Approaching Destination (2020)
3 channel Gan generated video
Latent spaces are the inner worlds of neural networks. In a generative adversarial network, or short “GAN”, every multi-dimensional coordinate in such a space translates to a unique image. By travelling between these locations one can explore these spaces and discover real and unreal views.
In my work an algorithm tries to find the painting “Wanderer above the Sea of Fog” by Caspar David Friedrich in the latent space of BigGAN, a model released by Google researchers, which has been trained on millions of real-world photos in 1000 different categories. I call this model a “public latent space” since it is open to everyone who has the technical means to run it. This model has never been trained on this or any other painting. By using a method called “gradient descent” the algorithm nevertheless tries to approach this target and to get as close as possible to it within its means, starting from different locations in the latent space.
On every step in the process the machine has to determine if an image it has generated has brought it closer to or further away from its destination. With the help of mathematical “loss functions” it has to calculate how similar two images are and then try to move in the direction that promises to increase this similarity. However, the landscapes in latent spaces are littered with an endless sea of high mountains and deep valleys which make it impossible to see the entire panorama at once and so every search is also like a gradual groping through the fog, trying to orient itself by only perceiving its immediate surroundings.
I see my work with AI as a journey in which I might set out with a certain goal in mind but do not always arrive at the destination I had hoped to get to. But sometimes the places I end up are more interesting than those I had imagined.
Mario Klingemann is an artist who uses algorithms and artificial intelligence to create and investigate systems. He is particularly interested in human perception of art and creativity, researching methods in which machines can augment or emulate these processes. Thus his artistic research spans a wide range of areas like generative art, cybernetic aesthetics, information theory, feedback loops, pattern recognition, emergent behaviors, neural networks, cultural data or storytelling.
He was winner of the Lumen Prize Gold 2018, received an honorary mention at the Prix Ars Electronica 2020 and won the British Library Labs Creative Award 2015. He was artist in residence at the Google Arts & Culture Lab and has been recognized as a pioneer in the field of AI art. His work has been featured in art publications as well as academic research and has been shown in international museums and at art festivals like Ars Electronica, the Centre Pompidou, ZKM, the Barbican, the Ermitage, the Photographers’ Gallery, Colección Solo Madrid, Nature Morte Gallery New Delhi, Residenzschloss Dresden, Grey Area Foundation, Mediacity Biennale Seoul, the British Library and MoMA. He is represented by Onkaos, Madrid and DAM Gallery Berlin.
Generative adversarial networks (GANs):
It is difficult to train an artificial neural network to be good at a task such as making images. One successful technique is to train two networks simultaneously with one learning how to make new images which look like those in the data set, and another learning how to tell the new ones apart from the original ones. The two networks are both bad to start with, but compete against each other, getting very good in the process. At the end, the generative network can be used on its own to produce art.