The Unknown Knowns of Machine Vision, or, How to Explain Pictures to a Self-driving Car

This paper proposes a media archaeology of autonomous vehicles developed during a fellowship at Bosch Research Campus near Stuttgart, Germany, excavating some of the 'situated knowledges' (Haraway) at work in one of Europe's foremost research centres on automated driving. As various departments at Bosch strive to technically overcome the safety implications of optical ambiguity in the world, the paper deploys the missing category in the Rumsfeldian taxonomy of uncertainty to propose art history as the 'unknown known' of machine vision. From the bounding boxes of now classical object detection to contemporary semantic segmentation, the paper probes the technical images used to train self-driving cars and traces some of the aesthetic prehistories that constitute the ‘optical unconscious’ (Benjamin) underlying their operative transparency within this corporate industrial milieu. This includes an analysis of the Daimler Cityscapes machine learning dataset comprising 25'000 semantically labelled photographs of urban street scenes throughout 50 European cities which is used by AI researchers at Bosch to preemptively test the capacity of machinic perception to be deceived in so-called 'universal adversarial perturbations', a form of steganography that alters what the machine 'sees' while remaining invisible to the human eye. The paper concludes by presenting the results of a creative collaboration with the AI team to reinsert a reflexive piece of photographic history into this novel and uniquely machine-readable class of images, closing a loop that situates autonomous vehicles within a teleology of photography.