Crafting the Weights of a Convolutional Neural Network to Make a Drawing

A growing number of visual artists use convolutional neural networks (CNNs) in their practice. While CNNs show promise as a form of representation in art, the lack of interpretability of CNNs limits creative control to high level decisions around datasets, algorithms, and hyperparameters. As an alternative, the field of computer vision presents a more immediate paradigm of control through the hand-crafting of convolutional kernels. This thesis investigates the hand-crafted approach as an additional creative lever for artists working with CNNs. It reimagines network weights as a continuous, spatial, and computational material supporting direct human interaction. Two experimental tools are proposed: one for parametrically

generating first layer kernels and the other for editing multiple layers. These tools attempt to transform the hand-crafting of features into “crafting” in a truer sense by bringing

CNNs and visual materials into a close feedback loop. The author extensively engaged with these tools and this serves as a case study that examines the affordances of hand-crafted CNNs. The results suggest that hand-crafted CNNs can be a viable form of representation for artists seeking to build simple, bespoke feature detectors, but that more complex CNNs would likely require a hybrid approach integrating data-driven methods.