The fundamental question about the function of sleep is why it needs to happen while we're asleep. In other words, why cannot whatever benefit sleep confers happen while the organism is awake and alert. Most theories of sleep function fail to address this fundamental question.

REM sleep

     Learning is the art of making predictions based on experience. Predictions are useful insofar as they either have direct value to the organism (association with more or less hunger, thirst, sexual gratification, pain, etc.) or make predictions about further consequences. Thus, often the greatest value of an association learned is not in the actual association experienced, but rather in its consequences for other anticipated scenarios, which were not experienced. The present theory states that sleeping is a mechanism to extract the consequences or predictions of what was learned for such anticipated scenarios. In order to discover these consequences, namely how an anticipated scenario would play out given the new knowledge acquired, the brain needs to simulate it. Since the brain does not have a spare simulator, this needs to happen while the brain believes it is in those scenarios. This explains the hallucinatory aspects and sensory disconnection of sleep, as well as the effect that its disruption has on learning.

      I call this the Dress Rehearsal Theory of Sleep Function, because it postulates that, like a dress rehearsal before a theatrical performance checking the effect of any changes in lighting, costumes, music or actors on the rest of the performance, dream sleep performs a similar verification of the effect of changes learned during the day on related memories. Both require the full theater --just like you would not expect to run a performance at the same time as a dress rehearsal in a theater, you cannot expect to perform the function of sleep in the theater of the mind while the person is awake.

     At the Humanity+ 2009 Conference, a question by Bill Lauritzen brought up a prediction of the theory, namely that, although it's well known that the formation of episodic memories is turned off during sleep, the Dress-Rehearsal Theory of REM Sleep predicts that other forms of learning will be active during sleep, for what good is it to simulate consequences if you will not remember them?

      N.B. Reading my friend Mariano Sigman's excellent book, El Breve Lapso entre el Huevo y la Gallina ("The Brief Lapse between the Rgg and the Chicken", a highly recommendable collection of stories about science), today I found a related idea in one of his chapters. So it came as some relief when he wrote that no ideas are original in the same book. And of course, Matt Wilson's work at MIT has long shown that the hippocampus replays brain activity of the wake state during sleep. That said, I think the concept above is different and more specific than Mariano's and the work I know from Matt: while Mariano suggests that sleep serves to simulate the events of the day and also to try out scenarios, my hypothesis is that we do not simply rehash the events of the day (that would not explain why we need to do it again rather than learn during the actual original awake occurrence) nor a simulation of random scenarios, but rather specifically computing the consequences of what was learned during the day for other bits of previously acquired knowledge, a process that I claim requires the brain to simulate both the day's lessons and its implications, perhaps by simulating potential consequences and probing their compatibility with the day's events or the day's events implications about them, something that cannot be done live (as during wake life we need to deal with reality, and cannot afford to have the brain wander off to ponder all the consequences of what we encounter) and thus requires the brain to turn behavior off --namely, sleep.

Non-REM sleep

      The function of non-REM, or slow wave sleep, remained elusive to me until after hearing a talk at Caltech this month by Giulio Tononi, the brilliant Italian scientist who, with Nobel-Prize winning Gerry Edelman, put forth what is perhaps the best quantitative measure of some of the properties of consciousness. Although his talk did not really address sleep more than in passing, after hearing it, and on my drive to have breakfast with him at Caltech's Athenaeum the following morning, it occurred to me that the function of slow-wave sleep, when neuronal connections become much more indiscriminate, could only be one of sensing averages across neuronal populations, and thus one of normalization. When talking with Giulio half an hour later, he confirmed that was his hypothesis too, and that he has published that much. He thinks without such normalization, learning would drive synapses to saturation. Why does such a normalization require turning the brain "off"? Because it requires changing the brain's connectivity pattern, which underlies its normal function in the wake state. Still, this theory has the unsatisfying aspect that it seems that shut-down during a third of its life sounds like a terribly high price to pay for normalization, so I suspect something more local and complicated is going on during slow wave sleep than global synaptic normalization.

Alex Backer, 10/28/2008

P.S. The theory on the function of REM sleep above was arrived at while sleeping during a night in Australia in the Southern winter of 2006.