In recent years, there has been an increasing number of studies on objective image properties in visual artworks. Little is known, however, about how these image properties emerge while artists create their artworks. In order to study this matter, I produced five colored abstract artworks by myself and recorded state images at all stages of their creation. For each image, I then calculated low-level features from deep neural networks, which served as a model of responses properties in visual cortex. Two-dimensional plots of variances that were derived from these features showed that the drawings differ greatly at early stages of their creation, but then follow a narrow common path to terminate at or close to a position where traditional paintings cluster in the plots. Whether other artists use similar perceptive strategies while they create artworks remains to be studied.
Deep neural networks for direct, featureless learning through observation: The case of two-dimensional spin models
