The Neural Code and Dreaming

A machine can now predict what you dream about.

Posted Apr 21, 2013

Dream research is bedeviled by its dependence on the subjective verbal report of the dreamer. We cannot independently verify if a person has been dreaming. Alls we have is his report that he has in fact had a dream. If we see the characteristic brain activity patterns of REM sleep on an EEG machine then our confidence that that person is dreaming is increased but never absolutely certain. When people are awakened from REM they do not always report a dream. Nevertheless, most people, most of the time, report a dream when awakened from REM sleep.

Even if we grant that a person is likely experiencing a dream when he or she enters REM sleep we have no idea what he is dreaming about. To find out what people dream about we have to ask them. Once again we are totally dependent on subjective report if we want to study dream content.

It would be nice if we did not have to ask the person what his dream was about. If instead we knew the basic neural code with which the brain processes cognitive content then we could simply consult the code displayed in brain activity patterns and then derive the content of the dream. Of course we are a long way from deciphering the neural code the human brain uses to represent cognitive content BUT a recent publication has moved us one step closer to doing just that.

Horikawa et al (Science, Vol 340, 4 April 2013) recruited 3 volunteers to study their brain activity (as measured by fMRI and EEG) during the sleep transitional state (S1) when many people experience hypnagogic hallucinations. When the volunteers transitioned into this S1 state they were awakened and verbally reported their visual experience during sleep. fMRI activity patterns were used as input into a computer simulation program that treated the fMRI patterns as code for words representing visual objects. Those visual object codes were labeled with words derived from the subject’s verbal description of his visual experience during sleep. For example if the person reported that he had seen a red house while the fMRI displayed a pattern X in the visual cortex then the machine learned that the pattern X or a pattern similar to X always indicated “red house”.

Once the machine learned a large number of these pattern-object associations it could predict what the subject was seeing during sleep based solely on fMRI patterns. The researchers did not have to ask the subject to report his visual experience. The machine could do that (at least 60% of the time) if the current fMRI patterns matched those in its repertoire of learned image-object pairings.

There are several significant implications and questions for dream research…First, when Horikawa’s volunteers confirmed that the machine was largely correct in its predictions of what they were seeing in their sleep it implied that we will eventually be able to look at neuroimaging records of sleep states and be able to tell what people dream about.

If, someday, we look at thousands of these neuroimaging records from a large group of subjects and then put together a summary of what these people dream about, we can then collect the associated dream reports from these people and compare the reports to the neuroimaging records. If that comparison matches up nicely (beyond what you would expect based on chance) then we can more easily trust people’s verbal reports concerning their dream content. We can be increasingly confident that people’s reports concerning their dreams are usually non-deceptive, veridical or true.

Second, as our knowledge of the impact of dreams on waking behavior increases we will be able to treat people with painful dream disorders (like repetitive nightmares) more effectively. If, for example, monster X always appears with fMRI pattern Y and drug z eliminates pattern Y from the fMRI AND the patient reports relief after treatment then we can infer that that dream image and its associated brain pattern was indeed causing the distress.

Third, as dream science begins to identify certain recurrent dream content images as strong predictors of waking behavioral patterns, the machine analysis of these dream images can be used to predict waking behavioral patterns.

Fourth, it will be fascinating to compare brain signatures for recurrent dream images with their waking counterparts. Does seeing a red house in a dream require the same neural signature as seeing a red house during waking life? Horikawa’s results suggest the answer is Yes but that might not be the case when we move away from the catalog of simple visual images studied in Horikawa’s subjects.

Fifth, the interesting content of dreams most often concerns emotions. Emotions are associated with neural signatures. Can the machine learn to predict emotional content of dreams based on neural signatures?