The Neural Net That Recreated 'Blade Runner' Has the Movie Stuck in Its Memory

Artist and machine learning engineer Terence Broad's Auto-Encoding Blade Runner is the project Philip K. Dick would have made if he were a scientist. In his presentation at SIGGRAPH 2017, a computer graphics and animation conference, Broad detailed how he trained a Convolutional Autoencoder--a type of neural network--to recognize patterns of data in Blade Runner and then reconstruct it, scene by scene. What results is an eerily-accurate full-length film that was so convincing that Warner Bros. issued a DMCA takedown notice to Vimeo when Broad first uploaded the footage in 2016. The real question behind Broad's Blade Runner parallels the themes of Philip K Dick's legendary novel Do Androids Dream of Electric Sheep: Where does one draw the line between human and machine, the real and the seemingly real? "I think the thing to understand about neural networks is that we don't really know how they work," Ruth West, SIGGRAPH's chair of art papers, told me in an interview. "They're black boxes, and they make these leaps that are kind of like the leaps we make internally.