Cloud Hosting

MK: Whats interesting is that in Anadols work, older and newer technologies come together in a kind of collage. He customizes a GAN, which is not even the most advanced type of generative AI now, and combines it with different types of rendering software and new mapping algorithms. And he does this in order to play with the degree of efficacy of these different algorithms: to explore different degrees of human intervention and control and different degrees of letting go, letting the machine conjure its own interpretations of the archive and of human memory.

The title of Unsupervised is very specific: it refers to a technical term for a type of machine learning that Anadol and his studio use. Most machine learning is supervised learning, where the AI needs to try to classify the information that is before it. (Even that is still difficult: We have autonomous cars, but they still cant distinguish between the moon and a stop sign.) In supervised learning, humans tag images, for example, or bits of information, in order to train a machine learning model. So, I would go through the data set, and I would tag an image of a pen with the word pen.

Unsupervised learning, on the other hand, is where the machine does the tagging itself. Its a whole other kind of black box where the machine is actually deciding not only how to tag something, what kinds of properties something should be classified as possessing, but its also deciding in many ways what is meaningful, what is of value, in terms of information. This is already opening up a kind of agency on the part of the machine that is very different from traditional processes of supervised learning.

Anadol is using unsupervised learning so that the work can actually generate something new based on its learnings, rather than just classify and process. And then the artist is in his studio working with this model, almost like an electronic musician with lots of different dials in front of him, adjusting what kinds of learning takes place, the rate at which the learning takes place, thousands and thousands of parameters around what kind of forms it could generate.

But at the same time, theres a huge gulf between that stage and getting to the point of creating something that looks the way the works in Unsupervised do. Theres so much intervention and, in fact, human collaboration. Because with machine learning, you often might get noise. It doesnt necessarily generate anything that we find meaningful or that we actually could perceive. And so, Anadol is working in concert with this quasi-organic, changing, adaptive systembut also guiding it away from what it might normally optimize, or think, to produce a series of morphologies that is unpredictable but neither simply a chance occurrence, nor fully automated. Theres an interplay between probability and indeterminacy.

And then, layered on top of that, in two of the works, is a diagram, a visualization, of the AIs movements in space, moving throughout that complex map or galaxy it has constructed based on everything its learned, and classified, and clustered according to patterns of affinities. Again, these are affinities that we may never even perceive or think of, building a very complex mapin this case, literally 1,024 dimensions. This is not perceivable by human eyes. But what Anadol is doing is creating a map of movement that we can perceive, either as a network of shifting, connected lines, or as four-dimensional fluid dynamics, which looks like a rushing waterfall. Its almost as if youre watching a dance unfold in real time, but the choreographic score is being overlaid on top of the dance.

Read the original:
Deep Learning: AI, Art History, and the Museum | Magazine - MoMA