This work explores how a feedback loop of generative adversarial neural networks (GANs) is transforming errors into meaning while at the same time confirming and emphasizing biases in the training data.
Three GANs and a camera form a closed chain in which they interpret and transform the input they receive amongst each other. Each model has been trained on thousands of painted “old masters” portraits harvested from various European collections. One model’s purpose is to generate a biometric semantic map of all faces found in an image, another model translates these maps back into portraits, a third model adds details and texture to the result. A camera filming the artist is mixed into the loop in varying degrees, disturbing and changing the face markers.
Since all the models’ knowledge consists of faces they will gradually transform any kind of incoming data, even noise, into their parameter space. The models’ misinterpretations of small errors accumulate due to the closed loop, while at the same time the errors and imprecisions introduced during the generation process prevent the system from ending up in a static state and result in a fluid creative process that traverses the possibility space of the system.
Website: http://quasimondo.com
