What is a "deep dream"?
Deep Dream of Electric Sheep by Calhoun Press. Original image by 3268zauber, CC-BY-SA.
"Deep Dream of Electric Sheep." From an image by 3268zauber, CC-BY-SA.
"Deep dream" images are made by artificial neural networks (a kind of computer program) that seek out and enhance patterns that they "see" in images.
Google uses artificial neural networks to find and sort images on the web. For example, whenever someone searches "sheep," they will get pictures of sheep, not goats or bushes or corgis or anything else, because of these networks.
Google also uses the surrounding text on a webpage to sort images, but that's not enough. For example, a page on sheep herding might feature pictures of sheep in a field, a border collie, a shepherd, and a picture of the author of the post. Artificial neural networks allow Google to discriminate between a picture of an eggplant and a picture of a human face.
How does the network know what a sheep looks like? Basically, the network is fed thousands of images of sheep and then looks for commonalities between the images. Each layer of the neural network looks for different things; the shallower layers look for patterns of light and shadow, while deeper layers look for shapes and more complex visual structures.
A few engineers at Google discovered that these artificial neural networks could not only recognize images, they could also generate them. If Google gave these networks an image of random noise, and then told the network to look for and amplify the sheep-like patterns in the image, these engineers could see what the network had identified as the defining visual characteristics of sheep. This process would allow the engineers to understand how the neural networks were processing images, and what each layer was "seeing."
The resulting pictures, called "deep dreams" by Google, looked a lot like the visual hallucinations of people on psychedelic drugs. Is that a coincidence, or are artificial neural networks giving us insight into how vision is processed in the human brain? It's too early to say, but computer and cognitive science can now "see what the machine sees." Where it goes from there is beyond my paygrade, but it's always interesting to see what happens next in this field.