The dream-reading machine

The film Inception starts with Leonardo DiCaprio and Joseph Gordon-Levitt attempting to infiltrate someone’s subconscious. They are trying to steal the target’s dreams. This wonderfully futuristic concept may be a thing of science-fiction movies, but researchers in Japan might just be on the road to seeing what you dream.

Research carried out by Yukiyasu Kamitani’s group at the Advanced Telecommunications Research Institute in Kyoto, published in the journal Science, used an fMRI (functional magnetic resonance imaging) brain scan to monitor volunteers’ brain activity whilst they drifted off to sleep. By creating a computer algorithm to predict what this brain activity meant, they were able to predict what a subject was dreaming about.

To begin, 3 volunteers were placed into an fMRI scanner and shown Google images of many different objects. The activity in the visual areas of the brain was monitored by the scanner and uploaded to a computer. Words associated with each of theses images were processed and arranged into groups of like-meaning words, called synsets. For example, words such as Structure, Building, House and Hotel would be grouped together in their own synset. Words within a synset were ranked depending on their importance, and the most important was used to describe that synset. For the example above, the word Building would be the highest ranked word in that synset. This allowed them to narrow down a number of possible words/objects to one word of ‘best-fit’.  Images of houses, hotels, offices, would all be narrowed down to Building.

The computer was given information of the synsets that related to each image, along with the brain activity at the time that image was shown. This allowed the researchers to match brain activity to certain images and words. The computer now knew that when the subject saw a picture of a house, their brain responded in a certain way. This brain activity was grouped together with activity when the subject sees an office, and a hotel etc.

Now came the real test. By categorising brain activity based on what a person sees, could they read what a person was dreaming about? The 3 subjects were placed in the scanner and told to fall asleep if they felt tired. The electrical activity of their brain was recorded by EEG (electroencephalogram) in order to see when they fell into the early stages of sleep. During these early stages, one does not normally have vivid ‘dreams’ but typically light hallucinations.

As these hallucinations started, the brain activity in visual areas was recorded and run through the algorithm. The algorithm came up with the synsets that were most likely to be represented by that brain activity, and used the Google images from before to present a video of what it thought the person was ‘dreaming’ about. This can be seen in the video below. To test how accurate the computer was at predicting the ‘dreams’, the volunteer was awoken and asked what they had just seen.

When the researchers compared what the participants reported they were seeing with the computers prediction, they found that the computer was correct in 60% of cases. This is significantly higher than getting it right by chance. The computer was able to use brain activity during the early stages of sleep to read and predict what the volunteer was seeing.

This study is not without its limitations. Firstly, what most of us see as ‘dreaming’ is not thought to occur in these early stages. We believe ‘dreaming’ occurs mainly during rapid-eye movement sleep, a stage of sleep that occurs around an hour later than these early stages of sleep (see below). What are measured here are hallucinations that occur when we are falling to sleep. Furthermore, an fMRI machine is incredibly loud due to its large, spinning magnets. It is questionable that the sleep stage observed in these participants is truly what we would regard as sleep.


A typical sleep cycle. Researchers recorded hallucinations in stages 1 and 2. We normally dream in REM (rapid eye movement) sleep. Image credit to Sleep 1102.

Secondly, a success rate of 60% is hardly news to excite those wanting to perform dream extraction. The crude prediction is not an exact match of what someone is seeing (as you can see from the video above). The computer is able to recognise that you were seeing a building, but not that you were cleaning windows of your own house for example. It is clear that it will take some time to really enter the realms of dream-reading. The interpretation of this crude prediction is also hampered by the fact that the study was based on only 3 participants. It is not clear whether this result will scale up to the larger public.

Despite these limitations, what the researchers have done is remarkable. They have shown that these early sleep hallucinations create very similar patterns of activity in the brain to when we are awake. They have shown a relatively accurate way to decode this activity into what the subject is seeing. And they have opened up the possibility of studying the function and nature of sleep in more detail. But don’t worry; the Thought Police won’t be after you just yet.


By Oliver Freeman @ojfreeman


4 thoughts on “The dream-reading machine”

  1. Howdy! I simply want to give an enormous thumbs up for the
    good info you have right here on this post. I will be coming
    back to your weblog for more soon.

  2. Stumbled across your blog while doing some research for a new dream sharing app we’re developing called SHADOW ( Really interesting stuff! Drop us a line if you’re ever interested in learning more or collaborating!

  3. Howdy very nice blog!! Guy .. Excellent .. Wonderful ..
    I will bookmark your site and take the feeds additionally…I am satisfied to
    find numerous useful information right here in the put up, we want
    work out more techniques on this regard, thank you for sharing.

Comments are closed.

Share This