(Local Maxima: SLFT2, Square, 2019)

Local Maxima, 2019, GAN-generated images and giclee prints

In classical AI research, a “local maximum” is a solution to a problem that cannot in itself be improved but that is not the best possible solution. Where a singular mathematical specification of a problem is possible, local maxima can be avoided using search strategies other than simple “hill climbing”. Art does not have such a specification, its search space consists entirely of local maxima.

Generative Adversarial (neural) Networks extract aesthetic surplus value from images, rendering it liquid. Training a GAN deterritorializes their form, generating new images with it reterritorializes it. Contemporary AI burns capital to produce discrimination, here discrimination between images that resemble the learned source material and those that do not.

The source material in this instance is my own previous work, playful compositions from the 1990s. A GAN was trained on them and then applied to simple geometric shapes, transferring the extracted “style” of the source images leavened with the perceptual artefacts of GAN training. Reducing (or promoting) artworks to flows of aesthetics that can be continued indefinitely is deeply challenging to art history and connoseurship.

GAN generated images tend to Pollockian all-overness, or to the same local sense but global semantic incoherence as Markov chains or a Francis Bacon painting. The conception of “style transfer” that all of this involves is deeply challenging to art theory.