Google fed their image recognition neural network images and instructed them to enhance whatever objects were recognised in the picture, resulting in some pretty weird stuff. Then they just started just inputting random noise and feeding the result back into the system over and over, ending up with some really interesting images.