Creative Adversarial Networks: GANs that make art

Generative Adversarial Networks use a pair of machine-learning models to create things that seem very realistic: one of the models, the "generator," uses its training data to make new things; and the other, the "discerner," checks the generator's output to see if it conforms to the model. Rutgers comp sci prof Ahmed Elgammal runs an Art and AI Lab where they use "Creative Adversarial Networks" to produce new artworks: CANs use a "discerner" that seeks out "novelty," not fidelity to the statistical predictions of the model. The underlying theory is that art evolves "through small alterations to a known style that produce a new one," which, as Ian Bogost (previously) points out, is "a convenient take, given that any machine-learning technique has to base its work on a specific training set." Elgammal recent exhibited a show called Faceless Portraits Transcending Time at Chelsea's HG Contemporary gallery; and his choice of portraiture as a means of showcasing the capabilities of CANs has proven to be controversial: as art historian John Sharp says, "You can't really pick a form of painting that's more charged with cultural meaning than portraiture." Portraits use extensive, coded symbology to say something about their subjects, and CANs do not, by themselves, understand or correctly use these symbols in the works they create.