Machine Learned Landscape 01

Single channel video. 2019.
Duration 8:41. In collaboration with Kıvanç Tatar.

During the production of prints for the exhibition Landscape Past Future, thousands of “mosaicked” variations were created, re-mixing masterworks of the landscape painting genre to create original images using algorithmic and machine learning techniques.

Nearly 5000 unused variations were then used as a corpus to train several neural networks which attempt to learn both the compositional structure of the images and the overall style. The networks then generates its own variations on the corpus, creating an ongoing sequence of speculative compositions.

The iterative process of feeding images created using ML techniques in to a neural network is in itself a barrier to the system’s understanding: the under-training of this network with a relatively low number of “mosaicked” images is difficult for a machine learning system developed primarily for image recognition purposes geared towards research and (eventually) commercial endeavours.

Rather than witnessing technologically supremacy, we are bearing witness to an ongoing performance of barely understanding a notion of landscape. What is a landscape, at bare minimum? The meeting of the ground and a sky stretching from the horizon; the occasional feature in the forefront coming into focus, shifting, disappearing as soon as it manifested into being. Learning, even when one learns iteratively from their selves, is a process that never ends.

Completed with support of the Canada Council for the Arts