Seeing as a Computer with Google Deep Dream

One step in the pursuit of artificial intelligence is to have computers recognize features within images. Facebook has a technology called DeepFace to recognize faces in the photos its users upload in order to suggest tags, for example.

Google recently shared its artificial intelligence program for image recognition called Artificial Neural Networks. The program is “trained” by looking at a huge collection of reference images that show the shape of buildings, or trees, or bugs, or dogs, etc. When you feed a new image into the program, the program puts that image through a series of layers. Each layer looks at a certain aspect of the image (e.g. edges within images, or orientation of shapes within images). The image is passed through these layers one-by-one and by the final output layer the program should have an “answer” as to what features are included within the image.

The Google Artificial Neural Network can take image recognition to strange new levels, though, if engaged in a loop. Designers at Google took a random image and asked the program to identify features within that image. Once identified, the program was asked to enhance that feature. (So if the program was given an image of a cloudy sky and one of the clouds looked a little like a bird, the program would make that cloud look more like a bird.) That newly enhanced image was then run through the program again so that the feature (e.g. the bird) became even more pronounced. The program repeated this over and over.

Google released this program under the name Deep Dream. Which is kind of a cool name, and evocative, but most of the applications look more like nightmares. The below video was created by starting with an image filled with random noise. Every four seconds a new layer of Google’s program is engaged to recognize and enhance features within the image. But the image is also constantly zooming in, so the image keeps changing and morphing in a nightmarish, psychedelic way. The video is silent so maybe queue up some Pink Floyd before starting. It really gets weird at about minute 1:30.

Inside an artificial brain from Johan Nordberg on Vimeo.

“What does this have to do with libraries?” you might ask. The technology could kick off a conversation about the ethics of artificial intelligence, I suppose. And it may also serve as an indication of the direction of the future Internet: towards automatic metadata description and (one could guess) hooking that metadata into a larger linked-data-type environment that continues to grow and describe new objects.