manetta berends index / H.I.T.md

H.I.T. - an algorithmic gameshow (2016)

H.I.T. is an algorithmic gameshow, developed in collaboration with Max Dovey during a residency at V2_, Institute for Unstable Media, Rotterdam.

Previous editions of H.I.T. took place at the 12th of February in Rotterdam (premiere), at the 19th of May in Rotterdam (during Kunstblock's Futuriosity conference) and at the 3rd of June in Weimar (during the Luxury Communism conference).

premiere of H.I.T., Rotterdam, February 2016

During H.I.T. (acronym for Human Intelligence tasks), six members from the audience perform in a visual recognition competition specialising in identifying the human labour involved in computer vision algorithms. It is a participatory performance where audience members compete to produce their own image recognition algorithm. It is styled on the annual ImageNet challenge, where groups of researchers compete to develop the most accurate image recognition algorithm. H.I.T. adapts this competition to reveal and perform the manual labour involved in training an algorithm on a public image dataset. Two teams play against each other in front of an audience where each round is a task taken from the production of an image recognition algorithm. Within the tasks, the cognitive labour is embodied by the participants as the often obscured process of producing a working algorithm is performed.

task 2 annotation interface

The starting point for H.I.T. is the auto-tagging application launched by Flickr in September 2014. The inaccurate tags caused great controversy and was featured widely in mainstream press. What received less attention was the debate surrounding the Flickr auto tagging bot that is still taking place on the users forum. Many Flickr users are expressing their dislike for automatic tagging and are forced to go through their individual images and delete the incorrect tags created by the auto tagging application. In doing so Flickr crowd sources the labour involved with training and improving the accuracy of their algorithm. But to what cost? The discussions on the forum frame the narrative as users ask pertinent questions about our role in the production of algorithms. Using this incident as a context to reveal the role of humans in the production of algorithms, each task in H.I.T. re-enacts the various procedures that were used to create the Flickr auto-tagging application from Flickr images.

premiere of H.I.T., Rotterdam, February 2016

The competition uses a popular image dataset, called 'Flickr8k', and re-enacts the selection-, annotation- and sorting tasks that occurs to produce an image recognition application. By doing this we aim not only to highlight and reveal the human labour in such algorithmic processes but also dramatize the radical reduction involved with such models of big data analysis. We want to create a counter narrative to the popular attitude that algorithmic processes are automated and the misleading claim that there are 'no humans involved'. Performing the story of the Flickr image dataset shows how the auto-tagging algorithm is produced by a mass of individual humans making subjective interpretations under extreme, unregulated conditions.

links


research and initiated work

commissioned work