Robot Aesthetics and Cultural Imperialism: the Double Hermeneutic of Computational Photography

Abstract

This paper attempts to investigate the consequences of the emerging field of Automatic Aesthetic Quality Estimation, where deep neural networks are trained to predict the average ‘aesthetic rating’ of a photo. I first investigate these algorithms from a Bourdieuian perspective: the most popular training data on which such systems are based in DPChallenge, an American amateur digital photography challenge (with an explicit demographic bias in terms of race, class, nationality, gender, and age). Such algorithms are used widely in industry, hidden in well-known social media, internet search, and smartphone systems: contributing to the choice of top-ranked images in search engines, the choice of images for automatic social media photo-collection suggestions, automatically-chosen ‘cover’ images for events, and automatic ‘image enhancement’ on smartphones, to name but a few. I propose that the ubiquity of such systems is leading to a ‘double hermeneutic’ of visual aesthetic hegemony. As media organisations (including the press and advertisers) increasingly make use of internet search to find digital images for mass distribution, the statistical bias of the machine-learned features for aesthetic quality are disproportionately reproduced and encouraged through institutionalised publications. Analogous aesthetic biases are introduced in noninstitituliased image production through cloud and social media systems, replicating the aesthetic value-systems of the machine. Such a universal statistical shift in visual aesthetics has an immediate consequence for visual aesthetic hegemony. At the same time, the largest corporate machine-learning systems now learn constantly (‘online learning’). As the visual-aesthetic hegemony is therefore both reinforced by, and reinforces, corporate machine-learning aesthetic systems, we can identify a positive feedback loop between society and algorithm.