Katja de Vries, Postdoctoral Researcher, Business IT Department, IT University Copenhagen
Machine Learning (ML) applications shape, and are shaped by, human values. In the last decade classificatory models have become the most widely applied type of ML. However, since the invention of Generative Adverserial Networks (GANs) in 2014, the field of generative ML is likely to become equally prolific in the years to come. GANs can generate convincing fakes of video footage, pictures, graphics, DNA strings, etc. Some have called their capacity to mimic and recreate any informational pattern 'machine imagination'. In this paper I explore what this new ML technique means in terms of how machines and humans co-create and co-organize life. I will discuss three important aspects of GANs. Firstly, what does it mean that GANs are trained on unlabelled data? Does that make them less prone to human bias? Secondly, what are appropriate success and optimization criteria for GANs? In classificatory ML prediction accuracy is the most common measure of success. GANs lack such an obvious criterion: assessing how good a creation is, is as difficult as assessing the quality of art produced by a human artist. Thirdly, how do GANs push the traditional opposition between fact and fiction out of joint? For example, GANs could offer a solution against unrepresentative datasets by their capacity of generating synthetic data. While a synthetically adjusted dataset could be more representative of 'reality' than a set merely consisting of 'real' data, this practice would challenge the crucial scientific principle that data should not be fabricated.
