Machine Learned Landscape 01
Kıvanç Tatar and Adam Basanta, 2019
premiered at gnration, as a part of Index Media Arts Festival
Braga, Portugal
Landscape Past Future is Adam Basanta’s previous work where he remixed Dutch Landscape painting using algorithmic approaches, to explore what makes a painting of a landscape, and question the ownership of the culture through the means of remix. Adam created thousands of “mosaicked” variations these remixes during the production of Landscape Past Future. In this collaboration with Adam, I have delved into Deep Learning algorithms to further question what a landscape painting is in today’s digital world.
I trained a type of Variational AutoEncoder (VAE) to attempt to capture the essence of the landscape paintings. In Computer Vision research, VAE based algorithms have been labelled as creating blurry images. However, my experience with VAE algorithms was that they are handy for capturing the shared properties of a dataset, in comparison to grasphing the fine details. Hence, I imagined that VAEs, maybe, could capture the essence of Adam’s dataset, as well as Dutch landscape paintings. I generated several videos using VAE networks that are trained on subsets and whole dataset of landscape painting remixes . I think, this particular application is an example of how Creative AI diverts from the idea of optimality and efficiency.
In my further experiments, I worked with another type of Deep Learning algorithm based on Convolutional Neural Networks. This second type is known in Deep Learning as the style imitation networks that are based on well-known VGG19 network. In the style imitation applications, the network copies the styles of edges, color gradings, color profiles, lines etc. from one image and applies that style onto another image. I have used five of Adam Basanta’s Landscape remixes as examples to create five distinct image styles using Deep Learning. The video files generated by VAE networks then further processed using these five distinct styles. All came together as a video piece.