*Blade Runner—Autoencoded* is a film made by training an autoencoder—a type of generative neural network—to recreate frames from the 1982 film *Blade Runner*. The Autoencoder learns to model images by trying to copy them through a very narrow information bottleneck, being optimized to create images that are as similar as possible to the original images. The network was trained on all the frames from the *Blade Runner* 20 times. After training, the autoencoder reinterprets each frame from the film in order, and the reconstructed frames are then resequenced back to create a reconstruction of the film. The resulting sequence is very dreamlike, drifting in and out of recognition between static scenes that the model remembers well, to fleeting sequences—usually with a lot of movement—that the model barely comprehends.

The film *Blade Runner* is adapted from Philip K. Dicks novel *Do Androids Dream of Electric Sheep?*. Set in a post-apocalyptic dystopian future, Rick Deckard is a bounty hunter who makes a living hunting down and killing replicants, artificial humans that are so well engineered that they are physically indistinguishable from human beings. Because replicants are indistinguishable from humans, Deckard has to issue the Voight-Kampff empathy test in order to determine whether they are humans or not. The technological advances of the Nexus-6 replicants makes it increasingly difficult for Deckard to determine what is human and what is not, and Deckard himself has the growing suspicion that he himself may not be human.

By reinterpreting *Blade Runner* with the autoencoder’s memory of the film, *Blade Runner—Autoencoded* seeks to emphasize the ambiguous boundary in the film between replicant and human, or in the case of the reconstructed film, between our memory of the film and the neural networks. Aspects of the flaws in its visual reconstruction are reminiscent of the deficiencies of our own, especially regarding memories of dreams. By examining this imperfect reconstruction of *Blade Runner*, the gaze of a disembodied machine, it becomes easier to acknowledge the flaws in our own internal representation of the world and easier to imagine the potential of other, substantially different systems that have their own internal representations.

Links: http://terencebroad.com/autoencodingbladerunner.html

Carried out on the Msci Creative Computing course at the Department of Computing, Goldsmiths, University of London under the supervision of Mick Grierson.

Terence Broad (UK) is an artist and machine learning researcher based in London. He works at the forefront of technological developments in machine learning, exploring both the perceptual capabilities and limitations of these techniques. He graduated in 2016 from the Creative Computing Masters programme at Goldsmiths, University of London. His work has been exhibited internationally at venues including The Whitney Museum of American Art, Art Center NABI, and The Barbican.