This artwork is a result of an ongoing experiment that attempts to mimic convergent evolution when training generative AI models.

In nature, evolution sometimes ends up shaping different species in similar ways. For instance, the ancestors of humans and octopuses both evolved eyes independently. Just like isolation may cause species to branch off, similar environmental conditions sometimes lead to convergence. Ancestors of whales evolved a streamlined body independently of fish, converging into a similar appearance despite taking very different paths.

This video was generated by an AI art model that emulated this process: isolation, branching off, and convergence. The model was first trained on all images in a large, generic image dataset. Then, copies of the model were further trained on separate datasets in isolation. Finally, a unified model was created by averaging the weights of the instances and it was once again trained on the initial dataset. Indeed, the weight averaging part does not really occur in nature but it was a convenient shortcut to implement a fast convergence. I call this approach Artificial Convergent Evolution.

The results of this process were striking: the model generated images that looked almost nothing like the images in any of the datasets yet they retained the aesthetics and regularity of natural images. To the best of my knowledge, this is the first time a generative model was trained through Artificial Convergent Evolution.

Leo Isikdogan is a researcher, engineer, and artist based in California. He has a Master's degree focused on computer vision for art analysis and a PhD from the Laboratory for Image and Video Engineering at The University of Texas at Austin.

Currently, Isikdogan works as a machine learning engineer at Apple. His practice spans a broad range of disciplines, including artificial intelligence, computer vision, computational photography, image processing, and augmented reality.