Every Image An Eigenimage

• In Event: Platform Practices and Predictive Seeing 1

Abstract

The practice and concept of the 'eigenface' has played a major role in allowing a 'system' of facial recognition to become operative. During the 1990s, cognitive science researchers
stabilised a face image that was compositionally centred, illuminated to distinguish background from foreground, standardised in terms of image dimensions and ideally set against some kind of constrained background environment such as an office or domestic space. These material and aesthetic constraints on face images allowed templatised faces to become the images upon which facial recognition data sets were then trained.

This paper argues that such practices of training constitute the
ontogenetic conditions for the production of a visual culture now committed to training all images. The eigenface has become an operation enabling the ongoing production of eigenimages. The eigenimage demands that any image belonging to a data collection must be trained as and alongside its eigenimage. A kind of pre-emptive prediction develops in which there are only ever eigenimage s and modes of imaging. Take style transfer, any image recognition task, or more sinister forms of securitisation based on facial recognition, for example. This paper will also inquire into the ways in which training in deep learning image recognition examples also breaks down, and the eigenimage becomes a deformation. Does this allow novel image production or are these simply reincorporated into eigenimaging?

Author

• Anna Munster, University of New South Wales