Grand Theft Autoencoder

Abstract

The implementation of generative models in deep learning, particularly those of Text-to-Image Synthesis (T2I), are essentially an exaptation of the cognitive processes of the transcendental imagination Kant outlined in his notoriously opaque schematism chapter of CPR. While such engineering feats mirror the liberating force of photography’s invention, they have also proven to be a significant engine for reproducing antediluvian ideologies of art pivoting on claims about what has been stolen by the machine. This paper argues that T2I presents an opportunity to instead reconsider what our models of the procedures of the imagination actually are or could be, and wagers that the interdisciplinary conceptual frameworks supporting machine learning enable us to recuperate from an “incommensurable” synthetic intelligence the necessary resources for revising our understanding of what creativity is and does, with pattern recognition providing the tools for a renewed elaboration of techné to pull a heist upon the transcendental itself.