some mechanism is necessary to identify the members of a representation as belonging to-gether and to distinguish them from other representations that may be present at the same time. what is the mechanism of the binding problem, or how neuron communicate would distinguish it from all other neuronal activity present simultaneously in the cortical net-work?
The past findings have suggested that temporal correlation may relate the communica-tions and the information flow between the distributed areas [8]. The temporal correlation hypothesis proposes that neural assemblies consist of active neural units that are grouped together based on temporal correlation [9–11]. The temporal binding is that signals of neu-rons that are to be grouped together are correlated in times. It has been suggested that the binding problem may be solve by a mechanism which exploits the temporal aspects of neu-ronal activity [9]. The synchronization phenomena predict by temporal binding hypothesis have been documented for several years [6, 11]. In both cortical and subcortical centers, neuron can synchronize their discharge with a precision in the millisecond range [6, 11–13]
in sensory-motor system and perceptual and cognitive functions.
Synchronization includes oscillation, phase synchronization and general synchroniza-tion where general synchronizasynchroniza-tion is a broad synchronizasynchroniza-tion. Oscillasynchroniza-tion is a common approach toward modeling such temporal binding to group the neural assemblies. The dy-namical linking of different neural structures via oscillatory coupling was demonstrated first by animal experiments [13]. It has also been noticed that responses often exhibit an oscillatory patterning which is best revealed by recording jointly the activity of several adjacent cells [14].
Recording the activity of neurons, it require electrodes to be inserted through the skull into the brain. Such electrodes can record extracellular single-neuron activity, multi-unit activity and local field potential (LFP). Single- and multi-unit activity reflects the action potentials. LFPs represent the aggregate activity of a population of neurons located close to the electrode (spatial average across many neurons) as shown in Figure 1.4, consistent effects across a local population of neurons are enhanced [7,15,16]. For several years, many animal studies report the synchronization phenomena at single neural or LFP levels [6].
On the contrast to invasive modalities, there are several ways to study on brain, through noninvasive imaging of the electrophysiological, hemodynamic, metabolic, and
neuro-6 Introduction
LOCAL FIELD POTENTIAL (LFP). LFPs represent extracellularly recorded voltage fluctuations of a local neuronal population.
2 8 6|APRIL 2005 |VOLUME 6 www.nature.com/reviews/neuro
R E V I E W S
Mechanisms of neural communication
The specific cytoarchitectonic structure and connectivity of individual brain areas endow each region with the ability to perform specific tasks. This specialization is important for optimal information processing, but neces-sitates integration through long-range communication between neural assemblies.
The exact mechanisms of long-range interactions are poorly understood. However, recent findings point towards some general principles that have been ele-gantly shown in studies of the insect olfactory system.
Although it is not yet clear to what extent these results apply to mammals, this system, owing to its relatively small size and reduced complexity, is ideal for studying the mechanisms that underlie oscillatory communi-cation, and their functional consequences, in detail9,10. The interactions between the antennal lobe, the mush-room body and the inhibitory neurons in the lateral horn are of particular interest11–13. The antennal lobe receives input from olfactory receptor neurons; it then transforms and reformats this input for transfer through projection neurons to the mushroom body, which is responsible for memory encoding and retrieval, and to the lateral horn (FIG. 1). In turn, inhibitory interneurons in the lateral horn project to the mushroom body.
Odours elicit global oscillatory activity of 20–30 Hz in the antennal lobe network (which is composed of local and projection neurons), and this, in turn, is reflected in the local field potentials (LFPs) of the antennal lobe, mushroom body and lateral horn. The action potentials of projection neurons are PHASE-LOCKEDto the LFP oscil-lations in a neuron-, odour- and time-specific manner, such that the antennal lobe output is an evolving 20–30 Hz sequence of synchronized projection neuron spikes.
Interestingly, the preferred firing time, relative to the oscillating field potential (taken, for example, from the mushroom body), is different in all three networks.
Lateral horn inhibitory interneurons fire about half a period after the phase-locked projection neurons.
Consequently, inhibitory postsynaptic potentials (IPSPs) to mushroom body neurons that arise from input from the lateral horn occur half an oscillation cycle after the excitatory postsynaptic potentials (EPSPs) that arise from direct input from the antennal lobe. The strong IPSPs lead to collective inhibition of mushroom body neurons during half the oscillation cycle. Therefore, mushroom body neurons can only integrate their input briefly once per oscillation cycle. The functional conse-quence of the oscillatory interactions of excitatory drive (from the antennal lobe) and phase-shifted inhibition (from the lateral horn) in this particular case is a spatial and temporal sparsening of the odour representation11, which is a prerequisite for fine odour discrimination9,10,14. This example illustrates the following rules, which govern interneuronal communication on different spatial scales.
Filtering and resonance. The effect of the action potentials that are generated in the antennal lobe on the neurons in the mushroom body depends on their timing relative to the global LFP oscillations. The oscillations add relevance This article reviews recent findings and perspectives
on the emerging concept of long-range neural syn-chronization and desynsyn-chronization in motor control and cognition under physiological and pathological conditions. These findings indicate that long-range oscillatory synchronization implements coordinated communication between various areas in the brain. In the first section, we summarize the basic structural and functional neuronal properties that subserve neural communication, and describe possible mechanisms for long-range communication. The second section covers new findings about the functional roles of long-range communication in cognition and motor control. The consequences of disturbed long-range communication and its involvement in clinical symptoms are presented in the final section.
Box 1 | Neural oscillations
Neural oscillations refer to periodic variations in the recordings of neural activity (left panel). Activity measures are related to the membrane potential of single neurons or populations of neurons (right panel) and thereby encompass action potentials (APs) and
LOCAL FIELD POTENTIALS(LFPs). LFPs represent extracellularly-recorded voltage fluctuations in the membrane potentials of a local neuronal population. LFPs originate from EXCITATORY AND INHIBITORY POSTSYNAPTIC POTENTIALS(EPSP/IPSP), mainly as a result of action potential input. As LFPs represent a population measure (spatial average across many neurons), consistent effects across a local population of neurons are enhanced. Oscillations in LFPs often represent regularities in the input to local neurons. These oscillations might have different relations to neuronal firing. They might result from a population of neurons, each firing consistently at every cycle of the oscillation (AP1, see panel). They could also be caused by neurons for which the firing probability is modulated at the frequency of the LFP oscillations. These neurons could fire at a different frequency from that of LFP oscillations, or at no particular frequency (AP2, see panel). Nevertheless, the modulation of firing probability leads (as a population effect) to oscillatory LFPs (see panel). LFP oscillations might be due to regularities in the firing of only a subset of neurons in a local area, as LFPs represent a spatial average. The members of the subset might even change over time. This has important implications for identifying correlated neural firing.
Because, technically, only a few neurons can be recorded simultaneously, it might be difficult to identify neurons that participate in correlated activity.
The emergence of oscillations and the frequencies of these oscillations depend on cellular pacemaker mechanisms and neuronal network properties25,129. In general, higher frequency oscillations originate from a smaller neuronal population, whereas low frequency oscillations encompass larger populations. Furthermore, in a given network, the frequency of inhibition-based oscillations might depend on the strength of the driving input and on the magnitude and timecourse of the IPSP26.
AP1
AP2
LFP
Figure 1.4: Action potentials are AP1 and AP2, and local field potentials (LFPs) are LFP which represent the spatial average of a populations of neurons (right panel) [15].
chemical processes, instead of anatomizing. ‘Noninvasive’ means without physical harm.
There are four major non-invasive modalities, which are Magnetoencephalography (MEG), Electroencephalography (EEG), functional Magnetic Resonance Imaging (fMRI) and Near-Infrared Reflectance Spectroscopy (NIRS). These techniques record a spatial summation of LFPs [15].
The functional magnetic resonance imaging (fMRI), measures blood oxygen level de-pendent (BOLD) component of the hemodynamic response that is associated with local neural activity with spatial resolution as high as 1-3 mm. Several studies shows some evi-dence the correlation between LFP and BOLD, and this evievi-dence indicates that the firing ac-tivity of a neuronal population will, in general, be proportional to the BOLD response [16].
Therefore, functional magnetic resonance imaging (fMRI) measure the spatial summation of LFPs. Even though it is a very promising approach to investigate the neural activity and the cortico-cortical correlation. However, the temporal resolution is insufficient to observe the details of the communications while the temporal resolution is limited by the relatively slow hemodynamic response, approximately 1s [17].
Another popular non-invasive technique, the scalp MEG/EEG/event-related potential
1.2 Magnetoencephalography 7
(ERP), is also thought to reflect the summed electrical effects of excitatory synaptic neu-rotransmission in large populations of neurons. MEG and EEG are two complementary technique that measure the magnetic induction outside the head and the scalp electric po-tential those produced by the neuron activities inside the brain. Therefore, the character of higher temporal resolution compared to fMRI allows the studies of the dynamics of neuron network on the order of tens of millisecond. These techniques non-invasively record the neural activity at high temporal resolution; Therefore, they are a proper modality to analyze the neural communication at cortical level.
However, the analysis of neural communication at cortical level based on non-invaseve MEG/EEG need some steps. We introduce the background knowledge in later section1.2 and then introduce how to estimate the cortical neural activity.
1.2 Magnetoencephalography
1.2.1 Background
At the end of the 18th century, the electric brain signal was found. When we perform a task, the neurons at corresponding cortex activate. The entire excited neuron can be thought of as a battery, and the potential difference causes a current flow, therefore the information interchange between the neurons. As neurons become active, they induce changes in blood flow and oxygenation levels, which is imaged by fMRI. fMRI can monitor the hemody-namic changes with spatial resolution as high as 1-3 mm; however temporal resolution is limited by slow hemodynamic changes. Therefore, fMRI has poor temporal resolution compared with MEG and EEG.
MEG and EEG are two complementary technique that measure the magnetic induction outside the head and the scalp electric potential those produced by the neuron activities inside the brain. Therefore, the character of higher temporal resolution compared to fMRI allows the studies of the dynamics of neuron network on the order of tens of millisecond.
Nevertheless, low signal to noise ratio (SNR), and inherent ill-posed problem are two major difficult in the studies of brain functionality by using modality of MEG and EEG.
First, electrical brain signal is very tiny compared to the environmental noise. Typical EEG
8 Introduction
scalp voltages are on the order of tens of microvolts, and characteristic magnetic induction produced by neural currents is extraordinarily weak, on the order of several tens of fem-toTeslas. Therefore, MEG measure induced magnetic field via superconducting quantum interference devices (SQUIDs), a highly sensitive amplifier, inside a magnetically shielded room. So, compared to EEG, MEG has higher signal to noise ratio (SNR). Second, the recording of MEG and EEG are induced by sources distributed the whole head. Even with infinite MEG/EEG sensors, a non-ambiguous solution to source localization of the neu-ronal activities inside the brain would be possible, let alone the number of electrodes of MEG and EEG sensors usually ten or a few hundreds.
Given the recording of MEG/EEG, the inverse problem involves estimation of the prop-erties of the current sources within the brain that produced these signals. We can acquire the concept of inverse problem from the Bayesian statistic framework.
P (x|y) = P (x)P (y|x)
P (y) (1.1)
P (x|y) denotes the conditional probability of x given y, also called the posterior proba-bility because it is derived from or depends upon the specified value of y. P (y|x) is the conditional probability of y given x. P (y) is the prior or marginal probability of y, and acts as a normalizing constant. P (x) is the prior probability or marginal probability of x.
It is ‘prior’ in the sense that it does not take into account any information about y. Apply to the MEG inverse problem and let x represent the distribution of the sources inside the brain and y represent the recordings from the MEG/EEG sensors. P (x|y) can describe the inverse problem that to get the distribution of the sources while given MEG recordings.
From the Bayesian equation, we can simplify the inverse problem as the form of the right side of the equal sign if we want to know parameters of the source distribution from the recordings. Therefore, P (y|x) is the key to the solution. P (y|x) describes the probability of the recordings when given parameters of the source distribution and that is the forward problem [18, 19].
The inverse problem is to estimate the neuronal activities in the brain based on MEG/EEG recordings [17]. As mention above, involving estimation of the properties of the signal in-duced by the current source inside the brain help to solve inverse problem. Before we can make such an estimate, we must first understand and solve the forward problem, in
1.2 Magnetoencephalography 9
which we compute the scalp potentials and external fields for a specific set of neural cur-rent sources.Therefore, the inverse problem can be transformed to the form Eq 1.1. But the inverse problem is a inherent ill-pose problem which has infinite solutions. Lots of algorithms with difference constraint had been proposed to solve the inverse problem.
1.2.2 Related work of the source activities at cortical level
Approximations such as the equivalent current dipole (ECD) model [17, 18, 20, 21], assumptions such as a fixed number of dipoles within an epoch is obtained by the least squares source estimation which is one of the most widely-used method. It finds out he best solution by nonlinear search to minimizing error between the induced electromagnetic field by ECDs and the MEG/EEG recordings. However, It is difficult to decide the prior number of the sources and it may trap in local minimum. Multiple Signal Classification (MUSIC) [22–24] is another kind of method which can avoid trapping in local minimum through by scanning the region of interest and determining the locations with peak projections of forward models in the signal subspace. Minimum Norm Estimation (MNE) [25–27] will estimate the brain activities on the cortical surface, so it set dipole orientation either to be on the tangential plane or normal to the local cortical surface. But the major problem is that because of the minimum norm constraint, the result will tend to emphasize the cortical regions closer to the MEG sensors [28].
Recently, beamformer is one of the most promising solutions to the inverse problem.
[29]. Beamformer performs a spatial filter on recordings of MEG/EEG to filter out the signal at the targeted location, acting as a virtual sensor to measure the signal with a specific orientation. Beamformer can obtain the activities of the targeted location and suppress the influence contributed from other sources by imposing the unit-gain constraint and minimum variance criterion. Given a unit dipole with specified position and orientation, we can calculate a spatial filter from the data covariance matrix and the lead field of the dipole.
The neuronal activity of the dipole at the specified position can be obtained by applying this spatial filer on the recordings of MEG/EEG. By repeating the procedure for each position inside the brain, we can obtain the neuronal activities of the whole brain.
Two kinds of beamformer, vector-type beamformer [30] and scalar-type beamformer,
10 Introduction
have be studied. [31, 32]. The vector-type beamformer decomposes the dipole orientation into three orthogonal components, each one with a fixed orientation. Every component has its own spatial filter calculated individually. Linearly constrained minimum variance (LCMV) [30] is one of the vector beamformer and it sums the results probed on three direc-tions. Only one spatial filter is used for each specific position in scalar-type beamformer.
Scalar-type beamformer determines the direction by maximizing the pseudo Z value. Com-pared to vector-type beamformer, the major advantage of scalar-type beamformer is that the activity distribution is more focal and higher signal-to-noise ratio [32,33]. But using vector-type beamformer is more efficient to calculate the spatial filter because all the procedures involved are deterministic.
In the scalar-type beamformer, only when the dipole orientation is accurate, the ef-fective spatial filter can be calculated. Therefore, it is essential to accurately determine the dipole orientation [33,34]. One way to determine the dipole orientation is to use the normal of cortical surface [34]. But the surface reconstructed very difficultly and the reconstruc-tion deviareconstruc-tion will decrease the accuracy of dipole orientareconstruc-tion. Only when the estimareconstruc-tion error is smaller than ten degrees, the spatial filter determined by the cortical surface normal is feasible [34]. Another way to determine the dipole orientation is to maximize the pseudo Z in the synthetic aperture magnetometery (SAM) method [31] by exhaustively evaluating all the possible candidates. Nonlinear optimization method is more efficient but only can guarantee the suboptimal solution.
Recently, a novel spatial filtering technique, called the maximum contrast beamformer (MCB), was proposed by Chen et al [35]. This MCB method has the advantages of good output SNR and focal activity distribution as in scalar beamformers, while the dipole ori-entation is determined accurately and efficiently in a close-form solution. The method exploits a maximum-contrast criterion that maximizes the ratio of the reconstructed neu-ronal activities in the active state to those in the control state and helps to analytically and accurately determine the dipole orientation in a closed-form manner.
The below is the result of MCB algorithm. There are three simulated sources and their strength and frequency are shown as Figure 1.5 (b), three different sinusoidal signals. Their locations in the cortical area are shown as Figure 1.5 (a). The electromagnetic mapping of brain activities calculated by MCB is shown as Figure 1.5 (c). It can be demonstrated that