Neural optimization in sleep

Jump to: navigation, search

This text is part of: "Science of sleep" by Piotr Wozniak (2017)

Sleep as a neural optimizer

All sleep theories that pass the shutdown test are compatible with what seems to be the chief function of sleep: neural optimization.

A series of theories indicate that sleep is a neural optimizer. Neural networks get rewired overnight. Memories move from a temporary low-capacity fast-encoding fast-association low-interference storage to areas where they can safely be used for months and years without much interference from newly encoded memories. Abstract patterns are extracted, while details obscuring the big picture are discarded.

Sleep and memory

NREM and memory

NREM sleep has magic powers! If you fall asleep for just a couple of minutes and manage to enter the deeper stages of sleep, you are likely to wake up with a brain that feels like brand new. Obviously, a short nap of that sort is only possible when it is properly timed and when awakening is natural. However, the impact of a short bout of NREM on learning is staggering. It takes many hours of heavy learning to make a brain homeostatically sleepy. It takes minutes of NREM sleep to take that sleepiness away.

Two-component model in SuperMemo shows how a 19 min. nap can nearly double the homeostatic component of alertness (green line)

Figure: Two-component model of sleep in SuperMemo shows how a 19 min. nap can nearly double the homeostatic component of alertness (green line). This theoretical model is backed by years of sleep and learning data.

This effect is so powerful that the whole myth of Uberman sleep was grounded on that foundation. Even a few minutes of sleep can provide sufficient refreshment for the brain to continue working for at least a few hours. Obviously, the power of NREM is only a fraction of the big picture. However, in this short section I would like to peek at what we know about the effects of NREM on memory.

NREM and hippocampal cleanup

The most convincing hypotheses on the function of NREM sleep depict it as a process in which the short-term novel memory stored in a temporary storage (primarily the hippocampus, the entorhinal cortex, and the adjoining structures) is written down optimally into the vast networks of the neocortical storage (Born and Marshall 2007[1]). The hippocampus connects various areas of the brain to form quick memory associations. For example, when we learn a new French word, it might provide a link between the concept or its expression in English with the new French sound, text, or a set of syllables. That hippocampal memory switchboard is, naturally, limited in size. This is why it needs to be emptied periodically, so that the new connection between the English and French words could be laid down at slower and more complex cortical storage. The hypothesis says that this happens in NREM sleep, and it happens pretty fast. Scientists noticed long ago that during slow-wave sleep, hippocampal and neocortical networks tend to replay the firing patterns associated with novel experience (Wilson and McNaughton 1994[2]). This led to the suspicion that the hippocampus might be rehearsing the cortex with the newly learned information, incl. the new French vocabulary. To understand how such training might be done, one needs to know more about the properties of neural networks and how they fire in synchrony with theta/gamma oscillations during explorative activity (e.g. learning) and SWP/200 Hz ripple bursts during consummatory activity (deep sleep)(Buzsaki 1996[3]).

NREM and declarative memory

Research into NREM sleep and memory is contradictory and does not fully align with the neat picture painted above. Everyone can see the power of sleep in SuperMemo's SleepChart, where even a short nap can bring recall and consolidation in learning back to the baseline. Sleep spindles associated with NREM sleep correlate well with the degree of feeling refreshed (Goetz et al. 1983[4]). This says little about the function of the hippocampus, and inner working of the brain in sleep, let alone NREM sleep. However, Takashima noticed that the duration of naps correlates positively with later memory performance, and negatively with the activity in the hippocampus registered at retrieval (Takashima 2006[5]). This might indicate that even a short nap can reduce the hippocampal memory load. Other researchers noted that after spatial learning, the amount of activity in the hippocampus in slow-wave sleep was proportional to overnight improvement in performance (Peigneux 2004[6]). It has been observed in many experiments that slow-wave sleep deprivation affects declarative memories more than procedural memories (Plihal and Born 1997[7]). Riding a bicycle is an example of a skill that requires procedural memory, while textbook knowledge is declarative in nature. In other words, cutting down on sleep before an exam effectively makes it harder to retain knowledge learned for the exam. This effect is particularly pronounced in the long term. This means that it is less pronounced on the exam day. This is why so many students keep making the same mistake over and over again. They get some more study time on the last night, at the cost of long-term retention of the learned knowledge. Obviously, extra study time has its benefits for the exam itself. Otherwise it is harmful for both health and wisdom.

Sharp wave bursts and long-term memory

In deep sleep, SPW bursts (sharp wave bursts) can be recorded in the hippocampus. Some researchers believe that this may be the critical moment of memory consolidation in which the hippocampus works as the neural trainer for the neocortex in which long-term memories are stored in cross-cortical connections. During SPW bursts, the experience of the day will optimally be transferred to neocortical networks via neural training. This will be followed by the initiation of gene expression and protein synthesis. Both these processes are needed for modifying long-term synaptic weights. Protein synthesis makes up the beginning of memories that will last for months and years (if sustained by a repetition/review, e.g. with SuperMemo). For more details see: Molecular correlates of the two-component model of long-term memory (Wozniak et al. 1998[8]). Those long-term memories cannot be formed without entering appropriate stages of the sleep cycle! You cannot build long-term memories without sleep. In addition, learning will be less efficient if it is cut short in the morning with an alarm clock.

Detecting memory optimization with interference tests

All research into declarative memories may be confounded by the neural optimization occurring in sleep. Sleep will often have a form of refactoring in which the same memories are stored differently. This way, it may not be possible to see the effect of memory change, but its internal representation will change. Such refactoring may not be detectable with behavioral tests or may be very difficult to test for. The same French word stored in working memory feels the same way as when stored in long-term storage. It does not seem to mutate overnight, and if it does, the changes are very hard to notice. Lack of sleep, however, will affect imperfectly stored memories more than those whose storage was optimized. Researchers found it hard to confirm the importance of sleep for declarative memories until they employed interference techniques to show how sleep prevents memories from being overwritten with new information (Ellenbogen et al. 2006[9]). Using SuperMemo, it is possible to collect sleep data that would make similar long-term determinations possible. Due to the scarcity of data with unambiguous sleep restrictions, at the moment, we only know that memory consolidation gets progressively worse as the waking day goes on (see: Sleep and learning).

REM and memory

There are many theories on the functions of REM sleep. It has long been known that most dreams occur in REM sleep, yet some scientists see dreams and REM sleep as separate though temporally overlapping phenomena. It has been found in a number of experiments that REM sleep is important for learning, yet some scientists question those findings pointing to experimental errors or to the fact that antidepressants do not damage memory even though they are potent REM suppressants. Some scientists believe REM is needed to reinforce little used synaptic connections, others that it weakens or deletes little used memories, others that REM helps the brain recover from slow wave sleep, or simply prepare the networks for the state of waking (Klemm 2011[10]). Still others believe that REM evolved just to fine-tune bifocal vision, to prevent corneal anoxia (eye movement stirs aqueous humor), or even restore the hydraulic properties of intervertebral disks (Fryer 2009[11]). Even a few advocates of the old psychoanalytical interpretation of dreams originated by Sigmund Freud can be sparsely found among scientific community. Some researchers believe that memory consolidation is possible during REM, others contest it, and yet others insist that REM has nothing to do with memory. On one hand, the percentage of REM sleep decreases with age which might indicate a correlation with the demand for learning. On the other, the percentage of REM during the night increases. Some researchers believe that if REM was to be involved in memory, it should rather begin quickly as we fall asleep. Others point to the fact that REM is phylogenetically younger and it is NREM that should play the most essential functions related to memory and learning. Historically, the importance of REM sleep for memory and learning was documented before we became truly aware of the role of slow-wave sleep. Consequently, articles and books on sleep are peppered with an overemphasis on the role of REM sleep in learning as compared with SWS. Over time, REM deprivation studies received lots of criticism. Today, we know that the natural harmonious interplay of uninterrupted NREM and REM sleep is essential for memory, learning and creativity (Salzarulo et al. 2000[12]). For more on the theories of sleep, incl. the function of REM, see What is the role of sleep?.

REM and learning

It was 1953 when Eugene Aserinsky and Nathaniel Kleitman published their famous article that demonstrated that sleep is composed of periods in which rapid eye movements occur and which might be associated with dreaming (Aserinsky and Kleitman 1953[13]). Little did they know how monumentally important that finding was. 60 years later, we know that REM and NREM sleep are two totally different brain states that are as different from each other as they are from the obviously different state of wakefulness. REM sleep shows a very different pattern of activity in various brain nuclei. It is characterized by a different direction of information flow. It is dominated by the release of a different set of neurotransmitters.

Carlyle Smith in 1991 showed how the administration of protein synthesis inhibitors during REM sleep windows in rats would prevent behavioral improvements that normally occur after sleep. This was a strong indication that REM sleep is important for memory (not all scientists agree). Moreover, an increase in procedural learning was accompanied by an increase in the density of REM, and the degree of that increase was proportional to the learning capacity of an individual (Smith et al. 2004[14]). The function of REM sleep is different than that of NREM sleep. Some researchers believe that REM may be more important for procedural memory (with declarative memories impaired more with loss of NREM sleep). However, the separation between declarative and procedural learning is more likely to be anatomical (e.g. the cerebellum vs. the hippocampus). It is important to note that fish, as an example, do not show any hallmarks of REM sleep and they definitely do lots of procedural learning after hatching (and probably also even before hatching).

REM deprivation diminishes the effects of learning in proportion to the complexity of the task. Some simple tasks do not seem to be affected (e.g. passive avoidance, simple maze, etc.). However, REM sleep deprivation affects more complex tasks (e.g. operant conditioning, probabilistic learning, complex maze, etc.) (see "Sleeping brain, learning brain. The role of sleep for memory systems" (Peigneux et al. 2001[15]) for review).

In the animal world, the rule of the thumb is that the more immature the newborn at birth, the greater the proportion of REM sleep in the first months. Human newborns are particularly immature in terms of the development of their central nervous system. This is why REM is very important for brain development in babies. REM deprivation in the neonatal period can result in a decreased brain mass, and various developmental and behavioral problems. As all forms of stress affect sleep structure, babies are particularly vulnerable to all forms of sleep disruption and the resulting negative effect on brain development (Peirano and Algarín 2007[16]). Leaving a baby alone in a cot to cry it out is a form of stress that will have long-term detrimental effects on the brain. Instead of a REM-first pattern that characterizes newborns, baby naps in conditions of stress can be REM-poor. Absence of the mother is a cause of stress in babies. For that reason, I advocate sleeping and feeding on demand, as well as co-sleeping for babies.

Some of my SleepChart data seems to tentatively suggest that REM sleep might also affect simple declarative memory (as in learning with SuperMemo). It is not possible to log REM sleep in SleepChart to know for sure, however, delayed sleep as well as sleep interrupted by an alarm clock are likely to be both REM poor. I do not know (yet) how a REM-poor night affects the learning that occurred before sleep. However, a REM poor night definitely reduces learning performance on the day after. In case of an alarm clock disruption, it is hard to say what is actually causing a decline in performance. However, with delayed sleep, the only conceivable alternative explanation is a lesser natural sleep total. I tend to believe that it is rather the scarcity of REM that causes the worse performance. This is because there are subsets of natural short nights that actually lead to excellent learning results.

As NREM and REM are two totally different brain states, what separate roles do they play? In the light of recent findings on the role of NREM sleep in learning, what could possibly be the role of REM, which bears no resemblance to NREM except for the outside appearance of being in the state of rest? One big clue comes from the fact that NREM and REM states keep flipping between each other overnight. Bouts of REM increase the demand for NREM and vice versa. The two stages of sleep show all the hallmarks of the complementary processes that abound in biological control systems. They behave in a flip-flop manner like waking and sleep, they counteract like synthetic and catabolic metabolic pathways, and they compensate. They act like the atria and the ventricles in the synchronous action of the heart. This hints at complementary functions, and the plethora of research findings seems to indicate that those functions revolve around learning and memory.

REM as a form of neural training

One of the hypotheses says that REM sleep is a form of training for the brain. While normal waking activities train the hippocampus with new patterns of activity, REM sleep does the same, only by using imaginary randomized hypothetical patterns. It is as if the brain did not get enough in waking, it needed more special training in sleep. The extra training would be beneficial for it would cost little (no need to expend behavioral energy). It would capitalize on the information already stored in the brain. For example, in REM sleep, the brain might generate a hypothetical situation in which we make a simple but costly mistake. The brain would then re-enact the hypothetical situation, and look for possible scenarios with possible beneficial conclusions. Perhaps we will wake up in sweat on the realization of the cost of the damage and take necessary steps. Frequently enough, the solution will be absurd, which should probably be interpreted as that we should not read too much into dreams. Scientists noticed that some networks replay their waking patterns of activity in REM, and still these are not simple re-enactments of episodic memories of the day, which only serve as a sparse inspiration for dreams.

As much as it is easier to program a computer than to make it learn from real life situations, it is easier and faster to load the hippocampus and other structures with new memories in REM sleep. After each load of new associations, the brain needs to redistribute the information in its long-term cortical storage. That is the function of NREM sleep as described earlier. After many hours of waking, we need over an hour of NREM sleep. However, only a few minutes of REM seem to swing the balance back to favor NREM. The cycle keeps repeating.

REM and creativity

As the night progresses, there is more REM and less NREM. If the "REM training" hypothesis was correct, it might mean that it is harder to generate new information as the night progresses. It might also mean that each NREM bout is incomplete and the remnants of unprocessed information keep blocking full swing REM until the very early morning hours. Perhaps the circadian REM propensity provides for a balance between the storage of old and the synthesis of new information with a gradual shift to favor the latter in later stages of sleep? If all the above scenario was to be true, we might wonder why the brain does not tend to wake up with a clean slate by terminating sleep with the last final NREM episode? Perhaps the transition to waking is all that is needed to clean up the remnants of newly "discovered" information lingering in the working memory? Or the new loads produced by the last bout of REM might have some special survival value? These could be the building blocks for that creative morning insight that the history of science is so rich in. It is possible that it is the very last segment of sleep that ends with REM sleep that provides the morning brain with that next big idea. Some evidence supporting this notion was gathered by Walker when researching performance in anagram puzzles in subject woken at different stages of sleep (Walker et al. 2002[17]). If the reasoning about the creative contribution of the last REM episode is correct, we could arrive at a dramatic conclusion that the alarm clock might be the primary killer of big ideas in the modern world! Stress or rat race are guilty of undermining human creativity too, but it is easier to keep stress in check to get good sleep than to produce great ideas in a stressless world without sufficient sleep. Researchers and educators should be very cautious when diminishing the damage inflicted by alarm clocks (see: Jim Horne and Daniel Kripke). Parents should also show more tolerance for kids who cannot wake up for school (see: Sleep and school).

REM as a neural optimizer

If REM sleep was just a training or creative option in sleep, why would we run REM deficits? Circadian REM propensity might provide for a balance between NREM and REM. Homeostatic NREM pressure might result from learning (in waking or in REM). However, the homeostatic REM sleep propensity, and REM sleep rebounds after REM deprivation both seem to indicate that REM is far more than just an option. The intricate impact of REM deprivation on complex tasks and procedural memory may be an expression of the more important function of REM sleep in which neural optimization is based on the reversal of the direction of the flow of information in the brain as compared with NREM sleep. It is possible that NREM merely serves as the long-term memory storage tool without much ability to optimize the network layout. NREM might simply be an anti-interference tool. However, sleep helps organize memories, increase their abstractness, and reduce the cost of storage. Perhaps that optimizing role rests solely with REM sleep. Perhaps a pseudopattern training makes it possible to relocate wide-network expensive memories into a smaller more generalized circuits. A big clue comes from baby sleep. As newborn's ability to explore its world is limited, the exploratory function of REM sleep might play an essential role in development. However, Dr Siegel noticed that the time spent in REM in humans does not correlate well with their learning abilities. I believe that such inconsistencies are well explained by individual differences that not only express themselves in the learning ability, or the average amount of REM per night, but also in the efficiency of REM (i.e. learning-to-REM ratio). This is analogous to, for example, our digestive abilities. Some people can gorge themselves on food and let it all go out. Others eat very little and are still able to extract all the nutrients down to the last milligram (and bloat). If we look further at various species, we will see even a lesser link between the amount of REM and "animal IQ". This could be explained that smart animals will extract far more value from REM sleep. In other words, no amount of optimization can do wonders in a small capacity network. Even more troubling might be the claim that it is hard to detect any REM in cetaceans, esp. infants (Castellini 2002[18]). If REM was as essential for procedural learning as depicted in this chapter, it would seem indispensable in young intelligent predatory swimmers. Obviously, as a relatively new evolutionary entity, REM-based neural optimization might have its variants with different phenomenology that might not be instantly apparent to researchers. Moreover, cetacean sleep featuring the miracle of unihemispheric slow-wave sleep had over 50 million years to develop characteristics that would set it apart from REM sleep in humans. That's a significant proportion of REM sleep's existence.

Synaptic changes in sleep

Some research shows that synapses get strengthened in sleep while other research finds the opposite effect. Overall synaptic strength tends to increase in waking[19], while the learning capacity keeps declining. Wakefulness increases cortical firing frequency in all behavioral states (Tononi et al. 2009[20]). The simultaneous weakening and strengthening of selected synapses in sleep could best be explained by some kind of memory reshuffling taking place overnight. This does not contradict the memory consolidation function of sleep. This also does not stand in contradiction with the fact that our learning ability tends to decline during the waking day.

When investigating the changes in synaptic strength in sleep, we always need to differentiate between:

  1. short-term memory: short-term increase in synaptic conductivity that is a result of a day's learning
  2. long-term memory: the ability to recall older memories (e.g. as measured with SuperMemo)

To put it metaphorically, the brain is like a computer that keeps loading chunks of data to its memory during the day (short-term memory). As the memory fills up, the computer slows down, and all applications crawl into a halt. However, if you test individual memory cells, you will notice that they strongly cling to their new data. In the night, the computer will gradually organize these chunks of data, remove discrepancies and duplicates, write down memories to the hard disk (long-term memory), and run a defragmentation process for easy and fast access. We need to look at neurophysiological correlates of that metaphor, and for the most likely explanation for the weakening of the recall during the waking period as both the increase in synaptic conductivity in wakefulness, and the decline of learning capacity during the day are well documented. The most coherent, attractive and best-supported hypothesis says that the overload of short-term low-interference networks is responsible for a declining capacity of memory during a waking day (see: NREM and memory). This decline cripples the working memory, and in consequence, it affects the entire spectrum of human cognitive capabilities. The main function of sleep would then be to redistribute, reconsolidate, and optimize those short-term memories that slow down further learning.

As for the decline in synaptic strengths during sleep, it also fits well into the present models of sleep and learning. One of the main functions of sleep should be to optimize the memory storage. This entails representing memories in the most efficient way, so that they are most abstract, consume least space, generate minimum interference, and so on. That process should indeed result in reducing the overall cost of memories, and result in weakening of redundant synaptic connections.

Dr Tononi believes that waking activity produces an overall increase in synaptic weights, and sleep may be necessary to counterbalance that increase. The hypothesized downscaling would occur in slow-wave sleep (Tononi and Cirelli 2006[21]). Dr Tononi clusters disparate components of the memory hierarchy from short-term (phosphorylation), to long-term (AMPA trafficking) to remoulding (sprouting), while I would rather stake my bets on daily learning and short-term memories. Overall cortical downscaling could be a beautiful expression of the post-learning clean-up congruent with the ideas of Crick and Mitchison. The clean-up could be combined with selective synaptic strengthening governed by short-term memory structures (e.g. the hippocampus, the amygdala, etc.).

Let us consider the famous Halle Berry neuron, i.e. a hippocampal neuron that might respond selectively to all-things Halle Berry after an exposure to Halle Berry pictures in a training session. All cortical neurons potentiated during the training would best be silenced in the course of the SWS with the exception of sparsely encoded Halle Berry representation refactored from the hippocampal association, incl. the HB neuron, to a cortical shortcut. This process would free hippocampal learning capacity, produce an overall downscaling, and still retain sparsely encoded pieces of newly learned information.

You may know that SuperMemo is based on the claim that memories get weakened and deleted overtime. That weakening does not refer to the loss of short-term memories in sleep, but to a long-term decline in memories over months and years. Dr Tononi proposed a variant of the theory of forgetting by suggesting that synaptic downscaling in sleep is done in proportion to the existing synaptic strengths. This way the weakest synapses would lose their memory trace. Tononi's proposition may find it difficult to pass the shutdown test unless it shows how the downscaling process requires a network-wide computational operations as opposed to a simpler "molecular forgetting clock" as described in Molecular correlates of the two-component model of long-term memory (Wozniak et al. 1998[8]). However, it is important to note that Tononi often speaks of short-term memory traces registered on the day preceding sleep, not of what, using the two-component model of memory terminology, we call memory retrievability, which tends to decline exponentially between reviews of the learned material (Wozniak et al. 1995[22]). Tononi found that the activated portions of the brain show most slow wave activity in the following night. Both in declarative and procedural learning, increases in cortical SWA are locally specific and proportional to the degree of learning and overnight improvements (Tononi et al. 2004[23]). Tononi explains those findings with an idea that downscaling affects mostly those portions of the brain that are subject to most change. However, another possible explanation is that those portions of the brain get reactivated in sleep as a result of short-term storage changes in the hippocampus to reflect the experience of the day. The hippocampus would represent a short-term memory network used in the training of cortical circuits. Instead of getting weakened though, selected synapses might actually get strengthened while reduced propagation of the stimuli in the cortex (as documented by Massimini (Tononi et al. 2005[24]) could be explained by the need to lay out memories without the following creative and associative propagation of stimuli that might activate more synapses. Overall downscaling would affect all newly potentiated synapses that would not be subject the hippocampal reinforcement.

For an excellent take on the mechanics of sleep see Dr Tononi's lecture at 2011 Allen Institute for Brain Science Symposium.

Ribeiro and Nicolelis believe that experience-dependent plasticity-related gene expression in REM is compatible with Tononi's synaptic downscaling. However, downscaling should affect only the circuits that have not been activated by the waking experience. In other words, upscaling would affect activated circuits, while downscaling would affect inactive circuits. This would increase the signal-to-noise ratio (SNR) in memory consolidation in sleep (Ribeiro and Nicolelis 2004[25]). Dr Walker believes that both the upscaling and downscaling processes might take place in sleep in a complementary manner: "homeostatic synaptic downscaling could result in the removal of superfluous neural connections, resulting in improved SNR. However, neural reactivation and strengthening of experience-dependent circuits, done without removing redundant synaptic connects, may equally improve SNR. Therefore, both mechanisms, while different, could produce a similar outcome: enhanced fidelity of the memory representation" (Walker 2009[26]).

This author believes that this process might be even more general in nature and involve long-term memories that would serve as a structural blueprint for newly optimized fabric of memories collected in short-term memory. This means that upscaling would also involve new neurons in the neocortex that have not been activated by experience. This belief comes from the simple need to recruit new cortical synapses for long-term memories that would become consolidated in an experience-dependent manner over months and years using a simple molecular mechanism that would not be sleep dependent. In other words, recruitment of new cortical synapses would largely be sleep dependent, downscaling would be associated with the post-learning clean-up, while the build-up of memory stability would be a process dependent on reactivation in waking and/or in sleep over the lifetime of a memory trace (Wozniak et al. 1995[22]).

Optimization of memory

After a day of hard work over a problem, if frustration sets in, and the problem seems unsolvable, or exceedingly complex, our working memory may feel like clogged up with pieces of information that do not fall into a coherent structure in which the solution might be found. However, after a night of refreshing sleep, we may suddenly hit upon an idea! This is a not necessarily a result of fresh mind and more morning thinking. Very often, the idea is already there upon awakening. As if the brain worked hard over night without our conscious participation. The process responsible for this magic insight is neural optimization.

Even though the list of biological functions associated with sleep is very long, sleep has evolved for one primary purpose: optimization of memories stored in the neural networks of the brain. This function is so essential that no complex nervous system can survive without it. This is why all complex animals sleep (which is not always easy to tell (Siegel 2008[27]). Even ants take naps.

The size of the cortex is fixed. This means that there are anatomical and functional limitations on how much information can be stored there. Don't believe mnemonic gurus who tell you "we can remember everything", all we need is a "way to access hidden memories". To maintain the ability to form new memories, the cortex must continually rework its representations in order to ensure that only the most salient memories are stored for long-term use. The belief that sleep is helpful in that process is as old as our understanding of the fact that the brain is involved in thinking and in memory. There are researchers though who still find it difficult to reconcile learning with unconscious states.

Hippocampal lesions

In 1953, Henry Gustav Molaison (aka H.M.) had portions of his medial temporal lobe removed bilaterally. After the surgery, H.M. lost his ability to form new long-term memories while retaining his pre-1953 memories, procedural learning and cortically-based working memory capacity. This led researchers to discovering the pivotal role of the hippocampus in the learning circuits of the human brain. The hippocampus receives rich connections from nearly all areas of the cortex, and it feeds back to targets in the same cortical areas. This means that its dense network of connections can maintain a snapshot of the current activation pattern in the cortex. It can also project the same pattern back to the cortex. Once memories are formed, they depend on the hippocampus for a period of days. This suggested that the memories might need to be relocated back to various areas of the neocortex at later time. As sleep deprivation affects this process, it has been suggested early that sleep might be playing a role in the process. Other richly connected areas of the brain play a similar role for various forms of specialized memories. The purpose of "memory transfer" became clear gradually over many years with contributions coming from various researchers coming from various fields.

Hippocampal lesions provide a very strong clue to the idea that the conversion of short-term to long-term memories is computational, and not just molecular/synaptic in nature. It also hints that the nature and the layout of memories will differ upon the conversion

Temporally graded retrograde amnesia

When there is an injury to the hippocampus, in addition to the inability to form new memories (anterograde amnesia), there is also a degree of loss to previously formed memories (retrograde amnesia). As early as in 1881, Théodule Ribot suggested that recent memories are more likely to be lost in retrograde amnesia (Ribot's law). Loss of memories proportional to their recency was termed the temporally graded retrograde amnesia. There have been many hypotheses for explaining this phenomenon, some of which, wrongly implicate the hippocampus in the process of prolonged storage of some of those memories, which are supposedly being gradually consolidated into the long-term storage. Other interpretations speak of a gradual physical transfer of memories in the network (e.g. one integrating neural cell layer after another), in the process that may last years. To anyone familiar with the two-component model of long-term memory (Wozniak 1995[22]), it is pretty obvious that the involvement of the lesions to the hippocampus in the temporally graded retrograde amnesia does not need to imply the involvement of this structure in storing memories in the long term. It could equally well be explained by the hippocampal involvement in the reconsolidation process that serves the build-up of memory stability. The hippocampus does not need to slowly consolidate memories stored in its connections. It is enough that it is involved in re-activation of those memories through review in the same way as it is done in the original establishment of cortical connections. This way, in retrograde amnesia, memories with lower stability will be lost in the first order. A simple way to verify this fact would be to track the course of the amnesia over years (e.g. with SuperMemo). If the hippocampus is necessary for the buildup of memory stability, the degree of amnesia should progressively get worse as implied by the natural process of forgetting. Possibly, the forgetting would not be as fast as it is the case in a healthy brain due to the lack of interference from new memories.

If the above interpretation of temporally graded retrograde amnesia is correct, it will provide a further clue as to the role of the hippocampus in establishing new memories and building memory stability over time. This would strengthen the concept of the hippocampus serving as the primary gateway for declarative memories stored in the cortex, and re-emphasize the computational aspect of this process.

Memory processing in sleep

In 1963, Sokolev suggested that the primary power of the brain rested in its ability to build a model of the surrounding world. However, it has always been hard to figure out how a chaos of data that arrives at sensory inputs gets reshaped into the magic world of abstract shapes and models of the human mind. In 1970, Marr suggested that memory consolidation in sleep might be an inductive process of sorting experiences into categories. This process would be based on statistical sampling of the environment. Marr proposed that the hippocampus stores experiences acquired during the day, and replays them back to the neocortex overnight. That replay would be the time when the category formation would occur. In 1979, Wickelgren suggested that the hippocampus is needed to assign cortical representations to novel conjunctions of inputs. The neocortex can then treat these conjunctions separately (e.g. like new items in SuperMemo). In 1992, French noticed that humans rarely forget "catastrophically". To explain this, he suggested that to prevent catastrophic forgetting, it was necessary for the neural networks of the brain to separate their internal representations during learning.

Catastrophic forgetting

In 1995, inspired by Squire's work on amnesia and his hypotheses on the involvement of the medial temporal cortex in memory consolidation (Squire et al. 1984[28]), McClelland, McNaughton, and O'Reilly capitalized on earlier theories and proposed that the brain copes with the problem of catastrophic forgetting by evolving two separate memory systems to separate novel representations from established memories: a dual network theory system. They gave their theory a computational framework. In the proposed complementary encoding mode, the hippocampus and the neocortex play the role of the dual network theory system. Using neural network simulations, the researchers showed their own interpretation of how temporally graded retrograde amnesia could proceed (still implicating the hippocampus as a possible long-term storage)(McClelland et al. 1995[29]).

In 2001, French et al. suggested how pseudopatterns could serve as a way of transferring information between neural nets (French et al. 2001[30]) (shortened):

"A neural network has learned a number of input-output patterns corresponding to some underlying function f. How can we get an approximation of the original function f? One solution: send random input into the network and observe the output for. We thus create a series of pseudopatterns, ψi, where each pattern ψi is defined by a random input and the output of the network after that input has been sent through it. Pseudopatterns were first introduced by Robins to overcome catastrophic interference. Robins suggested that when a network had to learn a new pattern, a number of pseudopatterns be generated. Then, instead of learning just the new pattern, P, the network would be trained on the new pattern plus the set of pseudopatterns that reflected what it had previously learned. In this way, the new pattern would be interleaved with patterns that, even though they were not the originally learned patterns, nonetheless reflected the original function learned. Robins showed that his technique did, indeed, reduce catastrophic interference"

Two-stage memory processing in sleep

It might seem an attractive and plausible proposition to envisage the hippocampus as a device for taking snapshot of the cortical activation status. Those snapshots could be taken in NREM and played back to the neocortex in the REM training mode. Some neuronal firing experiments seemed to even indicate such a possible sequence. However, Dr Georgi Buzsáki concluded that the process is actually reverse in nature. Again the sequence of neuronal firing measured experimentally was important to come to his conclusion.

In 1989, Pavlides and Winson demonstrated that hippocampal neurons active during a learning episode are more active during subsequent sleep (Pavlides and Winson 1989[31]. In the same year, Buzsáki suggested that hippocampal sharp wave bursts (abundant in deep sleep) may represent a mechanism for the consolidation of representations in intra and extrahippocampal circuits and for memory transfer from the hippocampus to the neocortex (Buzsáki 1989[32]). Buzsáki ideas led to a belief that SPW result in reactivation of cell assemblies which were earlier potentiated by exploratory activity (experience or REM sleep).

A student of Alexander Luria, and Sokolev's disciple, Olga Vinogradova, in her last publication, which was accepted a week before her death at the age of 72, summarized her 35 years of work over the hippocampal function in "Hippocampus as Comparator: Role of the Two Input and Two Output Systems of the hippocampus in Selection and Registration of Information", which paints the hippocampus as the primary novelty detector with far reaching implications for its role in memory optimization (Vinogradova 2001[33]).

Buzsáki, however, proposed a model in which the entorhinal cortex is the comparator (Buzsáki, unpublished, 2000). He hypothesized that the entorhinal cortex would help detect novelty in the system as follows:

"The EC functions as a "comparator" and evaluates the difference between neocortical representations and the feedback information conveyed by the hippocampus (the reconstructed input). The resulting difference or "error" is regarded as "novelty" and it is this novel information which initiates plastic changes in the hippocampal networks (error compensation). Alteration of synaptic connectivity in the hippocampus, in turn, gives rise to a new hippocampal output. In this process, the hippocampus generates separated (independent) outputs, i.e., it minimizes mutual information between its outputs. The output of the hippocampus, in turn, trains the long-term memory traces to minimize mutual information transfer amongst them. After long-term memories are trained properly, the hippocampal output will no longer affect synaptic weights in the EC. This feature is expressed by the auto-associative nature of the model: the hippocampus provides the reconstructed input (the auto-association) and the difference between the input and the reconstructed input drives the training of long-term memories. Because only the reconstruction error enters the hippocampus, a relatively limited computational network is sufficient for training long-term memories in the EC/neocortex"

According to Buzsáki, the memory optimization may be executed with the mediation of the hippocampus that would work in

  1. explorative mode during the REM sleep (in which neocortical information is used to train hippocampal circuitry), and
  2. in consummatory/sleep mode during the NREM sleep (in which the hippocampus is used to train neocortical circuits).

The network optimization hypothesis would explain why it is hard to detect rote learning deterioration in sleep deprivation. NREM/REM interplay defined as above should have less bearing on the output generated by the same inputs in reference to low-level associations (such as stimulus pairing). To detect the damage induced by sleep deprivation more complex tests should be used. Indeed some research by Dr Carlyle Smith has already been able to show the difference in the impact of REM sleep deprivation (REMD) on paired associate learning (which suffers little damage in REMD) and complex logic tasks which are most affected by REM sleep deprivation (Smith 1993[34]).

Buzsáki model is based on the changes in the direction of the flow of information in the brain during NREM and REM stages of sleep. The control of that flow is probably partly systemic/neural and partly neurohormonal. Anticholinergic drugs such as scopolamine can produce a delirious waking state with hallucinations, anxiety, and confabulations (Perry and Perry 1995[35]). The AIM model describes various brain states that pivot around the modulation axis and its cholinergic dimension, which is central to memory optimization in sleep (Hobson et al. 2000[36]). Cholinergic modulation in REM and wakefulness could help load the hippocampal circuits with new memories while aminergic dominance would reverse the flow to enhance the encoding of memories in the neocortex while inhibiting hippocampal encoding (Hasselmo 1999[37]).

Optimizing memories

We can now conclude that, in sleep, memories move from a temporary low-capacity fast-encoding high-associativity low-interference storage to areas where, on the basis of their novelty and applicability, they can safely be used for months and years without much interference from newly arriving memory data at little cost. Abstract patterns are extracted, while details obscuring the big picture are discarded. Rewiring of the network might bring some of the following advantages:

  • converting poorly associative memories into highly associative memories (the origin of the ancient phrase: let me consult my pillow)
  • eliminating knowledge interference to help avoid confusion between similar concepts
  • extracting common properties of objects and building models (pictorially: instead of holding 100 pictures of someone's face and searching on each encounter, recognize all common model characteristics and execute recognition in milliseconds)
  • optimizing procedural reflexes (some researchers even proposed that REM is mostly targeted on consolidation of procedural skills which seem to suffer most from REM deprivation, while NREM sleep would serve only the consolidation of declarative skills)
  • transferring memories from overloaded circuits (e.g. the hippocampus) to spacious areas of the neocortex

For more see: Neurostatistical Model of Memory

Garbage collection

Network optimization is not only a process that ensures long-term usability of memories, but also a vital cleanup mechanism that makes sure working memory storage is unaffected by the pile up of data. We can see a correlation between the activity in prefrontal cortex and the degree of sleepiness in memory tasks. In line with the dual network theory, once the short-term memory storage starts filling up, a compensatory mechanism is involved and the cortex may temporarily be used to assist in memory tasks. Giulio Tononi of the University of Wisconsin proposes that the network overload simply makes the brain more and more expensive to maintain (see: Synaptic changes in sleep). This is why the sleep is needed to do synaptic downscaling. Those downcaling ideas threw some confusion into sleep research. Some findings indicate that synaptic connections get strengthened in sleep, while other researchers noticed the opposite effect. If sleep was to be a neural optimizer, we should expect some connection to get weakened while others would get stronger. Marcello Massimini (Tononi et al. 2005[24]) used TMS to see how activation in one area of the cortex got transmitted to other areas of the brain in waking and in sleep. He noticed that in NREM sleep, the initial response was very strong, but would get quickly extinguished and did not propagate far beyond the site of stimulation. In line with Evans or Crick theories, Massimini noticed a weakening of synaptic connections in sleep. This could agree with the overall downscaling process observations of Tononi, however, the picture could be confused by the effects of neuromodulation that changes the modes in which the networks operate in sleep. Videos of the signal propagation in waking and sleeping brain can be seen here.

While Tononi and Massimini hypothesized on the synaptic downscaling in sleep, other scientists theorized and showed experimentally that slow-wave sleep can also enhance synaptic connections (Lee and Wilson 2002[38]; Sejnowski and Destexhe 2000[39]; Steriade and Timofeev 2003[40]). Sleep deprivation leads to a higher cortical activation, and increases the number of areas active when solving complex tasks (Drummond et al. 2000[41]). Tononi interprets those results as an effect of an overall increase in synaptic weights in the course of waking. Increased weights result in higher overall activation. However, the shifting patterns of activation (e.g. from temporal to parietal lobes) might suggest that the cause is a bit different. It seems like the brain recruits new areas of the cortex to compensate for overloaded networks that cannot keep up with the extra encoding and processing. This observation is essential in figuring out to what degree the sleep process is a network-emergent phenomenon as suggested by hypotheses by Krueger and Tononi, and to what degree it requires central control to fulfill its computational role in NREM-REM sequence interplay.

Unihemispheric sleep

An interesting question arises in the case of dolphins and birds that developed unihemispheric sleep. That form of sleep should be suboptimal for storage optimization. It could be compared to disk defragmentation in which only half of the disk space is available for disk housekeeping. Such a process could be highly beneficial, and yet to approach an optimum full-storage optimization it would require an exponentially more time for each increment in quality measurement. Most importantly, we could expect more hemispheric specialization and less inter-hemispheric communication. Obviously, unihemispheric sleep would be precious as a temporary measure in conditions of danger, migration, etc.

If the evolutionary step towards unihemispheric sleep seems complex, one only needs to note that cats with severed corpus callosum were able to sleep with one hemisphere in NREM sleep and the other hemisphere in the state of waking. In other words, blocking the inter-hemispheric communication may be all that is needed to produce hemispheric asynchrony. On a lighter note, efforts of polyphasic sleepers might actually bring a similar adaptation in humans if the adopters formed an isolated population. It would probably take thousands of years and many evolutionary casualties before polyphasic sleepers would finally "adapt" and be able to sleep unihemispherically.

Problem solving in sleep

Inspired by the concept of neural optimization in sleep, years ago, I developed my own formula for problem solving, which works for me each time I need a solution to a complex task marred by excess contradictory information. Sleep provides for a half of the solution! This is my formula:

  1. get good sleep: sleeping in the right phase and without any artificial control is vital! (see: Free running sleep)
  2. think about the problem: how can I solve it? what information can help solve it? This step requires 100% isolation from the outside world. It works great in nature or when just pacing a room up and down.
  3. read about the problem until my brain sizzles. With neural creativity and incremental reading, it can really take an hour or two to load the working memory up to its capacity, and push the brain to exhaustion. Again, this step requires minimum interruption to ensure 100% focus on the problem. Not a single brain cell should be absorbed with the conflicts of the world. All resources must go into problem solving.
  4. exercise: to stimulate circulation, to provide time for lazy unstructured thinking, and to fill the time before the next opportune sleep episode, I exercise. I know many people who solve problems successfully without ever exercising. So this might be just my personal optional favorite. I think it is important to keep the brain pure in its focus on the problem. I "pollute" the mind with irrelevant information only in cases of a major lockup, or mental block, where the solution to the problem is particularly elusive.
  5. go back to Step 1, only to discover that the previous round pushed my thinking by a country mile, and that sleep portion was essential for being able to see the big picture. Napping is great as it counts as much in the cycle as night-time sleep does. Pity we have been designed to nap only once per day. Perhaps multiple naps would provide for more creative steps during a day (if it was feasible).

For difficult problems, time is an ally! The more think-learn-sleep cycles you can run, the closer you can get to the target. Keeping the mind pure is vital, but taking occasional breaks for unrelated information processing can unclog prejudiced pathways in the brain.

For my problem solving formula see: How to solve any problem?


During the exploratory activity in waking, the associative networks of the brain (incl. the hippocampus) integrate information from various portions of the cortex with new information coming from various highly-processed sensory inputs. Cortical processing is responsible for the working memory and thinking, while the associative networks hold patterns of recent activity. During waking, cortical networks get overloaded with potentiated connections, while the hippocampus gets overloaded with new associative patterns. In NREM sleep, cortical processing is inhibited, the cortex is globally depotentiated and hippocampal patterns are used to integrate newly acquired information with previously stored cortical long-term memories. REM sleep is used to train the hippocampal network with new patterns garnered from the cortex in a process that can be likened to a "simulated waking". Those new patterns are then transferred back to the neocortex in the successive NREM episode. NREM-REM interplay is used to remould knowledge away from detail-rich patterns towards generalized patterns. This interplay, which repeats several times in the course of the night, is what makes us smart. This interplay helps us use little information for maximum effect. Frequently used patterns get reinforced in the cortex by gradually building their synaptic stability, while the synaptic retrievability decays in a negatively exponential manner to maximize the utility of memories and minimize the cost of storage (Wozniak 1995[22]).

Sleep is the peak achievement of neural computation. With good sleep, human brain is a highly efficient problem solving machine


  1. Marshall L. and Born J., "The contribution of sleep to hippocampus-dependent memory consolidation," Trends in Cognitive Sciences / Volume 11 / Issue 10 (October 1, 2007): 442-450, doi: 10.1016/j.tics.2007.09.001
  2. Wilson M.A. and McNaughton B.L., "Reactivation of hippocampal ensemble memories during sleep," Science / Volume 265 / Issue 5172 (July 29, 1994): 676-679, doi: 10.1126/science.8036517
  3. Buzsáki G., "The Hippocampo-Neocortical Dialogue," Cerebral cortex / Volume 6 / Issue 2 (March-April 1996): 81-92
  4. Goetz R.R., Goetz D.M., Hanlon C., Davies M., Weitzman E.D., and Puig-Antich J., "Spindle characteristics in prepubertal major depressives during an episode and after sustained recovery: a controlled study," Sleep / Volume 6 / Issue 4 (December 1983): 369-375
  5. Takashima A., Petersson K.M., Rutters F., Tendolkar I., Jensen O., Zwarts M.J., McNaughton B.L., and Fernández G., "Declarative memory consolidation in humans: A prospective functional magnetic resonance imaging study," PNAS / Volume 103 / Issue 3 (January 16, 2006): 756-761, doi: 10.1073/pnas.0507774103
  6. Peigneux P., Laureys S., Fuchs S., Collette F., Perrin F., Reggers J., Phillips C., Degueldre C., Del Fiore G., Aerts J., Luxen A. and Maquet P., "Are Spatial Memories Strengthened in the Human Hippocampus during Slow Wave Sleep?," Neuron / Volume 44 / Issue 3 (October 28, 2004): 535-545, doi: 10.1016/j.neuron.2004.10.007
  7. Plihal W. and Born J., "Effects of Early and Late Nocturnal Sleep on Declarative and Procedural Memory," Journal of Cognitive Neuroscience / Volume 9 / Issue 4 (July 1997): 534-547, doi: 10.1162/jocn.1997.9.4.534
  8. 8.0 8.1 Wozniak P., PhD and Gorzelanczyk E.J., PhD, "Hypothetical molecular correlates of the two-component model of long-term memory" (May 1998)
  9. Ellenbogen J.M., Hulbert J.C., Stickgold R., Dinges D.F., and Thompson-Schill S.L., "Interfering with Theories of Sleep and Memory: Sleep, Declarative Memory, and Associative Interference," Current Biology / Volume 16 / Issue 13 (July 11, 2006): 1290-1294, doi: 10.1016/j.cub.2006.05.024
  10. Klemm W.R., "Why Does REM Sleep Occur? A Wake-up Hypothesis," Frontiers in Systems Neuroscience / Volume 5 (September 2011), doi: 10.3389/fnsys.2011.00073
  11. Fryer J.C.J., "Is a purpose of REM sleep atonia to help regenerate intervertebral disc volumetric loss?," Journal of Circadian Rhythms / Volume 7 (January 5, 2009), doi: 10.1186/1740-3391-7-1
  12. Ficca G., Lombardo P., Rossi L., and Salzarulo P., "Morning recall of verbal material depends on prior sleep organization," Behavioural Brain Research / Volume 112 / Issues 1-2 (July 2000): 159-163, doi: 10.1016/S0166-4328(00)00177-7
  13. Aserinsky E. and Kleitman N., "Regularly Occurring Periods of Eye Motility, and Concomitant Phenomena, During Sleep," The Journal of Neuropsychiatry and Clinical Neurosciences / Volume 15 / Issue 4 (November 1, 2003): 454-455, doi: 10.1176/appi.neuropsych.15.4.454
  14. Smith C.T., Nixon M.R., and Nader R.S., "Posttraining increases in REM sleep intensity implicate REM sleep in memory processing and provide a biological marker of learning potential," Learning Memory / Volume 11 (2004): 714-719, doi: 10.1101/lm.74904
  15. Peigneux P., Laureys S., Delbeuck X., and Maquet, P., "Sleeping brain, learning brain. The role of sleep for memory systems," NeuroReport / Volume 12 / Issue 18 (December 21, 2001): 111-124
  16. Peirano P.D. and Algarín C.R., "Sleep in brain development rem in development," Biological Research / Volume 40 / Issue 4 (2007): 471-478, doi: 10.4067/S0716-97602007000500008
  17. Walker M.P., Liston C., Hobson J.A., and Stickgold R., "Cognitive flexibility across the sleep–wake cycle: REM-sleep enhancement of anagram problem solving," Cognitive Brain Research / Volume 14 / Issue 3 (November 2002): 317-324, doi: 10.1016/S0926-6410(02)00134-9
  18. Castellini M.A., "Sleep in Aquatic Mammals." In "Sleep-Related Breathing Disorders - Experimental Models and Therapeutic Potential" edited by David W . Carley and Miodrag Radulovacki (New York: Marcel Dekker, Inc., 2002), 279-294
  19. Fountain H., "Study Gives Key Role to Sleep in Helping Brain Learn Anew," The New York Times
  20. Vyazovskiy V.V., Olcese U., Lazimy Y.M., Faraguna U., Esser S.K., Williams J.C., Cirelli C., and Tononi G., "Cortical Firing and Sleep Homeostasis," Neuron / Volume 63 / Issue 6 (September 24, 2009): 865-878, doi: 10.1016/j.neuron.2009.08.024
  21. Tononi G. and Cirelli C., "Sleep function and synaptic homeostasis," Sleep Medicine Reviews / Volume 10 / Issue 1 (February 2006): 49-62, doi: 10.1016/j.smrv.2005.05.002
  22. 22.0 22.1 22.2 22.3 Wozniak P.A., Gorzelanczyk E.J., and Murakowski J.A., "Two components of long term memory," Acta Neurobiologiae Experimentalis / Volume 55 (1995): 301-305
  23. Huber R., Ghilardi M.F., Massimini M., and Tononi G., "Local sleep and learning," Nature / Volume 430 (July 1, 2004): 78-81, doi: 10.1038/nature02663
  24. 24.0 24.1 Massimini M., Ferrarelli F., Huber R., Esser S.K., Singh H., and Tononi G., "Breakdown of Cortical Effective Connectivity During Sleep," Science / Volume 309 / Issue 5744 (September 30, 2005: 2228-2232, doi: 10.1126/science.1117256
  25. Ribeiro S. and Nicolelis M.A.L., "Reverberation, storage, and postsynaptic propagation of memories during sleep," Learning & Memory / Volume 11 / Issue 6 (November 2004): 686-696, doi: 10.1101/lm.75604
  26. Walker M.P., "The Role of Sleep in Cognition and Emotion," Annals of the New York Academy of Sciences / Volume 1156 (March 2009): 168-197
  27. Siegel J.M., "Do all animals sleep?," Trends in Neurosciences / Volume 31 / Issue 4 (April 1, 2008): 208-213, doi: 10.1016/j.tins.2008.02.001
  28. Squire L.R., Cohen N.J., and Nadel L., "The medial temporal region and memory consolidation: A new hypothesis." In "Memory Consolidation: Psychobiology of Cognition," edited by Weingartner H. and Parker E.S. (New Jersey: Lawrence Erlbaum Associates, 1984), 185-210
  29. McClelland J.L., McNaughton B.L., and O'Reilly R.C., "Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory," Psychological Review / Volume 102 / Issue 3 (July 1995): 419-457
  30. French R.M., Ans B., and Rousset S., "Pseudopatterns and Dual-Network Memory Models: Advantages and Shortcomings." In "Connectionist Models of Learning, Development and Evolution: The 6th Neural Computation and Psychology Workshop, Liège, Belgium, 16-18 September 2000," edited by French R.M. and Sougné J.P. (London: Springer-Verlag London Limited, 2001), 13-22
  31. Pavlides C. and Winson J., "Influences of Hippocampal Place Cell Firing in the Awake State on the Activity of These Cells During Subsequent Sleep Episodes," The Journal of Neuroscience / Volume 9 / Issue 8 (August 1989): 2907-2918
  32. Buzsáki G., "Two-stage model of memory trace formation: A role for “noisy” brain states," Neuroscience / Volume 31 / Issue 3 (1989): 551-570
  33. Vinogradova O.S., "Hippocampus as comparator: Role of the two input and two output systems of the hippocampus in selection and registration of information," Hippocampus / Volume 11 / Issue 5 (October 2001): 578-598, doi: 10.1002/hipo.1073
  34. Smith C., "REM Sleep and Learning: Some Recent Findings." In "The Functions of Dreaming," edited by Moffitt A., Kramer M., and Hoffmann R. (New York: State University of New York Press, 1993), 341-362
  35. Perry E.K. and Perry R.H., "Acetylcholine and hallucinations: disease-related compared to drug-induced alterations in human consciousness," Brain and Cognition / Volume 28 / Issue 3 (August 1995): 240-258
  36. Hobson J.A., Pace-Schott E.F., and Stickgold R., "Dreaming and the brain: Toward a cognitive neuroscience of conscious states," Behavioral and Brain Sciences / Volume 23 / Issue 6 (2000): 793-842
  37. Hasselmo M.E., "Neuromodulation: acetylcholine and memory consolidation," Trends in Cognitive Sciences / Volume 3 / Issue 9 (September 1, 1999): 351-359, doi: 10.1016/S1364-6613(99)01365-0
  38. Lee A.K. and Wilson M.A., "Memory of Sequential Experience in the Hippocampus during Slow Wave Sleep," Neuron / Volume 36 / Issue 6 (December 19, 2002): 1183-1194, doi: 10.1016/S0896-6273(02)01096-6
  39. Sejnowski T.J. and Destexhe A., "Why do we sleep?," Brain Research / Volume 886 / Issues 1-2 (December 15, 2000): 208-223, doi: 10.1016/S0006-8993(00)03007-9
  40. Steriade M. and Timofeev I., "Neuronal Plasticity in Thalamocortical Networks during Sleep and Waking Oscillations," Neuron / Volume 37 / Issue 4 (February 20, 2003): 563-576, doi: 10.1016/S0896-6273(03)00065-5
  41. Drummond S.P., Brown G.G., Gillin J.C., Stricker J.L., Wong E.C., and Buxton R.B., "Altered brain response to verbal learning following sleep deprivation," Nature / Volume 403 / Issue 6770 (2000): 655-657