Transformations in Image Extraction Regimes: from Marey’s Photography to the Latent Vector

Abstract

This article explores the operational status of contemporary images linked to machine learning, revealing their historical continuity with data extraction methods developed in 19th-century scientific photography. The analysis begins with the experiments by Gilles de la Tourette and Étienne-Jules Marey, who transformed bodies into diagrams, eliminating the figurative to isolate patterns of movement and behaviour. These techniques, designed to transcend human perception and access "objective" knowledge, established a paradigm where the image functions as a surface for data inscription. The study demonstrates how this paradigm has reached its maximum expression in contemporary artificial vision systems, where images have ceased to be fundamentally objects of perception and become instruments of automated extraction. Unlike 19th-century photographic records − explicit in their technical mediation −, data registration now operates in opaque computational layers, shifting from the phenomenic to the statistical plane. Through the analysis of latent spaces in deep learning models, we show how images are reduced to mathematical vectors that encode both visual information and power structures. This transformation operates on three levels: technical (data compression into dimensions that are not humanly legible), epistemological (knowledge production through mass correlations), and political (naturalisation of social categorisations through seemingly neutral interfaces). The research concludes that this regime of post-retinal visuality inverts the traditional logic: the visible becomes an accessory residue, while real agency occurs in the calculation of the imperceptible.