Infinite Descent

"INFINITE DESCENT is the result of my early experiments with custom-trained models and coupling them in feedback loops. A CycleGANmodel that I had trained on tens of thousands of engraved portraits found in the public domain collection of the British Library and the Internet Archive is feeding its outputs back into its inputs.

My fascination with feedback loops in neural networks stems from the fact that their emerging patterns tend to result in a distillation of the most prominent features they have learned—in this case, elements that make up faces: eyes, noses, and mouths in line-art fashion. 

Another aspect that attracts me to them is their instability: often these systems are rather fragile, meaning they can only stay "alive in a very narrow range of hyperparameter settings, and finding those zones is not a scientific process but rather an intuitive one. The color is added by a second CycleGAN model that I had trained for photo colorization by pairing black-and-white and color images."

– Mario Klingemann

Mario Klingemann on 25 June 2017: "Two CycleGAN models trained on face transformations are passing their outputs back and forth creating an infinite loop of ever-changing Turing patterns."