adversarial models.