Past studies have found that rewarding participants during a visual perceptual task leads to performance gains.
However, new research suggests that these performance gains occur only if participants follow up the task with sleep.
The new findings may have particular implications for students tempted to sacrifice sleep in favor of late-night study sessions, said study corresponding author Yuka Sasaki, a professor of cognitive, linguistic and psychological sciences at Brown University.
“College students work very hard, and they sometimes shorten their sleep,” Sasaki said. “But they need sleep in order to retain their learning.”
In the study, published this month in the Proceedings of the National Academy of Sciences, young adults were asked to identify a letter and the orientation of a set of lines on a busy background.
Some participants were told to refrain from eating or drinking in the hours leading up to the task and were then given drops of water as a reward for correct responses.
In contrast to groups that were not rewarded during training, rewarded participants exhibited significant performance gains – but only if they slept after the training session.
This finding suggests that reward doesn’t improve visual perceptual learning until people sleep.
The researchers believe that reward (or anticipation of reward) reinforces neural circuits between reward and visual areas of the brain, and these circuits are then more likely to reactivate during sleep to facilitate task learning.
Indeed, during post-training sleep in rewarded participants, electroencephalogram (EEG) recordings found increased activation in the prefrontal, reward-processing area of the brain and decreased activation in the untrained visual areas of the brain.
That pattern of activation can likely be explained by past studies, which suggest that the prefrontal, reward-processing area of the brain sends signals to inhibit some of the neurons in the visual processing area.
As a result, irrelevant connections are trimmed and the most efficient connections are preserved, and task performance improves.
The study also examined when the patterns of activation occurred. Untrained visual areas of the brain exhibited reduced activation during both REM and non-REM sleep, but prefrontal, reward-processing areas became active only during REM sleep.
REM sleep appears to be particularly important for task learning — likely because connections are reorganized and optimized during this sleep stage — and it may be linked to the activation of reward-processing areas of the brain.
Consistent with this theory, the rewarded study participants exhibited longer periods of REM sleep compared to those who did not receive a reward during training.
This finding suggests that reward doesn’t improve visual perceptual learning until people sleep. Ima
Sasaki added that physical-based rewards, like food and water, may have a stronger impact on neural circuits compared to rewards such as money.
“Water deprivation may be fundamental,” she said. “When you’re really thirsty and you get water as a reward, the impact of that reward may be more prevailing to the brain.”
Future research could examine whether other types of learning, such as motor and associative learning, also benefit from the interaction between reward and sleep.
Going forward, Sasaki hopes the study will encourage collaboration between sleep researchers and scientists studying reinforcement learning.
“Reinforcement learning is a hot topic in neuroscience, but it hasn’t interacted much with sleep research,” she said. “So this could lead to more interdisciplinary work.”
In addition to Sasaki, other Brown University authors on the study were Masako Tamaki, Aaron V. Berard, Tyler Barnes-Diana, Jesse Siegel and Takeo Watanabe.
Funding: The study was funded by the National Institutes of Health (R21EY028329, R01EY019466, R01EY027841 and T32EY018080) and the United States-Israel Binational Science Foundation (BSF2016058).
Each evening we fall into an offline state defined by a diminished responsiveness to the environment, attenuated movement, intrinsically organized brain activity and bizarre thought patterns. It is during these sleep epochs that brains stabilize important, and erase unnecessary, information gleamed from waking experience in an attempt to ensure adaptive future behavior and efficient management of neural resources.
What follows is a general overview of sleep, oscillations, synaptic plasticity, neuromodulation, known interactions between these processes, memory consolidation and forgetting. Following the general overview of relevant topics, the remainder of this article is dedicated to reviewing what is known about the link between non-rapid eye movement (NREM) and rapid eye movement sleep (REMS) oscillations, memory consolidation and forgetting.
As well as, advocating for a candidate cellular mechanism of sleep-dependent memory consolidation, speculating on potential sleep-dependent forgetting mechanisms and discussing a largely notional model for NREM-REMS cycling wherein sequential sleep stages serve to first lay down and edit and then stabilize and integrate engrams.
Taken together, the aim of this article is three-fold: (I) to convince the reader that sleep oscillations act and interact in a coordinated manner to activate windows of synaptic plasticity; (II) to promote the notion that sleep oscillations, by activating transient windows of synaptic plasticity, mediate bi-directional memory processing; and (III) to advance a primarily notional model for NREM-REMS cycling compatible with objectives I and II.
Sleep is irresistible, necessary for survival and neurobiologically complex. Sleeps heterogeneous architecture can be deconstructed into NREM and REMS stages occupying 80% and 15%–20% of the offline period, respectively. NREMS is defined by a high-voltage, low frequency, synchronous electroencephalographic (EEG) pattern, vague, disconnected, mundane mentation and parasympathetic dominance (low heart rate, slowed breathing; Cabiddu et al., 2012; Carley and Farabi, 2016). Neuroanatomically, brain activity is near globally dampened during NREMS with stage prototypical neuronal activity programs arising from neocortical, thalamic and hippocampal neurons (Staresina et al., 2015). Comparatively, REMS is characterized by a wake like low-voltage, high frequency, de-synchronous EEG readout, vivid emotional mentation, muscular atonia and sympathetic dominance of peripheral physiological processes (elevated heart rate and ventilation). Neuroanatomically, REMS is distinguished from NREMS and wakefulness by strong activation of midline structures, collectively termed the REMS activation area, including the amygdala, pontine nuclei (Peever and Fuller, 2017) and the temporo-parieto-occipital junction, a region believed to be important for dream imagery. REMS also involves a hypoactivation of the frontal lobes (Nir and Tononi, 2010) and theta rhythmicity of hippocampal neurons (Spoormaker et al., 2013).
The ubiquitous nature of sleep in the animal kingdom indicates a robustly preserved biological process. Yet, a concise and accepted core function of sleep has yet to be elucidated. Proposed functions of NREMS include: (I) brain waste clearance by the perivascular glymphatic system (Plog and Nedergaard, 2018); (II) cellular prophylaxis and restoration (Vyazovskiy and Harris, 2013), accomplished by the window of decreased electrophysiological activity/metabolic demands (Van Cauter et al., 1997; Sharma and Kavuru, 2010) and heightened growth hormone secretion (Van Cauter and Plat, 1996) provided by NREMS; (III) systems level memory consolidation (Hardt and Nadel, 2018); this process is discussed in detail in the “Consolidation” section below; (IV) gist extraction, which functions to remove noisy, irrelevant information while preserving important data (Feld and Born, 2017); and (V) synaptic homeostasis, a process organized to renormalize net synaptic strengthen to some baseline following its elevation during waking (Tononi and Cirelli, 2014; Lewis et al., 2018).
Proposed functions of REMS include: (I) aqueous humor distribution (Modarreszadeh et al., 2014). Specifically, ocular movements during REMS may stir up eye fluids to provide nutrients and oxygen to the avascular cornea; (II) an ontogenetic/developmental hypothesis (Marks et al., 1995) suggests that brain activity during REMS aids in connection formation. The ontogenetic hypothesis is compatible with research indicating that REMS is most prevalent during development and decreases with age (Skeldon et al., 2016); (III) the restoration of monoaminergic receptors (Rotenberg, 2006), perhaps accomplished by the period of low monoaminergic tone provided during REMS; (IV) simulation and subsequent habituation of emotional scenes (Cunningham et al., 2014); (V) emotional memory consolidation (Hutchison and Rathore, 2015). Hypotheses IV and V align with the tendency for emotional brain areas such as the amygdala to be hyperactive during REMS; (VI) non-declarative memory consolidation (Cedernaes et al., 2016), a process which may benefit from the ability of procedural memories to be “rehearsed” without consequence during REMS owing to the atonia/paralysis of skeletal muscle (McCarter et al., 2013); (VII) synaptic consolidation, discussed below in the “Consolidation” section (Diekelmann and Born, 2010); (VIII) the integration of newly encoded information into existing information frameworks (Sterpenich et al., 2014); (IX) the formation of novel, non-intuitive connections capable of facilitating future divergent thought via abstraction and extrapolation (Cai et al., 2009; Lewis et al., 2018); (X) the “remembering” or stabilization of not yet consolidated memories; and (XI) the forgetting of previously consolidated memories (Poe, 2017). Hypotheses X and XI will be discussed in detail in sections: REMS Oscillations and Memory Consolidation and REMS Oscillations and Forgetting, respectively. Generally, sleep functions can be delegated to one of two categories: maintenance and repair or information processing. I focus here on the latter category, and in particular on how the sleeping brain accomplishes this information processing function. Which, appears to be by co-opting offline brain oscillations to simultaneously “download/save” important, and “delete/erase” noisy, un-adaptive information.
While neuronal activity in the awake brain largely reflects the processing of information pertaining to the external environment, the electrical activity of the sleeping brain results from internally generated oscillatory patterns (Olbrich, 2010). Oscillations are rhythmic fluctuations in brain activity (Thut et al., 2012) and are crucial for sleeps function as an information processing state. During NREMS the cortex is dominated by low frequency, synchronous delta (0.5–3.5 Hz; Harmony, 2013) and slow waves (0.5–1 Hz; Bellesi et al., 2014), which are intermittently interrupted by several irregular waveforms. These irregular waveforms include sleep spindles (SSs; 11–15 Hz; Purcell et al., 2017), K-complexes (<1 Hz; Lucey, 2017) and sharp wave-ripples (SPW-Rs; 150–200 Hz; Liu et al., 2017).
Of the slower, more synchronous NREMS oscillations, delta waves are primarily generated in the thalamus (Crunelli et al., 2011) while slow waves arise predominately from periodic fluctuations in the membrane potential of cells in the superficial layers of the frontal cortex (Nir et al., 2011; Halgren et al., 2018). Delta and slow waves, sometimes collectively termed slow wave activity or slow oscillations (SOs), consist of alternating depolarizing “up” and hyperpolarizing “down” states (Neske, 2016). The depolarizing SO up state is believed to serve a memory function (Heib et al., 2013) and is the phase of the SO considered throughout much this article. Comparatively, the hyperpolarizing SO down state may provide a period of cellular “rest,” allowing for prophylaxis and restoration (Vyazovskiy and Harris, 2013).
SSs and K-complexes are generated by the thalamus and cortex (Amzica and Steriade, 1998; Mak-McCully et al., 2017). In comparison to the synchronous NREMS oscillations, the role of these irregular electrophysiological phenomena is less clear. SSs are often implicated in memory processing (Mednick et al., 2013; Schonauer, 2018) while K-complexes seem to be important for suppressing neuronal responses to external stimuli that the brain has deemed irrelevant and non-dangerous, ensuring consolidation of the rest phase (Peng et al., 2014). A suppressive function of K-complexes is supported by the finding that these waveforms are evoked by external stimuli (Forget et al., 2011). Further, K-complexes produce a strong hyperpolarization, observable as a large deflection on the EEG readout, well suited for rapidly suppressing cortical excitation (Cash et al., 2009). The other irregular NREMS waveform mentioned, the SPW-R, is comprised of two, interacting waveforms, the hippocampal CA3 sharp-wave (0.01–3 Hz) and the CA1 ripple (Csicsvari et al., 2000). SPW-Rs are considered to have a memory function (Buzsáki, 2015).
During REMS neuromodulation from REMS inducing pontine nuclei, namely the pedunculopontine and lateral dorsal tegmental nuclei, increases neuronal excitability (Ye et al., 2009; Van Dort et al., 2015). This increased neuronal excitability, when combined with the interconnectivity of the cerebral cortex, produces a fast, de-synchronous cortical activity readout. The hippocampus is spared from this de-synchrony and instead exhibits theta rhythmic activity during REMS (Brown et al., 2012); there is some debate as to whether REMS theta rhythm is intrinsic (Goutagny et al., 2009) or is a product of projections from the medial septum behaving as a pacemaker (Stewart and Fox, 1990). Functionally, REMSs cortical de-synchrony and hippocampal theta rhythm are involved in memory processing (Buzsáki, 2002; Sara, 2010). Those NREM and REMS oscillations described as being involved in memory exert said involvement by organizing the brains activity to produce windows in time where the conditions are optimal for synaptic plasticity, the cellular basis of memory processing.
Neurons possess the unique capacity to store information by modifying the strength of their interconnections. In his seminal work, Hebb (1949) suggested that it is the co-activation of neurons by experience that triggers some fundamental mechanism capable of changing the brains connections and, in the process, learning information from the environment. Specifically, Hebb postulated that when two neurons fire simultaneously the connection between them is strengthened, by some “cellular growth” or “metabolic change,” such that the activation of one cell is now more likely to spread and activate the other. The ability of neurons to modify their connection strengths is termed synaptic plasticity.
Since the time of Hebb much has been learned about the neurobiology of learning and memory (Brown and Milner, 2003; Fox and Stryker, 2017). It is now known that the timing of neuronal activations, particularly the relative timing of presynaptic glutamate release and postsynaptic spiking, determines the extent (and directionality, discussed below) of connection strength change.
Activation timing controls the extent and directionality of plasticity by shaping N-methyl-D-aspartate (NMDA) channel open times and thus the magnitude of calcium influx (Feldman, 2012); kinases activated by high levels of dendritic calcium (protein kinase A) and phosphatases activated by low levels of dendritic calcium (calcineurin/protein phosphatase I) are key regulators of synaptic strengthening (Roberson and Sweatt, 1996) and synaptic weakening (Isaac, 2001), respectively (Castellani et al., 2005). The series of timing rules that shape synapse strength changes are sometimes termed Hebbian spike-timing-dependent plasticity (STDP), in honor of the contribution of Hebb’s theories to our understanding of the neurobiology of learning and memory.
Although not a focus of Hebb’s writings, it is now known that synaptic plasticity is a bi-directional process consisting of the synaptic efficacy augmenting long-term potentiation (LTP) and efficacy weakening long-term depression (LTD; Bear, 2003). At the cellular and molecular level synaptic plasticity involves changes in synapse structure and in the number and/or phosphorylation status of membrane channels. Increasing the number and size of synapses increases the total surface area connecting two neurons, while increasing the number of and/or phosphorylating receptors increases a neurons conductance potential (Citri and Malenka, 2008).
LTP is believed to be the principal cellular mechanism by which neurons encode complex information gleamed from experience (Miller and Mayford, 1999; Langille and Brown, 2018). Comparatively, LTD seems to be important for a simpler form of learning, habituation (Glanzman, 2009). Habituation is the process by which an animal decreases their behavioral responsiveness to a repeatedly presented, innocuous stimulus (Rankin et al., 2009). All said, modern experimental research on synaptic plasticity has confirmed Hebb’s (1949) theory of learning.
Additional forms of synaptic plasticity not mentioned here include non-Hebbian and homeostatic plasticity. For more information, readers should consider the following detailed reviews (Turrigiano, 2012; Piochon et al., 2013).
Sleep Oscillations and Synaptic Plasticity
Oscillations are fluctuations in membrane excitability. As eluded to in the previous section, changes in synapse strength typically require the activation/inactivation of membrane channels sensitive to voltage, such as the NMDA receptor. Therefore, oscillations have the capacity to alter the likelihood that various biochemical cascades, acting downstream of changes in membrane channel conductance and upstream of synaptic plasticity, are activated (Hölscher, 1999). As an example, moderate activation of a cortical synapse during the down state of a SO, when membrane excitability is low, may lead to a brief NMDA receptor open time, low magnitude calcium influx and LTD.
Comparatively, the same moderate activation, of the same cortical synapse, during the SOs depolarizing up state may elicit sufficient excitation to cause long NMDA channel open times and high magnitude calcium influx, culminating in the maintenance of connection strength at the activated synapse. A relatively recent study conducted in humans provides support for the above example by demonstrating that SO up state duration is related to the magnitude of post-sleep memory performance enhancement (Heib et al., 2013).
The authors of the study suggest that the correlation between SO up state duration and the degree of subsequent memory improvement results from changes in the amount of time available for the conversion of memory from recent to remote. Heib et al.’s (2013) work bolsters the above example as prolonged SO up states would cause NMDA channels to remain open longer and produce the high magnitude calcium influx necessary for memory stabilizing potentiation. However, recent research by González-Rueda et al. (2018) shows that both maintenance and depression of synapse strength can result from SO up states, adding an additional tier of complexity to the aforementioned example.
These authors suggest that during the up states of NREMS SOs synapses participating in memory associated cell assemblies cooperate to aid in postsynapse firing and as a result maintain their synaptic strengths while synapses which do not assist in postsynapse firing are subjected to weakening (González-Rueda et al., 2018). Notably, dendrite sub-region specific subthreshold input cooperation (Lee et al., 2016) may contribute to the synaptic weight maintenance during SO up states.
Taken together, these findings indicate that during sleep brain waves organize synaptic plasticity such that new or strong memories are safeguarded from, while weak memories are subjected to, SO up state driven synaptic downscaling (González-Rueda et al., 2018). As further exemplification of the relationship between oscillatory phase and plasticity, the REMS hippocampal theta rhythm is, like the SO, a sinusoidal waveform wherein the directionality of plastic change has been documented to depend on the phase (up vs. down state/limb) of the oscillation (Pavlides et al., 1988).
Oscillations may be necessary for synaptic plasticity during sleep. In support of this claim, Durkin et al. (2017) demonstrate that NREMS oscillations are crucial for the synaptic plasticity involved in transferring information captured by the visual thalamus during experience to the primary visual cortex. At least two reasons for the dependency of certain plastic changes on sleep oscillations can be levied.
First, changes in neuromodulation during sleep can alter membrane channel conductance properties. Second, sleep associated changes in transcription can alter the level of intracellular and membrane proteins and thus the composition of the synaptic proteome (Timofeev and Chauvette, 2017). Both of these changes lead to shifts in membrane excitability and thus re-shape the regional activity patterns necessary for activating synaptic plasticity pathways. Sleep-dependent oscillations may be well suited to satisfy these altered plasticity requirements.
Oscillations do not occur in isolation (Steriade, 2006; Jensen and Colgin, 2007). By coupling specific oscillations, the brain can shape what synaptic plasticity pathways are facilitated, in what brain regions and at what times. In the earlier example activation of a synapse during the SO up state was described as producing a “maintenance” level of excitation, not weak enough to trigger LTD nor strong enough to trigger lasting LTP.
However, if a SS is coupled to the SO up state the same activation can produce sufficient excitation to trigger consolidative plasticity (Kim S. Y. et al., 2017), such as LTP. These observations beg the question, why is it that during wakefulness information is readily encoded and during sleep precisely timed oscillatory interactions are necessary to trigger similar synaptic plasticity? The answer may lie in state specific differences in neuromodulation, discussed below.
Neuromodulation Organizes Oscillations and Synaptic Plasticity
Both arousal state and synaptic plasticity are shaped by neuromodulators acting to produce shifts in membrane excitability (Lee and Dan, 2012; Nadim and Bucher, 2014; Palacios-Filardo and Mellor, 2019). Specifically, research by Ding et al. (2016) indicates that changes in the ionic composition of the brains extracellular milieu, shaped by neuromodulators acting on ion channels and transporters, is what shifts the brain through various arousal states by altering neuronal membrane excitability.
During wakefulness, high cholinergic and monoaminergic (histamine, serotonin, dopamine and norepinephrine) neuromodulation increases membrane excitability and overall brain activity producing a neuronal environment conducive for potentiating plasticity (Vyazovskiy et al., 2009; Schwartz and Kilduff, 2015). During NREMS, neuromodulation dampens (Brown et al., 2012) and membrane excitability decreases producing an unfavorable environment for potentiating synaptic plasticity.
Yet, NREMS has a well-documented role in memory consolidation (Cox et al., 2012; Rasch and Born, 2013), a process clearly involving synaptic strengthening. In order to reconcile these seemingly opposite qualities of NREMS one must consider that something other than neuromodulation is acting to gate synaptic plasticity during sleep. Coupled brain waves have been shown to organize cortical dynamics in favor of synaptic plasticity (Zarnadze et al., 2016) and may, through the use of mechanisms akin to those described in the previous section, offer a solution.
In line with this suggestion, neural activity during sleep is primarily driven by intrinsically generated brain waves orchestrated by the principal neurotransmitters glutamate (Hughes et al., 2002) and gamma-aminobutyric acid (GABA; Sun et al., 2012). As an example, during wakefulness thalamo-cortical relay neurons are tonically depolarized by neuromodulation from the ascending reticular activating system, a diverse series of neuromodulatory neurons dispersed within phylogenetically older brain regions, including the brainstem (McCormick and Bal, 1997).
During NREMS, the ascending reticular activating system is silenced and the membrane potential of thalamo-cortical relay neurons moves further from spike threshold. Freed from waking neuromodulation, these thalamo-cortical neurons assume their intrinsic, slow, oscillatory activity patterns driven by hyperpolarization activated ion channels on the neuronal membrane as well as by reciprocal interactions between the thalamus, cortex and GABAergic thalamic reticular nucleus (Brown et al., 2012).
These state-dependent, neuromodulation driven changes in thalamo-cortical neuron membrane potential is what allows the thalamus to readily relay information from the environment during wake and to disconnect one’s mind from the outside world during sleep (Lam and Sherman, 2011; Sun et al., 2012).
During REMS, increases in cholinergic neuromodulation elicit increases in neuronal membrane excitability (Picciotto et al., 2012). REMSs increase in cortical membrane excitability, driven primarily by the modulation of potassium channel activity by acetylcholine (Gottesmann, 2002), produces wake like neural activity (Vakalopoulos, 2014) and an environment favorable for potentiating plasticity (Rasch and Born, 2013).
In sum, during wakefulness and REMS neuromodulation primes the brains neurons for potentiating plasticity. While during NREMS, when excitation promoting neuromodulation is absent, interacting oscillations may replace the role of neuromodulation in the gating of synaptic plasticity.
Information captured by the brain is inherently unstable upon initial encoding and ascends along a trajectory of increasing stability over time by moving through successive neurobiological embodiments. A declarative memory adopts three neurobiological embodiments, or traces, during its tenure: (I) reverberation (activity-dependent or working memory; Constantinidis and Klingberg, 2016; Riley and Constantinidis, 2016); (II) weighting plasticity engram; and (III) wiring plasticity engram.
Working memory is a transient form of information preservation dependent on the reverberation of neural circuits (Nyberg and Eriksson, 2015). Information held in working memory that is either salient, thoroughly processed or both may trigger stabilizing synaptic plasticity, crystallizing held information into activity-independent storage (Dubnau et al., 2003). Weighting plasticity describes the processes through which the hippocampus accomplishes the rapid, initial, limited capacity, interference susceptible, temporary, activity-independent storage of information (Cheng, 2013; Frankland et al., 2013; Preston and Eichenbaum, 2013). Engrams derived from weighting plasticity mechanisms are comprised exclusively of metabolic change (i.e., these engrams result from a re-weighting of existing synaptic connections; Frankland and Bontempi, 2005). Hippocampal anatomy enables rapid information encoding through weighting plasticity as a large portion of all possible inter-neuronal connections exist at any given time (Kali and Dayan, 2000) and can “simply” be differentially weighted to encode information (Takeuchi et al., 2014). The neocortex, by comparison, contains a paucity of all possible connections at any moment in time, owing to energetic, material and network stability constraints. Because of this feature of neocortical anatomy information storage in this region can use both weighting plasticity as well as slower, but more stable, wiring diagram changes to encode information activity-independently (Lisman and Morris, 2001). These wiring changes, which include alterations in synapse morphology and or the number of synapses, are produced through the mechanisms of wiring plasticity (Chklovskii et al., 2004; Holtmaat and Svoboda, 2009). In sum, hippocampal weighting plasticity enables the rapid, temporarily sustained capture of reverberating information while wiring plasticity, in the vast circuits of the cortex, affords a stable, interference resistant, quasi-infinite capacity, long-term storage of information (Alonso et al., 2005; Mednick et al., 2011).
Memory consolidation is the term given to the processes used by brains to move information through the described embodiment or trace succession. These consolidative processes stabilize important information gathered during waking so that it can be used to adaptively guide future behavior. Memory consolidation can be synaptic or systems level and online or offline. Synaptic consolidation describes the biochemical and biophysical stabilization of reverberating information into an activity-independent engram (Diekelmann and Born, 2010). Systems consolidation describes the time protracted process through which information moves from the hippocampus to the cerebral cortex (Born and Wilhelm, 2012; Hardt and Nadel, 2018). Both synaptic and systems consolidation can occur on or offline. Though, it can be argued that synaptic consolidation is favored while an animal is online and systems consolidation is favored once the animal goes offline (Gais and Born, 2004a,b). Mechanisms for offline systems and synaptic consolidation are described below; see “NREM Oscillations and Memory Consolidation” as well as “REMS Oscillations and Memory Consolidation” sections.
Importantly, not all information captured by the brain undergoes the consolidative transcendence from a reverberation to a weighting plasticity engram or from a weighting plasticity engram to a wiring plasticity engram. Information that is not particularly salient or processed meaningfully in working memory, along with information that is encoded in weaker forms of LTP, rapidly deteriorates (Frey et al., 1993; Raye et al., 2002; Curtis and D’Esposito, 2003), often before consolidative processes occur. This kind of “passive” or “natural” forgetting differs from the “active” or “sleep-dependent” forgetting discussed below.
If all the information an animal encountered was retained, and forever, the brains circuits, material limitations, energetic capacities and the ratio of adaptive to irrelevant, non-adaptive information (herein termed the signal to noise ratio) would quickly become saturated. Such saturation would prevent the useful application of acquired knowledge and diminish any capacity for future learning. For these reasons, information stored by the brain needs to be regularly and selectively cleared. The clearance of information, characterized by a decreased accessibility and/or fidelity of said information and termed forgetting, is therefore crucial for healthy brain function, adaptive behavior and learning.
Forgetting can occur in an active or passive manner. Active forgetting is deliberate, organized and acts to selectively clear unnecessary information from the brain. Comparatively, passive forgetting is unintentional and disadvantageous, often causing adaptive information to become decreasingly available and veridical. Notably, the active forgetting described here refers to a series of adaptive, sleep-dependent processes organized for gist extraction and the clearance of unimportant memories and differs from other forms of active forgetting described elsewhere, including intrinsic and motivated forgetting (Davis and Zhong, 2018).
As an example of active forgetting during sleep, Saletin et al. (2011) demonstrate that pre-sleep instruction influences which memories are stabilized or forgotten during a subsequent sleep session. Specifically, the authors found that memories participants were instructed to forget did not experience the same benefit from sleep that memories participants were told to remember did. These findings support the thesis that sleep can carry out forgetting in an active and organized manner, as opposed to simply resulting from global, passive, unintentional clearance operations.
As described earlier, the brain stores information through weighting (post-translational modifications and receptor trafficking; Hardt et al., 2013) and wiring (growth of existing and activation/formation of new synapses; Kandel et al., 2014) plasticity. Neurobiologically, forgetting results from the disintegration of these changes. During forgetting, phosphatases, including calcineurin and protein phosphatase-1 (Sachser et al., 2016), remove post-translational marks while the Rac1/Cofilin pathway allows for the disassembling of structural changes (Davis and Zhong, 2018).
There are at least eight brain mechanisms capable of triggering the neurobiological change associated with forgetting: (I) neurogenesis-dependent circuit rearrangement/overwriting (Gao et al., 2018); (II) pro-active interference (Alves and Bueno, 2017); (III) retro-active interference (Alves and Bueno, 2017); (IV) engram instability (transience of weighting plasticity; Davis and Zhong, 2018); (V) sharp wave replay-induced depotentiation (Norimoto et al., 2018); (VI) homeostatic synaptic downscaling (Tononi and Cirelli, 2014); (VII) theta trough replay (Poe, 2017); and (VIII) low frequency oscillation evoked depotentiation (Tononi and Cirelli, 2006); a summary of these forgetting mechanisms is included as Table 1.
Of the forgetting mechanisms described above, the latter four are sleep-dependent. These sleep-dependent forgetting mechanisms can be realized by the same oscillations known to be important for sleep-dependent memory consolidation: hippocampal sharp-waves (V), cortical SOs (VI and VIII) and hippocampal theta rhythm (VII), as discussed below. Notably, mechanisms VI and VIII are similar in that both can be realized by SOs. The distinction between the two is that homeostatic synaptic downscaling can also arise from mechanisms which are at their core biochemical (Maret et al., 2007; Siddoway et al., 2014; Diering et al., 2017) and independent of network activity during sleep. The four sleep-independent mechanisms listed above, I, II, III and IV presumably also operate during sleep.
NREMS Oscillations and Memory Consolidation
Learning during wakefulness changes the brains activity patterns during sleep, largely in favor of memory processing. The SPW-R waveform, when observed during sleep, reflects the offline, temporally-compressed replay of neuronal sequences active during recent, online learning (Atherton et al., 2015). These rhythms increase markedly in the sleep episode following associative learning, relative to other forms of learning or recall (Eschenko et al., 2008), and improve subsequent performance (Ramadan et al., 2009). Accordingly, disrupting SPW-Rs impairs later performance (Girardeau et al., 2009) and new learning (Norimoto et al., 2018). The memory improvement observed following increases in SPW-Rs comes from the involvement of these oscillations in the inter-regional communication component of memory consolidation (Buzsáki, 2015). Other NREMS oscillations, including neocortical SOs (Marshall et al., 2006), thalamic SSs (McDevitt et al., 2017) and cortical ripples (Khodagholy et al., 2017) have also been implicated in this veridical preservation process.
Like SPW-Rs, SOs and SSs are more numerous in the sleep that follows learning, and act to improve memory processing and performance. More specifically, after learning SO depolarizations increase while down states modify (Mölle et al., 2009). Marshall et al. (2006) and Schneider et al. (2015) demonstrate that augmenting SOs by transcranial stimulation improves memory retention on paired-associate and picture recognition learning tasks.
SSs occur rhythmically and, in addition to becoming more numerous following learning (Gais et al., 2002; Morin et al., 2008), increase after exposure to memory cues (Antony et al., 2018; Cairney et al., 2018). In Antony et al.’s (2018) study, to improve memory cues were presented to sleeping subjects immediately following the inter-spindle period, when new spindles ought to be occurring.
The authors suggest that the rhythmic organization of spindles, into active and refractory (inter-spindle) periods, may serve to deliberately separate memory reactivations in time; how SSs regulate memory reactivations is discussed below. Structuring memory reactivations in this way would help preserve memory fidelity by preventing overlap or undesired integration of reactivated memories, resulting from coincident or lingering plasticity pathway activation. In another study, increases in SS frequency at the same time as cued-memory reactivation were found to be correlated with improved performance on a motor learning task (Laventure et al., 2016).
In regards to ripples, Khodagholy et al. (2017) found that the coupling of hippocampus and association cortex ripples is beneficial for systems consolidation. Similarly, Axmacher et al. (2008) found a correlation between ripples in rhinal cortices and performance on an item recall task learned prior to sleep, when administered again after waking. Thus, SPW-Rs reactivate memories while other NREMS oscillations, including SOs and SSs, are auxiliary to these memory reactivations and facilitate memory processing. As discussed below, these auxiliary oscillations organize windows in time during which information is transferred and engraved into cortical circuits.
Cortical SOs and thalamic SSs cannot consolidate memories independently as they do not have access to recent memories stored in the hippocampus. Yet, SOs and SSs, like SPW-Rs, increase following learning (Gais et al., 2002; Eschenko et al., 2008; Morin et al., 2008) and improve memory (Marshall et al., 2006; Ramadan et al., 2009; Kaestner et al., 2013). These conflicting observations are explained by an interactive model wherein NREMS oscillations couple and this coupling mediates memory consolidation (Staresina et al., 2015). Memories formed in the auto-associative circuitry of the CA3 field spontaneously reactivate as SPWs and influence CA1 ripples, which nest in the troughs of SSs triggered by SOs (Csicsvari et al., 2000; Staresina et al., 2015; Latchoumane et al., 2017). By triggering SSs SOs ensure that the memory reactivations (SPW-Rs) carried in spindle troughs reach the cortex during the depolarizing phase of the SO. Bergmann et al. (2012) have shown that SS activity is highly correlated with hippocampal BOLD signal and that this pairing is concomitant with activity at distinct neuroanatomic loci involved in category-specific prior learnings. An interactive model is compatible with the studies mentioned above, which demonstrate the influence of learning on sleep oscillations and memory performance. Specifically, after learning an increased number of SOs and SSs would have more opportunities to couple or transmit SPW-Rs, resulting in enhanced consolidation and improved performance. In short, during sleep a systems consolidative dialog between memory encoding (hippocampus) and long-term storage (cerebral cortex) structures manifests as an interregional, cross-frequency, EEG signature. Namely, SO up states coupled to spindle trough nested ripples.
Offline, SPW-R memory reactivations couple with SOs and/or SSs to mediate systems consolidation (Mölle and Born, 2011; Staresina et al., 2015; Latchoumane et al., 2017). However, online, isolated SPW-Rs have also been observed (O’Neill et al., 2006) and proposed to, amongst other things, consolidate memory (Joo and Frank, 2018). However, several lines of evidence suggest that these waking SPW-Rs may be insufficient for meaningful consolidation:
I. If waking SPW-Rs were sufficient for consolidation an increased duration of waking, i.e., sleep restriction, should improve memory performance, the opposite of which is observed (Kamphuis et al., 2017);
II. During wakefulness an acetylcholine rich neurochemical milieu prohibits SPW-R mediated hippocampal-neocortical dialog (Vandecasteele et al., 2014);
III. SPWs, absent SOs and SSs, may cause synaptic weakening through engagement of the decoupling force of STDP (Norimoto et al., 2018);
IV. Online consolidative memory replay would be biased by immediate waking experiences. Comparatively, during sleep a brains knowledge can be sampled in an unbiased way (Tononi and Cirelli, 2014);
V. SPW-Rs in awake animals typically occur during exploratory pauses and have been suggested to have a function in trajectory planning and memory retrieval (Gupta et al., 2010; Carr and Frank, 2012; Pfeiffer and Foster, 2013; Joo and Frank, 2018) and;
VI. It is likely that ongoing sensory experience would contaminate, and be contaminated by, consolidation occurring during wakefulness, suggesting this process best occur during sleep (Gais and Born, 2004a,b; Chokroverty, 2017).
Thus, even if waking cortical activity is strong enough to stand-in for NREMS oscillatory couplings and consolidate information overcoming the cholinergic repression on hippocampal-neocortical dialog it is unlikely that waking SPW-Rs serve this function. More likely, is that systems consolidation depends primarily on interacting oscillations during NREMS (Ladenbauer et al., 2017; Latchoumane et al., 2017). In which case, a cellular mechanism connecting NREMS oscillatory interactions to memory consolidation is needed (Clemens et al., 2007; Wei et al., 2016).
A Cellular Mechanism for Systems Memory Consolidation
The synaptic change underlying memory storage requires the upstream activation of dendritic signaling cascades, initiated by high magnitude calcium transients, for its instatement (Ismailov et al., 2004; Yamanaka et al., 2017). Problematically, the properties of both SPWs and SOs suggest that these rhythms produce low magnitude calcium transients and synaptic depression or LTD (Tononi and Cirelli, 2006; Norimoto et al., 2018).
Therefore, a mechanism operating during NREMS to augment or prolong calcium influx and mobilize potentiating cascades is required. The temporal coupling of NREMS SPW-R, SO and SS rhythms can satisfy this requirement by spatially and/or temporally summating on dendrites to:
(I) increase depolarization magnitude and the number of conducting calcium channels; and/or (II) prolong depolarization duration and calcium channel open time.
In either case, coupling NREMS oscillations ensures that calcium levels surpass the threshold for engaging the coupling force of STDP and consolidating memories; a summary of this model is included as Figure 1.
The known properties of NREMS oscillatory interactions are compatible with those suggested as being necessary for modifying calcium dynamics and consolidating/interregionally re-locating memories. SPW-Rs often occur during the down-up state transitions of neocortical SOs (Battaglia et al., 2004). By reaching the cortex during the depolarizing portion of the SO signaling pathways initiated by the SPW-Rs can be augmented and prolonged.
Thus, by coupling with the depolarizing phase of SOs SPW-Rs may increase the number, or prolong the conductance, of membrane calcium channels, changing calcium’s kinetic profile in favor of higher dendritic concentrations and, in turn, consolidative plasticity. In support of this, Batterink et al. (2016) show that the presentation of cues associated with prior learning around the time of SO up state onset resulted in better post sleep memory performance than if cue presentation occurred around the time of SO down states.
SSs, when evoked by corticothalamic volleys triggered by SOs (Contreras and Steriade, 1995; Mölle et al., 2011), nest ripples in their troughs and in doing so transfer information about reactivated memories from the hippocampus to the cortex (Born and Wilhelm, 2012). Upon reaching the cortex, SSs are coupled to the depolarizing SO up state (Mölle et al., 2002).
Through these cross-frequency interactions SSs could augment/prolong calcium conductance and trigger the signaling cascades required for consolidative plasticity. Seibt et al. (2017) demonstrate that SS associated sigma activity elevates dendritic calcium, providing experimental evidence that systems memory consolidation associated oscillations can evoke changes in dendritic calcium kinetics/dynamics. In sum, interaction amongst dynamical brain rhythms orchestrates membrane properties to gate the synaptic plasticity necessary for consolidative mnemonic processing during sleep.