A prominent property of neural networks is their tendency to engage in oscillatory activity. Hans Berger (1873–1941) was one of the first scientists to observe the brain's rhythms which he recorded in the form of electrical activity on the scalp of healthy, awake participants.1
He also introduced the current nomenclature naming different rhythms by Greek letters. Rhythmic electroencephalographic (EEG) activity is commonly subdivided in 5 major frequency bands, the delta- (0–3 Hz), theta- (4–7 Hz), alpha- (8–12 Hz), beta- (13–30 Hz), and gamma-band (30–200 Hz). (We did not include oscillations in the delta-frequency range [0–3 Hz] in the review, as evidence on the functional role of delta oscillations is so far limited.)
Since the early discoveries of Berger and others, EEG has become a valuable tool for the study of brain functions as well as for clinical investigations of various neurological and psychiatric disorders. More recently, a method for recording the brain's magnetic activity noninvasively—magnetoencephalogram (MEG)—has been developed that has significantly improved the spatial resolution of extracranial recordings and, thereby, also the detectability of low-amplitude, high-frequency oscillations.
In the last 2 decades, the investigation of brain activity in EEG and MEG data has experienced an important paradigm shift because the focus has moved from the analysis of event-related averaging of neuronal responses (event-related potentials [ERPs]) to methods that investigate the power, the coherence, and the phase locking of nonaveraged oscillating signals. The reasons for this paradigm shift are in part technical; the access to computational power has increased, and the analysis techniques have proliferated.2
However, the most important reason for this shift is conceptual: the discovery of synchronized, oscillatory in neuronal spiking activity led to novel hypotheses about the putative functions of the phenomenon,3–5
and the testing of these hypothesis required the analysis of nonstimulus-locked internally generated temporal patterns.
Oscillations that reflect self-generated activity are referred to as induced oscillations and preclude evaluation of averaged responses and, therefore, require single-trial analyses because their latency varies from trial to trial. Typically, the latency in which induced oscillations occur is within the time window of 150–400 ms. In contrast, evoked oscillations are strictly phase locked to the onset of the stimulus and are measured by stimulus-triggered averages of responses. Evoked oscillations typically occur within a latency window of 50–150 ms and have been related to early, stimulus-driven encoding processes.
An important link between oscillations and cortical computations was the discovery of the role of oscillatory rhythms in the beta/gamma range (20–80 Hz) in establishing precise synchronization of distributed neural responses. Gray and colleagues6,7
showed that action potentials generated by cortical cells align with the oscillatory rhythm in the beta and gamma range, which has the consequence that neurons participating in the same oscillatory rhythm synchronize their discharges with very high precision. Thus, it is a central role of cortical oscillations in the beta/gamma range to enable neuronal synchronization.4,5
Self-generated oscillations and synchronization are highly dynamic phenomena and depend on numerous conditions, such as central states,4,5,8
The occurring strength of synchrony is closely correlated with perceptual processes such as feature binding, subsystem integration, brightness perception, and interocular rivalry.4
In addition, the strength of synchronization predicts whether an animal will give a correct response in an upcoming trial of a perceptual decision task,10
suggesting its important functional role.
Synchronization patterns between pairs of neurons have a particular topology that exhibits the properties of small-world networks. Networks with “small-world properties” are characterized by a combination of local clustering of activity and a short path length as an index of global integration. The hubs of these networks—ie, neurons synchronized most strongly with the rest of the network—exhibit most prominently their cortical functions, ie, orientation selectivity in the early visual areas, suggesting a close relation between neuronal oscillations and synchrony on the one hand and the organization of network interactions on the other.11
Possibly the most important function of synchronized, oscillatory activity is the implementation of a mechanism that can exploit the relative phase of oscillations. Recent data show a systematic relation between the phase offset of synchronized spiking and stimulus properties,12–14
suggesting that information is encoded in the relative firing times of the discharges of distributed neurons.
As the empirical data indicate, cells that are excited more strongly tend to fire earlier than those excited less strongly.14
This offers a particularly efficient mechanism for encoding stimulus-related information, which can be either redundant or complementary to that already provided by the rate responses. Redundant information would provide more precision to the encoding process, and complementarity would provide information additional to that coded by firing rates. Either way, an important advantage of such a temporal coding scheme is the processing speed with which participating neuronal structures may conduct computations.
The information about a neuron's firing rate can be assessed accurately only after observing this neuron's activity for some time (eg, 100 ms or more). In contrast, the information about the phase of an action potential is present instantaneously (ie, within a single cycle of a beta/gamma oscillation) and, hence, can be, at least in principle, also extracted within such a short period of time. Such accurate and readily available information can enhance numerous cognitive processes, ranging from object recognition to decision making.
The mechanisms underlying the generation of these spatiotemporal patterns associated with beta/gamma oscillations are likely to be similar to those responsible for the phase precession of pyramidal cell firing in the hippocampus relative to theta oscillations.15,16
Cortical neurons not only generate such precise temporal information but also appear highly capable of reading out the same information when receiving inputs from earlier processing stages. The possible readout mechanisms have been discussed most commonly in the context of the varying spike latencies of the sensory inputs.16–18
In addition to the high-frequency oscillations in the beta- and gamma-band, oscillatory rhythms in the theta- and alpha-band also play an important role in cortical computations. Alpha activity (8–12 Hz) has been not only associated with an inhibitory function19
but also with the long-distance coordination of gamma oscillations,20
and theta activity has been proposed to support large-scale integration of subsystems serving the formation and recall of memories.21–23
In general, there is a correlation between the distance over which synchronization is observed and the frequency of the synchronized oscillations. Short-distance synchronization tends to occur at higher frequencies (gamma-band) than long-distance synchronization, which often manifests itself not only in the beta- but also in the theta- (4–8 Hz) and alpha- (8–12 Hz) frequency range.23,24
Thus, oscillations and their synchronization are important correlates of neuronal processing and provide valuable measures for the assessment of normal and pathological functions. (It is important to note that the remaining part of this review focuses on studies based solely on field recordings [EEG, MEG] rather than the activity of individual neurons. Consequently, it can never be directly resolved whether an increase in the power of these field recordings is due to the increase in the strength of synchrony or in the strength of oscillations.).
Theta oscillations occurring in the frequency range of approximately 4–8 Hz represent one of the best-studied rhythms in the mammalian brain (for a review, see Buzsaki,21
Kahana et al,25
and Lengyel et al26
). In mammals, theta oscillations are particularly prominent in the hippocampus but occur also in extrahippocampal regions, such as the ento- and the perirhinal cortex, the prefrontal, somatosensory, and visual cortex, and superior colliculus.27,28
In the hippocampus, theta oscillations are generated by an interplay of glutamatergic and gamma-aminobutyric acidergic (GABAergic) neurons.29,30
In addition, GABAergic inputs are modulated by cholinergic inputs from the septum that possibly acts as a pacemaker for theta activity.31
Studies in rodents revealed a close relationship between the occurrence of theta oscillations in the hippocampus, locomotion, and the place-specific firing of hippocampal pyramidal cells (place cells).15
These place cells have spatially selective fields and discharge when the animal is at the location corresponding to the receptive place fields.32
In a seminal study, O'Keefe and Recce15
reported that the phase of the place cell firing relative to the ongoing theta oscillations advances gradually as the rat passes through the cell's place field. This phenomenon, called phase precession, has been interpreted as a mechanism to increase the precision of spatial coding and to bind assemblies of place cells for the representation of movement trajectories.21,16
Spatial navigation, a process requiring evaluation and temporary storage of relations, is only one of the numerous functions of the hippocampus. A large body of evidence indicates that the special ability of the hippocampus to process relations is used in a much wider context and plays a crucial role in the formation and recall of episodic and declarative memory.33,34
Accordingly, theta oscillations have also been assigned functions in memory-related processes. (1) Manipulations that eliminate theta activity in the hippocampus produce deficits in spatial and nonspatial working memory (WM) tasks in rodents (for a review, see Gevins et al35
); (2) theta oscillations correlate with WM load35,36
and link neural assemblies in frontal and parietal cortices during the maintenance of items in WM22
; (3) significant increases in theta power during encoding of information predict subsequent recall of information37
; and (4) in human EEG data, a theta component that is largest over midline frontal regions (frontal midline theta) is enhanced in tasks involving WM and requiring focused attention.35,36
An involvement of theta activity in memory processes is also supported by the evidence that rhythmic stimulation in this frequency range (theta burst stimulation) is particularly effective in inducing long-term potentiation (LTP) and that LTP induction is highly sensitive to the phase between stimulation relative to the ongoing theta rhythm. Stimulation at the depolarizing peak of the theta cycle favors LTP while stimulation in the through causes depotentiation. Thus, both in vitro and in vivo data suggest that theta oscillations act as a windowing mechanism determining the threshold and polarity of synaptic modifications.38–40
Alpha oscillations, the frequency of which is centered around 10 Hz, were the first rhythm to be discovered by Berger in 1924.1
Alpha activity is very prominent in the thalamus and can be sustained by isolated thalamic networks,41
which led to the assumption that cortical alpha is driven by thalamic peacemakers (see Basar et al42
for a review). However, da Silva et al43
demonstrated that cortical alpha results from synergistic interactions within thalamo-cortical-thalamic reentrant networks (see Sauseng et al44
for a review). In addition, alpha oscillations have also been recorded in subcortical areas, such as the hippocampus and the reticular formation.42
It is well established that alpha rhythms result from reciprocal interactions between excitatory and inhibitory neurons whereby the synchronization is in addition stabilized by gap junctions among inhibitory interneurons.45,46
The susceptibility of these networks to engage in alpha rhythms is in turn modulated by cholinergic and serotonergic mechanisms and by glutamatergic afferents acting via metabotropic receptors.47
Alpha activity is most prominent over occipital cortex when the eyes are closed and subjects are in a relaxed state and has therefore been regarded as reflecting cortical idling.48
Typically, opening the eyes results in an alpha blockade which has been linked to active stimulus processing.44,49
It is currently under debate, whether alpha oscillatory activity is related to functional inhibition of task-irrelevant processing44
or whether it is a direct and essential constituent of the active network (for a review, see Palva and Palva20
). Results from studies investigating visuospatial attention50–52
especially when capacity limits are reached,44
have been interpreted in favor of the inhibition hypothesis.
In contrast, evidence for an involvement of alpha activity in information processing has been provided by studies on mental imagery,55
conscious somatosensory perception,56
and WM tasks.20
Furthermore, there is an evidence for a relationship between long-range coherence in the alpha-frequency band and perceptual and cross-modal binding.54,57,58
Several lines of research indicate a relationship between baseline alpha activity and the expression of ERPs. Low prestimulus alpha amplitudes are associated with larger P1 responses in a visual discrimination task and enhance good performance. By contrast, high prestimulus alpha is associated with enhanced P3 activity and enhanced retention in WM tasks.44
The underlying assumption is that the sensory stimulus induces phase resetting of ongoing rhythmic activity in each trial and that averaging these phase-coherent rhythms produces the ERP.59,60
Another possibility is that the ERP reflects a transient response to the stimulus that is superimposed on the background EEG. The current literature suggests that these 2 mechanisms coexist (for a review, see Sauseng et al44
Nevertheless, it is still a technical challenge to disentangle the ERP response from ongoing oscillations.61
The available methods do not allow to distinguish conclusively between phase resetting and transient activity. Phase-coherent activity specifically in the lower frequency bands has been taken as evidence for the former model. However, the contrary may be true as well because transient bursts of neural activity may likewise result in similar phase-coherent activity. Thus, many of their physical characteristics may be overlapping and should not be used as criteria for the presence of either kind of activity (for a discussion, see Yeung et al61
Frequencies between 12 and 30 Hz are termed beta activity. Beta oscillations occur in all cortical areas and numerous subcortical structures including nonspecific thalamic nuclei, the hippocampus, the basal ganglia, and olfactory bulb.
Generation of beta-band oscillations has been linked to neurotransmitter systems including metabotropic glutamate as well as N
-aspartate (NMDA) receptors and GABAa receptor activity.62,63
Of particular clinical relevance is the dopaminergic modulation of beta oscillations in the basal ganglia, the subthalamic nucleus, and the motor cortex.64
In Parkinson disease, reduced dopaminergic control leads to the instability to desynchronize these beta oscillations that prevail in the rest condition and need to be replaced by gamma oscillations in order to enable the initiation of voluntary movements.65
This central beta oscillation has been related to the rolandic mu rhythm that is generated by the sensorimotor cortex and is most prominent at rest and attenuated or abolished by moving or observing biological movements.66
There is evidence for enhanced synchronized beta-band activity prior to movements in sensorimotor cortex influencing descending motor commands to contralateral hand muscles.67–69
Such corticomuscular coherence is the strongest during steady hold periods after movement but is abolished during the actual movement.68
In addition, beta activity has been implicated in a variety of cognitive tasks, such as learning in the mammalian olfactory bulb,70
novelty detection in the auditory system,71
and reward evaluation.74
It has been suggested that the common denominator of beta oscillations is to highlight a stimulus as novel or salient that warrants further attention.72
Furthermore, beta-band activity is involved in large-scale coordination of distributed neural activity. Kopell and colleagues29
showed that beta oscillations support more effectively coherence over large distances than gamma oscillations because their synchronization is less susceptible to long conduction delays. Consistent with this hypothesis, Tallon-Baudry and colleagues75
observed increased oscillatory activity in the beta-frequency range between 2 sites in inferior temporal cortex during the maintenance period in a WM task. Further evidence for a relationship between long-range synchronization and beta-band activity was reported by Schnitzler and Gross.76
The authors investigated the neural networks underlying attention control in visual processing in the attentional blink task. The results revealed that communication within the fronto-parieto-temporal attentional network proceeds via transient long-range phase synchronization in the beta-band.
Rhythms >30 Hz are addressed as gamma oscillations and occur in virtually all brain structures, including the olfactory bulb and the retina. These oscillations cover a broad frequency band ranging up to 200 Hz. Gamma-band oscillations were first recorded by Adrian and Matthews48
from the olfactory bulb following odor stimulation and have been later observed in the visual cortex following visual stimulation.
Generation of gamma-band activity is critically dependent upon several neurotransmitter systems. The networks of chemically and electrically coupled GABAergic neurons play a pivotal role in the primary generation of high-frequency oscillations and local synchronization77–81
while more far-reaching glutamatergic connections appear to control their strength, duration, and long-range synchronization.79
Cholinergic modulation via muscarinic receptors plays a crucial role both in the fast, state-dependent facilitation of gamma oscillations and associated response synchronization as well as in the control of use-dependent long-term modification of cortical dynamics that favor synchronization of responses in the gamma-frequency range.82,83
First indications for a functional role of gamma-band oscillations for information processing in cortical networks have been obtained in studies investigating the relations between stimulus-induced synchronization of gamma oscillations and feature binding in cat primary visual cortex (V1) by Gray and Singer7
(see ). This relationship was then confirmed in a series of studies with EEG and MEG that found robust support for the role of gamma oscillations in the grouping of stimulus elements into coherent object representations.84,85
Fig. 1. Synchronous Gamma Oscillations and Perceptual Binding. Responses are recorded for a pair of multiunit (MUA) activity with nonoverlapping receptive fields in primary visual cortex of an anesthetized cat. In the 2 different visual stimulation conditions, (more ...)
Further research has found a close relationship between attention, oscillations, and synchrony in the gamma-band range. Fries and colleagues86
recorded neurons in cortical area V4 while macaque monkeys attended to behaviorally relevant stimuli and ignored distracters. Neurons activated by the attended stimulus showed increased gamma synchronization (35–90 Hz) compared with neurons at nearby V4 sites activated by distracters. Consistent evidence for the relationship between attention and gamma-band activity both during visual and auditory perception has been also found in human EEG and MEG recordings.87–89
Taken together, these findings suggest that gamma-band oscillations have a general computational role in dynamically selecting neurons that communicate information about sensory inputs effectively.
Gamma oscillations have been also implicated in higher cognitive functions, such as memory. Tallon-Baudry and colleagues90
examined EEG-activity during a visual short-term memory task. Induced gamma-band activity was observed during the delay over frontal and parietal electrodes, indicating that gamma oscillations are involved in the maintenance of information in WM. Additional evidence for the role of gamma oscillations during WM has been obtained from intracranial recordings91
as well as from studies that examined the relationship between auditory WM and gamma oscillations in MEG data.92
Finally, gamma oscillations may also be involved in long-term memory because there is a relationship between the amount of gamma-band activity during the encoding of information and the subsequent recall.37,93,94
Recent data suggest that synchronized gamma-band activity may also be related to consciousness. A large body of evidence suggest that consciousness has to be understood as a function of numerous interacting systems that require a mechanism that transiently synchronizes a number of widely distributed neural assemblies.95
Melloni et al96
examined EEG data in response to the processing of visible and invisible words in a delayed matching to sample task. Both perceived and nonperceived words caused a similar increase of local (gamma) oscillations in the EEG, but only perceived words induced a transient long-distance synchronization of gamma oscillations across widely separated regions of the brain.
In addition to the role of gamma-band oscillations in perceptual organization, attention, memory, and consciousness, gamma-band activity has also been related to other cognitive phenomena, including language processing and motor coordination (for comprehensive reviews, see Fries et al4
As has been discussed for theta oscillations, gamma oscillations are also involved in the gating of synaptic plasticity.82
The mechanisms are similar, but because of the high frequency of gamma oscillations, the temporal resolution of the gating process is much higher and capable of distinguishing temporal offsets between pre- and postsynaptic activity in the range of milliseconds.