Researchers at Google have been developing a computer software program that may shed new light on certain aspects of dream formation. Titled “Inceptionism: Going Deeper into Neural Networks,” the Google project involves the clever use of computer networks designed to recognize images. The engineers created a program that could “learn” how to identify various kinds of images with increasingly greater precision and accuracy. Through millions of trial-and-error attempts, the program developed the ability to determine which features of each image are most likely to represent a particular object.
The engineers then put this program to novel, “what would happen if…” uses. In some cases they gave the network an image and asked for an amplified analysis of some specific feature of the image:
“We ask the network: ‘Whatever you see there, I want more of it!’ This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird. This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere.”
(Included with their announcement was an “Inceptionism gallery” with some of their initial images, to give a sense of what they are talking about.)
They go on to comment that this kind of narrowly iterative process “can be used to over-interpret an image.”
A further level of experimentation involved letting the network run free, with no particular inputs at all:
“If we apply the algorithm iteratively on its own outputs and apply some zooming after each iteration, we get an endless stream of new impressions, exploring the set of things the network knows about. We can even start this process from a random-noise image, so that the result becomes purely the result of the neural network…”
Google has made the code for this process public, posting it on Github for free download. The technical requirements are considerable for putting the program to use, but already hundreds of people have created fascinating images that are posted under the hashtag #deepdream.
The implications of this technology are potentially enormous for dream research. The Google engineers mentioned the possible relevance of their network for illuminating “the roots of the creative process in general,” which is ambitious but seems justified given the intriguing parallels between artistic innovation and the processes modeled by this kind of computer program.
However, an even better case can be made for the relevance of the Google network for the science of dreaming, as the #deepdream hashtag seems to recognize. The ability to explore the self-organizing activities of networks at different levels and scales of abstraction holds enormous promise for helping researchers understand what people typically do and do not dream about. One of the central mysteries of dream research is explaining the paradox of intense, often highly detailed visual perceptions in dreaming while the eyes are closed, external sensory input is shut off, and the mind descends into deep sleep. The Google program may contribute to efforts to tease out which specific features in “dream vision” are related to amplified versions of waking vision.
A caution about this kind of research: it does not provide evidence for the idea that dreaming is merely random neural nonsense, which many people assume is the proper conclusion to draw from the use of new technological tools for studying dreams. The Google project is exploring one particular area of overlap between its network and dreaming experience, namely the identification of visual images. There are many, many other ways in which dreaming is not like the Google network. For one thing, dreams are not just images; they typically also involve feelings, thoughts, interactions, movements, memories, conversations, dramatic sequences, and many other aspects of form and content that lie outside the range of this computer program. For another thing, dream formation is influenced by a complex, dynamic, and interactive array of “inputs” from each individual’s life including physiological processes, personal experiences, social activities, and cultural meaning systems. Perhaps future computer networks will incorporate more of those elements into their algorithms, and then we can begin talking about contributions toward a general dream theory. But for now this kind of research is best understood and appreciated in a more focused way, as an exciting new method of studying the creative powers of visual imagination in dreaming.