Convolutional Aesthetics

A cultural and philosophical analysis of the perceptual logic of machine learning systems

PhD defence by Naja le Fevre Grundtmann.

Assessment committee

• Associate professor Bjarkí Váltysson, Chair (University of Copenhagen)
• Professor Celia Lury (University of Warwick)
• Associate professor Winnie Soon (Aarhus University)

Head of Defence

• Associate professor Ulrik Ekman (University of Copenhagen)

Copies of the thesis will be available at:

• The Copenhagen University Library South Campus (KUB Syd), Karen Blixens Plads 7
• The Royal Library at Søren Kierkegaards Plads 1 (the Black Diamond)

Abstract

Machine learning systems have reached a performance level comparable to humans in certain recognition tasks, such as object recognition in images. This thesis analyses these historically unprecedented and contemporarily influential technologies via the proposition that their perceptual logic constitutes aesthetics. Focusing on current theorisations and practices of state-of-the-art computer vision, the thesis examines how the use of visual material in the field of computer science informs the understanding of the perceptual logic of machine learning systems. Computer scientists often use artistic material or other kinds of images as resources to provide evidence for various kinds of hypotheses regarding the inner workings of artificial neural networks. Drawing on several disciplines, this study assesses the aesthetic and technical premisses of these claims by a close reading of high-level descriptions of machine learning systems together with cultural, philosophical and aesthetic theory. It demonstrates that there is a discrepancy between how some computer scientists mobilise art to theorise machine learning and the actual functioning of the processes they engineer. Theorising the perceptual logic of artificial networks in terms closer to their technical description, this thesis contends that machine learning systems foster a generative environment in interaction with data. By engaging with the processes performed by convolutional neural networks, this thesis claims that the particular ways in which input data is folded into the architecture of these systems constitute part of this new and idiomatic mode of aesthetics, which it proposes to be a “convolutional aesthetics.” In identifying these convolutional aesthetics, this project makes claims toward understanding the expanded ways in which vision is inhabited and operates in machine learning systems and their wider contexts.