Figure 6.

Schematic of Coupled Generative Adversarial Networks. Top: Schematic of a standard generative adversarial network applied to the problem of camouflage. A generator (blue) neural network generates data x_g with the goal of emulating the potential distributions of real data x_r as best as possible. In this example, inspired by Talas et al. [139], the generator generates stripe patterns that, when emerged in the habitat of a Siberian tiger, are aimed at fooling the discriminator in not being able to discern generated tiger patterns from tree patterns in the background. A second neural network called discriminator receives unlabeled data from both real and generated distributions and aims to label them correctly by origin. Bottom: Schematic of a coupled generative adversarial network. When fitness landscapes of two populations are coupled, each generative adversarial network type agent tries to internally generate and discern trait patterns that correspond to the observed traits of the other population, thereby learning and predicting the traits of the other population. Because in a coupled generative adversarial network weights between the first few generator layers, and last few discriminator layers are shared, this configuration allows the coupled generative adversarial network to learn a joint fitness distribution underlying traits without correspondence supervision [140]. Tiger and birch tree images are adapted from images taken from flickr.com (last accessed 02/09/2022) under the CC BY-NC-SA 2.0 license.