Sleep and learning
- 1 How sleep affects learning?
- 2 Studying sleep and learning with SuperMemo
- 2.1 Long sleep results in poor learning?
- 2.2 Learning reduces the demand for sleep?
- 2.3 Approximating the sleep phase
- 2.4 Timing of repetitions
- 2.5 The impact of SleepChart
- 2.6 Recall vs. Consolidation
- 2.7 Alarm clock vs. learning
- 2.8 Learning in free running sleep
- 2.9 Alertness multiplier
- 2.10 Learning overload
- 2.11 Alertness vs. learning
- 3 How learning affects sleep?
- 4 Sleep and school
- 5 Learning in alpha state
- 6 Learning during sleep
- 7 References
How sleep affects learning?
Why is sleep important for learning?
If I was to bet on the top two factors that hinder learning in industrialized nations, these would be:
- Stress which takes away your focus, stifles creativity, saps motivation, and which can contribute to poor sleep
- Sleep which is needed for optimizing memories. Without sleep, you cannot even experience the sense of a "good day"
Health is important too, but, statistically, it is stress and bad sleep that affect nearly everyone, and take the largest toll. Reduce stress and improve sleep, and you might see a society changed beyond recognition!
For healthy people, all other factors in learning seem to be somewhat secondary. Self-discipline improves greatly if you are rested and happy. The fun of learning follows. The way you approach learning, tools and techniques, the way you represent knowledge in your mind, and other factors can all be improved gradually and consistently. If you are on a steady path ahead, success is nearly guaranteed. Metaphorically speaking, your brain comes with a solid warranty of progress that you can easily void with stress and/or poor sleep.
Given the importance of sleep, unless you are a "natural" and rarely get a bad night sleep, you should understand the basics of sleep physiology and the impact of your sleep habits on learning. Moreover, even if you sleep well today, you are always in danger of ruining your sleep patterns through the use of computers, Internet, mobile phones, SuperMemo, etc. In short, the human brain has not yet got enough time to evolve and adapt to the stimuli of the modern lifestyle. That's why we witness an epidemic of sleep disorders in industrialized nations.
In the following sections, I will try to show that the impact of sleep on learning goes far beyond the simplistic concept of "rested mind".
Sleep and learning research
Everyone knows that without a good night in bed, the next day can be ruined. When sleepy, you can easily shovel the garden in fresh air, but if you try some creative work in front of your computer in a warm room, your brain will tend to switch off and stifle any creative progress.
It is quite evident that cognitive functions and learning are the primary victims of sleep deprivation. Scientists have for long suspected that the main function of sleep is related to learning and memory. Even in the 17th century, John Locke campaigned for good sleep for kids for those reasons. However, only recent decades and years brought an exponential increase in evidence demonstrating the role of sleep in memory. There are still prominent sleep researchers that dispute the link. Some insist that only a conscious brain can be involved in memory. Others claim that sleep is like eating, if you can get more, you will always want to get more. Outside the scientific community, sleep is held in an amazing disregard. Many people do not want to waste time on sleep to economize more time for work and "creativity". Others try to get "best" sleep in minimum time (see: Polyphasic sleep).
The worst part of that disregard is that little kids worldwide are woken up early in the morning to go to school to "learn". Not only does their learning suffer, or even becomes futile; not only do those kids get stressed and cranky; their health can be affected. Their immune systems undermined. Their long-term development stunted. Some sleep researchers try to battle the establishment for more rational school schedules (hats off to Dr Mary Carskadon and Dr Amy Wolfson; see interview). At the same time, the ever-present rat race produces forces in the US, in Europe, and beyond, that insist on even earlier school hours. That comes from both parents and from the authorities. They all bring up a spurious and biologically untenable excuse: the kids can just go to sleep earlier.
In this gloom and doom scenario, there is still a ray of hope though. Science is slow to percolate into social awareness; however, in the end, it wins most of the time (except where it needs to combat stronger forces; e.g. intelligent design theories still keep doing well with the backing of religious doctrinaires). My optimistic prediction is that, sooner or later, governments, school authorities, and parents will realize that the use of an alarm clock to rip kids from their beds contradicts the goals of education!
Studying sleep and learning with SuperMemo
For three decades now, I have been interested in the negative impact of modern lifestyle on sleep and learning. I have suggested that a large proportion of sleep disorders can be remedied with simple techniques such as chronotherapy, free running sleep, etc. The first step towards a solution to a sleep problem is the understanding of one's own sleep patterns. For that reason, I have encouraged people with sleep problems to collect their sleep data with SleepChart freeware that was released in 2003 (download). When SleepChart was created it was not clear what benefits it would bring. I have suggested that SleepChart might in the future be used to investigate the links between sleep and learning, and that SleepChart could become a tool for the optimization of learning, esp. when used in conjunction with SuperMemo. One of beautiful things about SuperMemo is that it keeps a detailed record of memory performance while you learn. If that record could be combined with measurements of sleep quality before and after learning, an ocean of research opportunities would emerge. It was the SleepChart application that provided the missing link. With SuperMemo and SleepChart, we can collect data that can provide answers to a virtually infinite set of questions about sleep and learning. However, my suggestion that SleepChart and SuperMemo be integrated, raised a lot of opposition, primarily from users of SuperMemo who have always complained that the program fell into an endless spiral of mounting complexity and that few users will ever need or make use of the new functionality.
Long sleep results in poor learning?
As of 1996, SuperMemo makes it possible to keep a detailed record of all repetitions. You can check which piece of knowledge was reviewed, when, and with what outcome. As of January 2000, I kept a detailed record of my own sleep timing. I was always curious how sleep affects learning and how learning affects sleep. With learning and sleep data at hand, I could look for correlations between the two. My first, most atavistic and raw intuition was that it should be easy to show that short sleep produces poor learning. "Does more sleep help learning?" I took my own sleep-and-learning data to quickly investigate such a correlation. However, nearly a reverse relationship could be demonstrated. In retrospect, the paradox is very easy to explain: in free running sleep, which I practise religiously, there is a correlation between the quality of sleep and its length: the better the alignment of the sleep episode with the circadian rhythm, the shorter the sleep, and the better its quality. Unless they are sleep deprived, healthy people sleep long only if they sleep in a wrong phase. Optimum sleep is usually very short. In other words, length of sleep is no measure of sleep quality.
Learning reduces the demand for sleep?
An analogous question to ask was "Does learning increase the demand for sleep?" When I tried to investigate this mirror question, I was equally unsuccessful. Again an inverse correlation could be noticed. This time, the reason for that surprise was that insufficient sleep discourages learning. This way, less sleep means less learning, and longer sleep on the following night to repay the sleep debt. In other words, lots of learning would paradoxically be followed by little sleep! For more details see: Impact of learning on sleep.
Those failure made it apparent that little evidence can be garnered on the relationship between sleep and learning without considering the circadian timing, i.e. the time in which learning takes place in reference to the sleep phase (e.g. as determined by the natural waking hour).
Approximating the sleep phase
In the next step, I was hoping to see a correlation between learning and the disparity between sleep time and sleep phase. However, for this correlation to be computable, one needs a good estimation of a circadian rhythm phase. SleepChart uses a rough heuristic algorithm that attempts to do just that. However, this algorithm was too weak to interpret major disturbances in the sleep rhythm caused by delayed sleep, stress, exhausting exercise, etc. That algorithm was replaced in SleepChart 2.0, which uses a recursive phase response curve (rPRC) to estimate the circadian acrophase. rPRC is a variant of a phase response curve that is based solely on the outward expression of the circadian rhythm as documented by sleep logs, and whose only phase shifting stimulus is the delay in bedtime (in reference to the optimum bedtime).
Timing of repetitions
Another stumbling block in further research was a feature used in SuperMemo called Midnight clock shift. It makes it possible to use circadian time for repetition record as opposed to clock time. For example, if the student keeps working after midnight, repetitions are recorded for the previous day, not the new calendar day. That could cause misalignment of sleep and learning data by an entire day. Sadly, earlier versions of SuperMemo kept only the date of the repetition, not its precise time. This was changed only in SuperMemo 13 (2006), in which the clock time of each repetition is recorded. This makes it possible to compute the exact circadian timing of each memory recall act. At last, it was possible to correlate sleep data with learning in precise time frames! I had the alpha release of SuperMemo 2006 available as of July 17, 2006. This means that, as of this writing, we have passed the fifth anniversary of data collection, and the data set is getting bigger and more meaningful with each passing day. The circle of people logging their sleep in SuperMemo is increasing.
The impact of SleepChart
The application of SleepChart in SuperMemo surpassed all expectations in its value and is now a unique tool for investigating sleep and learning. As of the release of this article (winter 2012), this is the only tool in the world that makes similar investigations possible. The employment of SuperMemo in this research is essential as it effectively aims at the same level of knowledge retention at each review. This provides for a steadier comparison platform between different levels of circadian and homeostatic sleep propensity. Developers of other spaced repetition applications have never expressed much interest in investigating sleep. Moreover, the extra accuracy of the newest SuperMemo Algorithm SM-15 provides for extra sensitivity that should yield faster clarification of trends and correlations even for smaller datasets.
You can also join the research effort! In SuperMemo 15, you only need to log in your sleep and send the data with just one button push. More details on the functionality of SleepChart in SuperMemo can be found here.
Recall vs. Consolidation
In studying the impact of sleep on learning, we have to separate two important measures of memory: recall and consolidation.
Recall measures the proportion of pieces of information that can be recalled from memory at any given circadian time. In SuperMemo, recall can be simply measured as the average grade received in learning within a selected subperiod of circadian time. Grades can be converted to percent recall, or can be used as an equivalent measure of recall. The conversion to recall may be of all-or-nothing type (successful recall is treated as 100% recall, while a recall failure is treated as 0% recall). The conversion can also rely on the expected and/or estimated forgetting index in SuperMemo to provide a more precise reflection of recall difficulty. The conversion that uses the forgetting index may be based on the correlation between grades and the expected forgetting index, or can use a heuristic based on the subjective estimated forgetting index assessment (note that the estimated forgetting index, unlike the expected forgetting index, is not part of repetition history in SuperMemo). That latter, seemingly less precise approach, provides sharper contrast between recall levels and is accomplished by depressing the Exp FI button in alertness graphs in SleepChart.
Consolidation measures how well we consolidate or re-consolidate memories with repetitions executed at any given circadian time. Recall measurements are fast. We get our data on the day of learning. We instantly know if we can or cannot answer questions at the selected time. However, consolidation data may take years to collect. We may review an item today, and need to wait several years before the outcome of the review (consolidation) can be verified. As sleep-and-learning options in SuperMemo are relatively new (timing of repetitions is collected as of 2006), only very large sets of data collected over the periods of many years provide a basis for meaningful consolidation measurements. For that reason, memory consolidation graphs are currently not part of data analysis in SuperMemo. The plan is to introduce those options in a few years when they become usable for a larger proportion of long-term users of SuperMemo.
Exemplary illustration of the speed in which recall drops during a waking day. In this example, the average grade drops from 3.3 early in the day to less than 3.0 after 16 hours of waking.
Interestingly, even a short nap seems to bring the recall back to the baseline level. In other words, there seems to be a direct link between recall and alertness. Recall seems to be inversely correlated with the homeostatic drive to sleep. A slight increase in recall around the 12th hour of wakefulness is a reflection of the circadian component of alertness. The waviness at later waking hours seen in the graph comes from the scarcity of data as learning at later hours makes less sense (of total 31,000 repetitions used to plot the graph, only 684 fell beyond the 10th hour of waking).
An exemplary recall graph displayed by SleepChart shows the decline in grades scored in learning during a waking day. This graph also shows a slight increase in the grades in the second half of the day due to circadian reasons.
Note that both graphs above show a similar time constant of 178 and 172 respectively (half-life of 124 and 119 hours). For calibration reasons, half-life becomes meaningful only when actual recall percentage data is used (in SuperMemo, grade 3.0 is a sharp border between recall success and recall failure).
The decline in the ability to consolidate memories during the waking day follows a curve that mirrors the decline in the ability to recall things from memory!
Exemplary relationship between the circadian time (hours from waking) and the ability to consolidate memories (expressed by an average grade scored in the next repetition)
As the day goes on, the ability to store facts in memory declines. A repetition in SuperMemo is a single effort to recall previously learned information from memory. The graph has been constructed by correlating the circadian time of one repetition (in reference to waking time), and the grade scored in the successive repetition of the same piece of information. The successive repetition often takes place months or years after the repetition for which the consolidation time was registered. Again, short naps seem to restore the memory consolidation power to baseline. As much as recall, consolidation seems to be inversely correlated with the homeostatic drive to sleep. A slight increase in the quality of learning can also be seen around the 12th hour since natural waking (in the presented case).
The conclusion is that in free running sleep (i.e. primarily in the absence of an alarm clock), we can get best learning results if we learn early in the morning. The same holds for exams. The recall and exam results will be best if the exam is held in the morning even though some time for pre-exam cramming may skew the outcome.
Correlation between recall and consolidation
The fact that both recall and consolidation curves seem to follow a very similar course during a waking day seems to indicate that they both may depend on the same underlying mechanism. This conclusion is amplified by the fact that recall is a passive process, while consolidation is an active process of forming new or reconsolidating old memories. We can hypothesize that the underlying mechanism is therefore not molecular. The decline in recall and consolidation might simply be caused by a decline in operational efficiency of the neural networks involved in learning. That efficiency, expressed as alertness (see: Alertness in SuperMemo), depends on both homeostatic and circadian components of the sleep drive. The homeostatic component determines an overall decline in network efficiency over the course of a waking day, while the circadian component allows of a small bump in the second half of the waking day, presumably due to a neurohormonal impact of the circadian cycle on the overall function of the central nervous system.
Good learning days
The correlation between recall and consolidation can also be seen in abstraction from the circadian time. If the overall recall and consolidation data are taken from individual days of learning process, they correlate pretty well too:
Exemplary graph that shows that learning days that are good for recall are also good for memory consolidation. Recall is expressed as a fraction of correct answers on a given day. Consolidation is expressed as a fraction of correct answers on the day of the next repetition that follows the one on the day for which the consolidation is measured.
In future versions of SuperMemo, the user will be able to see the strict correlation between his or her own recall and memory consolidation:
Exemplary graph showing how good memory recall improves memory consolidation. The relationship between recall and consolidation is nearly linear. The graph was plotted using over 800,000 repetitions in SuperMemo, with 538,000 of these contributing their data to the consolidation estimates. Recall levels with fewer than 3000 data points have been omitted from the graph. The Deviation parameter says how well the linear fit matches the data (the less the deviation, the better the fit). The deviation is computed as a square root of the average of squared differences between the approximation and the data.
Exemplary graph showing the average recall for days producing a given level of memory consolidation. The relationship between consolidation and recall is nearly linear. The graph was plotted using over 800,000 repetitions in SuperMemo. Consolidation levels with fewer than 3,000 data points have been omitted from the graph. Lowered recall for consolidation of 100% comes from the fact that this consolidation level is overrepresented by small sample days where lucky perfect recall in just a few items may result in perfect consolidation reading without actually saying anything about the recall on the day the consolidating repetition took place. Sufficiently large number of such cases will let consolidation category of 100% pass the 3,000 data points outlier limit set for this graph, and result in a recall level that is much closer to the average level.
Alarm clock vs. learning
There is an urgent need to collect sleep data from subjects who disrespect healthy sleep in various ways. The most interesting area for further investigation is how poor sleep hygiene affects learning. As an example, let's have a peek at an interesting graph showing the average recall of a teenager who often needs to get up early for school, far ahead of his natural waking time. If grades are converted to the forgetting index, we can see that this student forgets 53% more on schooldays when he needs to get up early. This is a very preliminary sample that should not be used to draw far-reaching conclusion (for example, more learning occurred in earlier hours on days free from school), however, it is my hope that with more data pouring in, we can tangibly demonstrate the disastrous impact of early school times on learning. In other data sets, it has also be found that later waking time (after 11 am) often correlates with lower grades as well (perhaps as a result of weekend late "partying" that results in poorer sleep and later awakening).
Learning in free running sleep
Everyone has his or her own optimum learning hours that depend on the circadian rhythm. For most people, optimum learning occurs in the morning and after a siesta. Non-nappers also improve their learning in the evening due to a circadian upswing. However, the exact timing of those optimum periods can only be determined on an individual basis. The disconnect between the optimum learning time and the absolute clock can be seen in a regular free running sleep rhythm as in the analogous graph below that does not show any hours (on the clock) in which learning is more efficient:
However, when the free running sleep data presented in the graph above is processed using the circadian time rather than the clock time, a typical two-peak circadian pattern re-emerges with good grades in the morning, siesta dip, and an evening upswing. The circadian phase estimations have been generated with SleepChart. The peak learning times are usually separated by 10-13 hours:
It is obvious that alertness improves learning. However, it is worth noting that even marginal improvements to high alertness can yield major benefits to learning. In other words, it is not enough to be alert. Crisp alertness might substantially improve learning as compared with just being ok. In the presented graph, sleep propensity has been estimated with SleepChart using the two-component model.
The more time we spend learning on a given day, the lower our learning capacity is. Recall decreases along a homeostatic increase in sleepiness. However, it decreases much faster when the learning process continues. In other words, learning increases sleep propensity. That observation agrees nicely with the complementary encoding theories that explain how the brain copes with catastrophic forgetting that occurs in artificial neural networks. Those theories speak of secondary memory systems used to redistribute knowledge originally stored in low-interference short-term networks. The act of storage redistribution is hypothesized to occur during sleep. In other words, as you keep loading your memory with knowledge, your brain turns on a defense mechanism, makes you drowsy, and sends you to an earlier sleep. This is why, against conventional advice of sleep experts, I recommend SuperMemo to insomniacs (if they must go to sleep early). Except where the circadian component of sleepiness is missing, learning is a good tool for boosting homeostatic sleepiness. Obviously, it will not work in cases like learning before an exam, which may subconsciously be associated with stress.
Average grade in learning with SuperMemo depends on the position of the tested item in the learning queue. Later items receive lower grades. To eliminate the impact of the homeostatic sleep propensity, all repetitions studied took place in the hours 5-7 of the waking day.
Sleep might be the chief anti-overload protection mechanism. The hypothesis says that sleep helps unload separated neural representations from the hippocampus. It optimizes the long-term neocortical overlapping representation. Learning with a fresh mind after a good night sleep will then be recommended. Learning in condition of sleep deprivation or mental fatigue would then be a mistake (unless employed as an anti-insomnia tactic).
Robin Clarke who hypothesized that too much learning can cause Alzheimer's (Wozniak 2002) writes: "Natural selection will favor further mechanisms, which enable local matrixes nearing overload, to signal their lack of spare capacity, thus activating diversion to other locations". This sounds exactly like the job of NREM-REM sleep interplay. Optimizing the storage is the simplest defense against memory interference. Sleep may act as an anti-overload and anti-interference mechanism that does not show the same destructive powers as forgetting. The signal on the "lack of spare capacity" might simply be adenosine-based homeostatic component in the two-process sleep model. See also: Neural optimization in sleep
Alertness vs. learning
As shown in the preceding sections, in healthy individuals who are not sleep deprived and who sleep in the correct phase, the best learning results are obtained early in the morning. This easily reproducible observation was an incentive to introduce two options in SuperMemo that help users of the program study their alertness throughout the learning day. The term alertness, in SuperMemo, is used interchangeably to describe two different measures of cognitive function: inverse of sleep propensity (or sleep drive) as derived from the two component model, and the average grade in learning with SuperMemo which corresponds with memory recall. Both expressions of alertness are closely correlated. SuperMemo measures alertness as well as attempts to predict changes in alertness in two different time frames intended to separate the homeostatic and circadian components of sleep propensity. Both approaches require a sleep log for the measurements and for the predictions to be possible. To demonstrate the homeostatic changes to alertness, SuperMemo measures the learning performance since the last sleep episode. To demonstrate the circadian changes to alertness, SuperMemo measures the learning performance in reference to the circadian time (i.e. time measured since the optimum natural waking hour) in periods that may or may not include intervening sleep episodes. As it can be seen in the enclosed pictures, it is not possible to fully deconvolve the impact of homeostatic and circadian sleep propensity on learning. Homeostatic graphs will always include a small circadian bump related to post-siesta learning, while circadian graphs will be affected by sleep habits that are closely correlated with the circadian cycle, esp. in free running sleep.
If you have already collected your sleep data with SleepChart, you can see your wake-recall correlations with the newest SuperMemo. Note that only repetitions executed with SuperMemo 13.0 (2006) or later will be included in the graphs as earlier SuperMemos did not store precise time of repetitions in repetition history.
You can see how fast your alertness, recall and grades drop during the day by inspecting the Alertness (H) graph in SuperMemo. In this graph, you can see the time that has passed since the last sleep block, and how your recall changes in waking:
Alertness (H) graph makes it possible to visually inspect how recall decreases during a waking day. It also shows the impact of circadian factors with grades slightly lower immediately after waking and slightly higher in the post-siesta period (i.e. in the 10-13 hour bracket). The Deviation parameter displayed at the top tells you how well the chosen approximation curve fits the data (in the picture: negatively exponential recall curve). The lesser the deviation, the better the fit. The deviation is computed as a square root of the average of squared differences (as used in the method of least squares).
In Alertness (H), the minimum length of a sleep episode in consideration is determined by Min. sleep block (h) box (0.2 hours, or 12 min. is the default minimum). Shorter sleep blocks are disregarded in plotting this graph. Homeostatic alertness half-life (in hours) tells you when your learning capacity drops by half after waking. You can modify this parameter to look for a better curve fit in your case (the Model button must be depressed). See Deviation to evaluate the fit. This half-life can differ between individuals. Notably, it is very short in narcoleptics, and very long in natural non-nappers.
The circadian changes in alertness can be seen in the Alertness (C) graph, which plots alertness throughout the day in reference to the circadian time measured from the actual waking time or from the optimum natural waking time:
Alertness (C) graph showing the powerfully biphasic nature of the human circadian cycle. The horizontal axis shows the circadian time, i.e. the time that elapses from phase 0, i.e. the predicted "end of the night" time (if Model is depressed). The prediction comes from the circadian model employed in SleepChart, and is derived from the sleep log data. The yellow line is the predicted circadian alertness derived from the same sleep log data using the two component model of sleep propensity developed for the purpose of sleep optimization in SuperMemo (inspired by similar work by Alexander A. Borbely and Peter Achermann). The overall alertness, not shown in the graph, is the resultant of the status of the two components of sleep propensity: the homeostatic component and the circadian component. The blue dots are recall data taken from the learning process in SuperMemo that correlate well with overall alertness
How learning affects sleep?
Impact of learning on sleep
In many of my older articles I often mention the fact that learning should increase the demand for the total sleep time. I read about the impact of learning on sleep yet in the 1980s. I have since lived with the conviction that this is a science fact that is as obvious as the fact that sleep is essential for learning. However, when I tried to prove the claim with data collected with SuperMemo, I discovered that it was not as easy as I thought.
When I tried to see if prior learning increases the length of sleep, I found the opposite. Again I started with my own sleep and learning data, which is rare in its size and the fact that the free running condition applies to both sleep and learning. I explained free running sleep earlier in the article. By "learning at libitum" I mean learning that, for the sake of efficiency, is more intense and long lasting on good learning days, and less intense on worse learning days. Good and bad learning days are primarily determined by the quality of sleep, and not, for example, availability of time. I thought that the free running condition is essential for such investigations, esp. sleep should not be controlled artificially so that to make sure that increased demand for sleep is reflected in total sleep obtained.
It appears that in a free running condition, the days with lots of learning were followed by less sleep in the night!
The amount of sleep obtained in the first 11 hours from bedtime as a function of the amount of learning in the last 8 hours preceding the bedtime.
Upon a closer inspection, it appears that the reason for this surprising outcome is that if learning is done on demand, i.e. more learning on good learning days, prior quality of sleep determines both the amount of learning and as well as the total sleep on the following night.
I tried to correct for prior total sleep to get a better picture. You may recall from the section devoted to napping that the amount of napping correlates well with the prior night's total sleep (the less sleep, the more napping). If I could find a similar neat relationship between the sleep on two successive nights, I could perhaps correct for sleep debt and reveal if more learning entails more sleep.
However, the relationship between total sleep on two successive nights is also pretty surprising. For example, the following U-shaped relationship shows the amount of night-time sleep depending on the total sleep in the preceding 20 hours.
Exemplary U-shaped relationship of total sleep and sleep on the preceding night. Total sleep on the vertical axis is taken as the consolidated night-time sleep (i.e. sleep in which short-lived nighttime awakenings are ignored). The horizontal axis represents total sleep whose termination point is embraced by the 20 hour margin preceding the bedtime in consideration. This margin was chosen to capture the preceding night sleep as well as follow-up naps without reaching into areas of sleep that should be considered two nights away from the period of interest.
The U-shaped graph shows that a simple sleep debt formula cannot be used to correct for sleep demand after a day of learning. However, a subset of normal-length nights could be used to filter out for varying sleep debt conditions.
As for the explanation of the U-shape obtained, it might be a combination of three main causes:
- Most obviously, short sleep on one night will often result in longer sleep on the successive night (e.g. as with total sleep from 0-3 hours on the preceding night).
- Some factors that determine the length of free running sleep may span over periods longer than a single day (e.g. health status, season, humidity, availability of sunlight, etc.). Those factors will result in a positive correlation between total sleep on successive nights (e.g. as seen in the graph in total sleep spanning 5-8 hours).
- Extremely long days, with more than 20 hours of wakefulness, will result in failing to register the preceding sleep on the graph (i.e. the preceding night sleep will equal zero).
Using data on the relationship between the length of sleep on two successive nights, we can apply a "band filter" on the data used to generate the first learning-vs-sleep graph. If we eliminate short-sleep nights by choosing only data points with total preceding sleep equal to five or more hours, we can reverse the downward trend and produce a nearly flat linear relationship between learning and the follow-up sleep:
If sleep on the preceding night is above 5 hours, then the amount of learning has nearly no impact on the follow-up sleep
If the "bandwidth" is narrowed to 5.0-6.5 hours, we get a perfectly flat line (slope=0.00). This data seems to indicate that an increase in learning does not increase the total sleep on the follow-up night.
As there are many lines of evidence that learning does affect the follow-up sleep, there could be many explanations of that conclusion. Sleep density might change instead of the length of the night sleep episode (as it is the case with REM density (Smith et al. 2004). In an active lifestyle, learning may not increase the demand for sleep much above the baseline. Last but not least, the result may differ between students. Some students can swear that more learning requires more sleep in their case. I am yet to receive an appropriately large set of data that could demonstrate this fact. As much as free running sleep makes it impossible to prove that short sleep is bad for learning, learning on demand may make it impossible to prove that lots of learning increased the demand for sleep. As much as alarm clocks can be helpful in showing their own bad impact on learning, forced learning may also be a more grateful research subject. Forced learning may be more costly for the brain and show a more pronounced impact on the density and length of sleep. Perhaps learning needs to be heavy enough to notice the effect due to the fact that all our waking experience is a form of learning, even if we do boring repetitive activities. A mere thought process, e.g. recalling a relative, will form new memory traces in the brain. These will be processed in sleep. For sleep demand to come well above the baseline, learning must come above its own baseline as well.
Sleep and school
Schools have changed the world for the better. Literacy is on the increase worldwide. However, there is one huge factor that holds schools back: sleepy kids!
Sleepy kids learn little!
Modern lifestyle results in an epidemic of delayed sleep phase disorder in the adolescent population. Millions of families nowadays struggle with putting their kids to sleep early enough, and to have them wake up fresh in time for school. It seems like we are losing this battle worldwide. Kids seem to be getting less and less quality sleep! Drs Amy Wolfson and Mary Carskadon study sleep in teenagers. They were horrified to find out that sleep latency during school hours was lowest for 10th graders and was a shocking 1.8 minutes! This latency is lower than the value a good sleeper usually achieves at bedtime! In other words, kids are more ready for sleep at school than a normal individual is ready for sleep in the night!
Learning in such sleep deprived state is worth little more than zero!
This is an alarming situation that can undermine the future of education as well as the physical and mental health of the next generation! Some sleep researchers ring the alarm bells, others look for remedies. I do not have a prescription for the problem. Hereby I would only like to appeal for more tolerance and understanding on the part of parents and schools. All my life I have worked for the purpose of better education for everyone. However, there is no learning without sleep. Sleep is important enough to often take precedence over the education itself! My appeal is:
The premise of this appeal is very simple. Waking up a semi-conscious kid for school implies a day that is practically wasted for learning, or literally crossed out from a young life's calendar. Adding those extra 2-3 hours of sleep means that the kid will only miss a class or two, with many additional productive hours left in the day! It is by far better to spend an hour on productive learning than to spend 8 hours on comatose "survival through the class". It amazes me how little this simple truth is appreciated! When I speak to parents, they always excuse early waking with "there would be consequences for missing the class"! There must not be any consequences! Sleep deprivation shrivels the brain! Sleep is the fundamental human right of a developing brain. If someone threatens the kid with "consequences", you need to combat that attitude. Sadly, for many parents, the timing of the early morning schedule is determined by work and other obligations that cannot be worked around.
Excessive school workload
One of my favorite journalists, Fareed Zakaria, spoke in his GPS program about his prescription for better education: "Some elements of the solution seem obvious. The writer Malcolm Gladwell says it takes 10,000 hours to get really good at anything. It's really just another way of making Thomas Edison's famous point that genius is 1 percent inspiration and 99 percent perspiration. Now if our kids spent two years less in school than in many other countries, they will find themselves behind in many areas. We don't have to go to the lengths that South Korea has gone to lengthen the school day and the school year, but we can't do the least work and hope for the best results" (source).
The problem with this "Korean solution" is that it fails to account for a dramatic difference between good learning and bad learning. Given a shortage of good teachers, good funding, good methodology, etc. we might as well pump up school hours in hope of converting quantity into quality. However, a good hour of self-learning or a good hour of customized one-on-one tutoring is worth more than 10 hours of boredom in an average classroom. Perhaps Finnish schools with their reliance on excellent teachers would show a better ratio. If we added just two factors to our school systems: (1) good sleep and (2) spaced repetition, we could safely cut school hours to 1-2 classes per day and still get better results!
School hours and homeschooling
Due to a well-documented sleep phase shift at adolescence, teenagers find it more and more difficult to solve the problem of sleep deprivation by just going to bed earlier. Instead of providing for longer sleep, early bedtimes may result in insomnia and a multitude of psychogenic sleep and emotional problems. Teens are simply unable to fall asleep at designed early time, and trying to force them to do so may actually backfire. Even a mild degree of sleep deprivation might be better than hours of tossing and turning, or nocturnal awakenings. Return to a farmer's lifestyle would remedy the problem of teenage body clock, however, this would mean many hours of physical work in the field from the early morning. Sitting in a school bench just won't do. Heavy load of schoolwork on its own contributes to the late sleep phase lifestyle!
When schools experiment with later class hours to accommodate the adolescent body clock, they get better learning results (Wahistrom 2002). Traffic accidents among young drivers on the way to or from school also drop (around 25% for a mere one hour shift clockwise). However, it appears that kids just tend to adapt and stay up later in the night. Later school hours are an imperfect remedy, esp. that kids differ by chronotype and each will have its own optimum window for the best learning performance. Callan (1998) reported that in high school, less than 10% of kids preferred the early school hours, while 15% preferred evening hours. Reported preference is often confused by the misalignment of circadian cycle with the waking period, which often makes evening types claim evening is better for learning, while in free running sleep, the same kids would prefer the subjective morning hours (with "morning" coming as late as mid-day). Moreover, as the kids get older the predominance of eveningness starts becoming more pronounced.
I do not know a universal solution, however, all parents should consider homeschooling, which could make a world of difference. Not every parent is qualified, and not everyone can afford it. Amazingly, some modern and progressive countries banned homeschooling altogether. It is hard to believe, but two leaders in the adoption of rational and scientific social solutions, Germany and Sweden belong to that group! In fear of dangerous ideologies, some governments block a return to a tradition that is as old as the human race. A tradition that could remedy many weaknesses of the school system: tutoring one-on-one under the supervision of the most loving people in existence: own parents or other family members. Homeschooling makes it easy to employ the most efficient of the learning methods: self-paced self-directed exploration based on passion and curiosity. This ideal solution solves the problem of matching learning hours with the circadian cycle.
Sleep deprivation in kids
Most kids wake up earlier than they would prefer to. This results in sleep deprivation and a set of negative consequences:
- bad learning: cutting down on sleep dramatically impairs learning. Some groups of schoolchildren show sleep latency of just 1-2 minutes at the time when they sit in the class. This is catastrophic! This is a latency that many would be envious of at evening bedtime! Kids should learn at a time when sleep initiation is near-to impossible! In free running sleep, that would be the first 3 hours of the waking day! After that, learning is impaired and PE classes and lunch could come in. Lots of kids sleep in the class, or are solely preoccupied with "surviving".
- bad emotions: sleep deprivation results in irritability. Without the neural network cleanup executed in sleep, the brain quickly gets overloaded and overwhelmed with the stimuli. This could be learning stimuli, or minor annoyances, such as a colleague's jokes. Bad temper follows even in otherwise well-mannered kids. Comraderie and social interaction are replaced by bullying, fighting, aggression, and sheer meanness. Over many years this can lead to psychological problems, depression, aggression, suicide, and the loss of ability to harmoniously integrate with the rest of society.
- hate of school: by the age of 10, most kids universally hate school! One of the first questions I ask any kid I meet is about his or her fondness for school and learning. If they claim to like school, it is often because of the chance to interact with friends or simply to break free from parental supervision or leave the home environment. Vacation countdown becomes a daily preoccupation. This has disastrous effects on the efficiency of school education, and long-term choices such as going to college. Universal hate of school shapes a generation and the way society copes with challenges of the modern world. And it all begins with the malignant device: an alarm clock!
Poor recall on schooldays
To illustrate the impact of school hours on learning, see the following exemplary graph. A 16-year old high school student logged his sleep patterns in SleepChart and his learning results in SuperMemo. By combining the two we can see the relationship between the waking time and the average grade obtained in learning with SuperMemo. The waking time for school was always ahead of the natural waking time and the teen compensated by sleeping longer on weekends:
Despite a decline in the learning performance on schooldays, the teen would do great at school, do his best learning during weekends, and would later get admitted to an Ivy League school. The dramatic impact of sleep deprivation on learning can be seen when grades are converted to the forgetting index. In this case, the students would forget 53% more on schooldays when he needed to get up early. Clearly, sleep deprivation is not likely to deprive someone of a chance to get to the Ivy League. However, it does affect the performance and undermines a young man's potential. At younger ages it may also have a significant impact on the brain development. Interestingly, in other data sets, I have also found that later waking up (after 11 am) often correlates with lower grades too. Perhaps that is a result of weekend late "partying" that results in poorer sleep and later waking?
Example #1: Long weekend sleep
A typical sleep pattern with short weekday sleep and long weekends sleep is shown in the following sleep log and the corresponding circadian graph.
Exemplary sleep log with weekday sleep deficits and longer sleep on weekends. Typically, Saturday morning sleep is longer than the Sunday morning sleep.
Circadian graph for sleep with weekday sleep deficits and longer sleep on weekends. The graph shows that a day of 16 waking hours and 8 hours of sleep would probably make the desired optimum. Instead, the 7 hour night causes an accumulation of sleep deficit with sleep cut short by one hour per day on weekdays.
Example #2: Phase 12 napping
A more troubling example shows a fragmentation of the sleep schedule caused by short night sleep episodes, and frequent Phase 12 napping. Here a student attempts to sleep in the exactly same brackets, i.e. 23:00 - 6:00:
Circadian graph for short night sleep with irregular napping. Naps are taken ad hoc in various phases. Early naps are short and do not cover for sleep deficits. Late naps cause a delay in night sleep, and possibly a phase delay that compounds the problem of sleep deficits.
Learning in alpha state
There are learning gizmos and contraptions out there, which are marketed as based on learning in a relaxed state. Proper cognitive environment is paramount to learning. However, for clarity, we should rather use the term concentration instead of an all-inclusive relaxation. Concentration in learning should be maximized by taking into account the following factors:
- being cut off from all sources of interference in learning (telephone, e-mail, conversation, radio, and perhaps even one's favorite music)
- finding the optimum circadian timing for learning (e.g. early in the morning in free running sleep cycle, late in the evening in DSPS individuals who cannot afford free running sleep, etc.)
- all aspects of mental and cognitive health (e.g. avoiding stress, substance abuse, etc.)
The concept of relaxation is often associated with alpha wave learning which has attracted lots of companies that are more interested in their bottom line than their customers' actual success in learning. EEG measurements can be used to roughly determine the current state of the brain in the same way as you could detect bustling activity in a major city by scanning the surrounding electromagnetic field. The usefulness of alpha wave scanning in learning can be compared to the usefulness of electromagnetic field scanning for social life of a city. You need to focus on the causes rather than on symptoms. Alpha waves appear primarily in the absence of visual processing and other intense mental processes. This is why they cannot dogmatically be considered a desired learning state. After all, the drowsy alpha state that precedes falling asleep is exactly the worst moment for learning during the day.
In evaluating the "relaxation products" you need to differentiate between the relaxation effect and the actual learning effect. The number of companies making false claims in this field is astounding. It is very easy to fall for a simple solution to a learning problem (e.g. get 10Hz binaural beat difference and your learning problem will go away for life, and perhaps your sex drive will improve at the same time, you will sleep better and you will look younger). The easy learning solution explains why false claims related to "learning in relaxation" are so hard to extinguish.
At the same time, if you need to cope with stress or insomnia, many products in the field may have a legitimate application. Customers of the Polish Sita system jokingly claim that the company would do better if they marketed their product as a napping system. A worthy application on its own. In the 1990s, I appealed to users of SuperMemo to let me know of relaxation products that might be worth mentioning as an effective help in learning. I do not think I have received any credible suggestions until now.
Learning during sleep
When Soviet researchers made a claim of sleep-assisted instruction, they started a powerful meme that could never be reproduced and is now pretty hard to extinguish. You may have heard of sleep tapes that offer effortless learning during sleep. They are a direct follow-up of the Soviet claims and only a part of the whole series of products for learning in sleep. Your investment in tapes for learning in sleep will not be money well spent. Attempts at learning during sleep should be discouraged! It is possible to occasionally recall a fraction of the material presented during sleep. Information may reach and register in memory during short periods of awakening or transition from REM to shallow sleep. There is also ample evidence that some circuits in the brain can be conditioned during REM sleep. However, the connection between the senses and the brain in sleep is rather focused on awakening in danger rather than on processing complex information.
Whatever you might gain from your sleep tapes will by far be offset by damage to the quality of sleep. If the learning stimuli do not reach a certain threshold, they will simply be ignored. However, past a certain value they may prevent the progression of NREM sleep toward stages 3 and 4. They can also shorten REM sleep.
Interestingly, memories acquired minutes before falling asleep do not get consolidated! Even a few minutes of sleep leave a short window of waking time that is totally erased from memory. Luckily, we rarely learn mission-critical information shortly before dozing off.
Counter-recommendation for learning during sleep, does not imply that falling asleep with TV or radio turned on should be discouraged. If you would like to get a dose of education yet before falling asleep, be sure your tapes, TV or radio meet these conditions:
- they turn off automatically no later than in 10-20 minutes
- they have no ability to wake you up from a properly timed sleep. If you wake up in the initial minutes of sleep you may experience a dramatic drop in homeostatic sleepiness that would delay the sleep onset. Awakening may also indicate that you went to sleep too early in reference to your circadian cycle
- they do not include highly emotional content, distressing messages, shrill sounds like doorbells, phones, timers and alarm clocks, as these all have been designed to most effectively interfere with sleep for the purpose of interrupting it
Moreover, if you find it difficult to fall asleep due to the stresses of the day, subtle news channel may actually help you fall asleep by keeping your mind away from the thoughts that might trigger the release of ACTH, cortisol, catecholamines, or other alertness hormones.
TV, radio or tapes in the morning are OK too, on condition you turn them on manually (i.e. they should not work as an alarm clock substitute). If you wake up slightly ahead of your expected waking time, turn on the news and stay in bed. Test your brain for signs of sleepiness. Occasionally, you may still be able to fall asleep and go through one cycle of sleep that will be beneficial to your intellectual performance.
Some self-help personal power gurus keep bringing up the concept of lucid dreaming as a tool for enhancing learning and creativity. Terms such as super-consciousness or hyperreality are tossed around. Polyphasic sleepers often claim that the Uberman sleep schedule helps them achieve lucid dreaming and an enhanced experience of reality. There might be a grain of truth in that claim. Read about the polyphasic rollercoaster to understand why polyphasic sleeper might experience euphoric highs that seem even higher due to the periods of total zombification. Lucid dreaming is as useful for learning and creativity as LSD. Striving at lucid dreaming is rather likely to disrupt the healthy sleep and negatively affect learning. During REM sleep, the prefrontal cortex should normally be de-activated. Hobson's AIM model of 3D sleep-wake space (Hobson et al. 2000) can be used to illustrate the state corresponding to lucid dreaming as a partitioning, in which the cortex and the rest of the brain occupy different points in the AIM space. Such partitioning is likely to interfere with the physiological function of REM sleep. It can be compared to eating your lunch while jogging (i.e. the situation where contradictory targets are fed to the nervous system). Using auto-suggestive tricks to change the AIM state may affect neural processes occurring in sleep with unpredictable consequences that are not likely to be positive. As for creativity, it is conceivable that LSD (and less so lucid dreaming) might boost non-specific creativity or help understand the creative process. However, most of the mankind's creative breakthroughs occur when a healthy refreshed mind focuses on solving a specific problem. Hallucinatory haze is not helpful in directing creativity towards a useful purpose. Creativity is a game of chance. You should look for ways of consciously directing the creative process rather than to increase its randomness indiscriminately (Wozniak 2001). For some tips, see natural creativity cycle and How to solve any problem?.
- Wozniak P.A., "Can too much learning lead to Alzheimer's?" (May 2002)
- De Koninck J., Lorrain D., Christ G., Proulx G., Coulombe D., "Intensive language learning and increases in rapid eye movement sleep: evidence of a performance factor," International Journal of Psychophysiology / Volume 8 / Issue 1 (September 1989): 43-47, doi: 10.1016/0167-8760(89)90018-4
- 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
- Lawton M., "Too Little, Too Late," Education Week
- Wahistrom K., "Changing Times: Findings From the First Longitudinal Study of Later High School Start Times," NASSP Bulletin / Volume 86 / Issue 633 (December 2002): 3-21, doi: 10.1177/019263650208663302
- 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
- Wozniak P.A., "The roots of creativity and genius" (2001)