Searching for Dream Logic in Google’s DeepDream

by Amanda Beth Peery

DeepDream ArchesIn this image produced by Google's DeepDream algorithm, a pattern of archways stretches across the screen in an Escher-esque eye-game, the intricate façade of a building with impossible architecture. This vision is the result of a computer program, a neural net that has been trained to recognize objects and, when looking at an image, amplify any hint or outline of an object it knows. So, in other DeepDream images, two dots might become two dark eyes or a leaf might become a dog's face. Here, DeepDream recognizes arches everywhere, and the original image is lost in arches that nest within and interrupt each other.

But what kind of meaning can we find in this image? Is there something essentially human—or inhuman—about it? In some ways, this image and others produced by DeepDream look like the hallucinations or dreams we know, but do these AI-produced images really mirror human visions? One way to think about this is to see DeepDream's images in the context of dream-art, art that follows the logic of dreams. The logic of human dreams and dream-art is different from the rational logic of our waking lives. Dream logic is what gives dreams a meaning of their own, making them more than a film-reel of leftover, twisted images from the day.

Many great artists, from the Lascaux cave painters to Tarkovsky to Lewis Carroll, created images and stories that follow dream logic. Dream logic is not arbitrary, and it's not an empty absurdity. One thing that makes a work of art like Alice in Wonderland powerful is that we, readers and viewers, recognize the logic of the story as the structure of our own dreams. The symbols, and the connections between them, feel coherent and right. Of course a vial of liquid or a little cake cause Alice to grow or shrink. Of course the Cheshire Cat's smile lingers in the tree. These are recognizable absurdities.

DeepDream's images do not contain dream logic because they do not understand the images they reproduce, so they cannot deal in the deeper meanings of the objects. They cannot see the objects arise, transformed, within the world of dream logic. They cannot see the objects through the looking glass. Many people who have looked at the DeepDream images compare them to drug-induced hallucinations. I do not doubt the similarity. But there is a difference between hallucinations and dreams, and this difference is important. Hallucinations are eye-games, tricks on the surface of our world (and across our screens), while dreams and dream-art immerse us in another reality, structured by an alternative logic.

In Werner Herzog's movie Fitzcarraldo, a missionary in the Amazon tells the protagonist, Fitzcarraldo, about his attempts to convert native people to Christianity. He says, "We can't seem to cure them of the idea that our everyday life is only an illusion, behind which lies the reality of dreams." Fitzcarraldo responds, "Actually, I'm very interested in these ideas. I specialize in opera myself." To him, opera is self-evidently like dreams. Maybe he means that opera follows the same poetic logic as dreams, as opposed to the logic of "everyday life." At a different point in Herzog's movie, Fitzcarraldo describes opera by saying that it "gives expression to our greatest feelings." This is another way that opera is like dreams. Sometimes in dreams we feel a depth of terror or joy that we rarely feel in life. Art strives to express this depth of feeling.

It might be useful to classify the pattern of archways in the Google image as a hallucination rather than a dream. That does not mean that the archways are uninteresting. Hallucinations are interesting in their own way, and this hallucinatory image is interestingly resistant to meaning.

Because DeepDream propagates and amplifies patterns it knows, it gives rise to symbols without meaning, repetition without the subtlety that is essential to aesthetic or emotional depth. But in this image, the repetition of archways does contain some nuance, some fascinating differences between arch and arch, and maybe that is enough to create meaning.

Near the lower left of the image, the archways open into dark, painterly depths, shadows that promise an endless hallway of dark arches. And towards the lower right of the image, there's a spark of what looks like sunlight and green growth behind a closed gazebo. Just above the gazebo, above another row of tiny arches, there's a dark smudge that could be a cross or a man perched on the edge of a balustrade looking down.

The philosopher Walter Benjamin spent a long time searching for the meaning of the arcades in Paris, eventually producing a volume of quotes, collected fragments circling the meaning of the arcades. The quotes explain the architecture of the archways, the symbolism of commodities such as those sold in the arcades, the politics of indoor and outdoor markets. By collecting knowledge that helped define elements of the arcades, Benjamin tried to pierce through the essential meaninglessness of the external world.

The Google image of archways is like Benjamin's book. It shows archways from every angle, composed in every way, with meaning just beginning to crystallize or leak through. I imagine that if the image didn't stop at the edge of the screen it would go on forever—an infinity of impenetrable archways, portrayed from all sides.

Benjamin's book, The Arcades Project, was published posthumously, with many of the quotes arranged by an editor. If Benjamin hadn't died, it, too, might have kept going on. (Maybe readers today write their own quotes into the margins and, in this way, the project continues). If Google's hallucinatory image were created by a human artist, it might be read as a desperate attempt to find meaning in the meaningless. Dreams contain symbols oversaturated with meaning, while this hallucination is oversaturated with objects—archways—devoid of meaning. It is appropriate, then, that there is no human mind at work here, but only a program originally designed to help identify and grasp objects. Rather than find the true nature of the arches, or of reality, it can only reproduce the object in its many forms.