Algorithms of Vision. Human and machine learning in computational visual culture

Current computer vision algorithms largely depend on the availability of images labelled by human annotators at very high speed. The mode of production of these annotations strongly resonates with an early experiment conducted in 2007 at Caltech by Fei Fei Li, initiator of ImageNet, one of the most popular visual datasets. In a laboratory, the subjects were asked to describe photographs shown for a few milliseconds and to filter them through a taxonomy. The Caltech experiment is used, in the thesis, to engage with the photographic elaboration of computer vision: the model of vision, the photographic alignments and the micro-temporal rhythm that subtend the modes of production of labelled data and the labour behind it. The written and practice components of the submission elaborate a novel method and document the path towards it. The method has developed in the context of practice-led research in collaboration with The Photographers' Gallery and crystallised into a project, Variations on a Glance, a series of re-enactments based on the Caltech experiment. The original experimental protocol is submitted to several variations, called re-experiments, exploring its potential to produce a time-critical model of vision and collective visual interpretations. The experimental protocol is re-designed iteratively to explore specific configurations of micro-temporal vision and different configurations of collectives of human and non-human participants. The thesis examines the dynamics of these collectives, in particular how they reach consensual interpretation, and how the taxonomic practices of the lab interfere in this process. The contribution of this research is a mapping of the entanglement of computer vision and photography and a method embedded in practice that does not attempt to resolve the differences and tensions between photography and computer vision but provides a device to explore the texture of their relation. The research complements and complicates the recent critiques related to bias and discrimination in machine learning and the exploitative work conditions it relies on. Finally it offers to the photographic institution and its public a mode of intervention into the making of computer vision.