| dc.date.accessioned | 2019-05-29T12:44:13Z | |
| dc.date.available | 2019-05-29T12:44:13Z | |
| dc.identifier.uri | http://95.216.75.113:8080/xmlui/handle/123456789/108 | |
| dc.description | Biography: Leonardo Impett is a PhD candidate in Computer Science in the Image and Visual Representation Lab (IVRL), EPFL, Switzerland. He is co-supervised by Sabine Süsstrunk (EPFL) and Franco Moretti (Stanford). His research focuses on the application of computer vision to human visual culture - particularly in Western art history – through a computational analysis of Aby Warburg’s Bilderatlas, and an ‘operationalisation’ of his theory of Pathosformel. As well as the use of computer techniques as tools for art history, his research attempts to use art-historical applications for the technical development of new computer-vision algorithms, including modifying Convolutional Neural Networks (CNNs) to be able to better generalise their understanding to non-photographic images, as humans can. He is interested in the epistemological and methodological overlaps between Digital Art History and Computer Science; to this end, he organises the Ways of Machine Seeing workshop in Cambridge, and the ‘Exploring Edges’ summer school at EPFL. He completed his undergraduate and masters degrees in Engineering at the University of Cambridge, where he remains a guest member of the Rainbow Research Group under Professor Alan Blackwell, with whom he has collaborated on a number of research projects, including a machine-learning robotic musical instrument for the Royal Opera House, London. He is a Member of the Institution of Engineering and Technology (IET) and a Fellow of the Royal Society of Arts. He has done computer-vision internships or consultancies for several companies, including Microsoft Research, Shell, and Boeing. | |
| dc.language.iso | en_US | |
| dc.title | Robot Aesthetics and Cultural Imperialism: the Double Hermeneutic of Computational Photography | |
| dc.contributor.author | Impett, Leonardo | |
| dc.description.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. | |
