Before A Computer Can Draw, It Must First Learn To See

Most computationally creative systems lack adequate means of perceptually evaluating the artifacts they produce and are therefore not fully grounded in real world understanding. We argue that perceptually grounding such systems will increase their creative potential. Having adequate perceptual abilities can enable computational systems to be more autonomous, learn better internal models, evaluate their own artifacts, and create artifacts with intention. We draw from the fields of cognitive psychology, neuroscience, and art history to gain insights into the role that perception plays in the creative process. We use examples and methods from deep learning on the task of image generation and pareidolia to show the creative potential of systems with advanced perceptual abilities. We also discuss several issues and philosophical questions related to perception and creativity.