http://www.brain.mpg.de/fileadmin/user_upload/Documents/Download/Singer_Emeritus_Group/13GR3__20MOSERV50.PDF

http://www.brain.mpg.de/fileadmin/user_upload/Documents/Download/Singer_Emeritus_Group/13GR3__20MOSERV50.PDF
13
Coordination in
Brain Systems
Edvard I. Moser, Maurizio Corbetta,
Robert Desimone, Yves Frégnac, Pascal Fries,
Ann M. Graybiel, John-Dylan Haynes, Laurent Itti,
Lucia Melloni, Hannah Monyer, Wolf Singer,
Christoph von der Malsburg, and Matthew A. Wilson
Abstract
This chapter reviews the concept of dynamic coordination, its mechanistic implementation in brain circuits, and the extent to which dynamic coordination, and specic
manifestations of it, have the power to account for functions performed by interacting
brain systems. In our discussions, we addressed how on-the-y changes in coupling
between neural subpopulations might enable the brain to handle the fast-changing recombination of processing elements thought to underlie cognition. Such changes in
coupling should be apparent, rst and foremost, in the statistical relationship between
activity in interconnected brain systems, rather than in the individual ring patterns of
each subsystem. Dynamic coordination may manifest itself through a variety of mechanisms, of which oscillation-based synchronization is likely to play an important but
not exclusive role. Also discussed is how modulation of phase relationships of oscillations in different brain systems, in neocortex and hippocampus of the mammalian
brain, may change functional coupling, and how such changes may play a role in routing of signals at cross sections between cortical areas and hippocampal subdivisions.
Possible mechanisms for oscillation-based synchronization, particularly in the gamma
frequency range, are explored. It is acknowledged that the brain is likely capable of
producing zero-phase lag between spatially dispersed cell populations by way of rather
simple coupling mechanisms, primarily when neuronal groups are coupled symmetrically. Synchronization with remote areas may be most efcient with phase differences
that match the conduction delays. Fast-conducting, long-range projecting interneurons
are identied as a potential substrate for synchronizing one neural circuit with another.
A number of research strategies are identied to enhance our understanding of dynamic
coordination of brain systems and how it might contribute to the implementation of the
functions of those systems.
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
194
E. I. Moser et al.
Introduction
In its simplest form, coordination is dened by the multiple interrelations that
can be drawn between elements of any given assembly, and its phenomenological expression is signaled by the reconguration of elementary dynamics.
The potential relations are viewed as functions of an externally dened context
or an internally self-generated goal. In their introductory chapter, Phillips et al.
(this volume) further constrain the issue by adding: “In general, coordinating
interactions are those that produce coherent and relevant overall patterns of
activity, while preserving the essential individual identities and functions of
the activities coordinated.”
We tried to dene the process of dynamic coordination by contrasting it
with non-vacillating alterations, ever aware of the fact that virtually all neuronal processes are coordinated in one way or another. Much of the coordination of activity required for the formation of specic response properties, such
as the receptive elds of neurons in sensory systems or the generation of sequences necessary for the execution of movements, can be achieved by appropriate neuronal architectures that allow for the recombination and sequencing
of signals in processing cascades based on divergence, selective convergence,
and feedback. Such architectures allow for highly complex coordination and
association of signals, even if these originate in separate processing streams,
provided that there are adequate connections between the various stages of
these processing cascades. To dene relations and support selective grouping,
it would be sufcient to increase jointly the rate of the responses that are to
be associated with each other. Sparse coding and topographic coding would
further enhance the salience of rate changes and reduce the risk of grouping
unrelated but simultaneously enhanced responses. The fact that a number of
phenomena can be predicted by ring rate-based models raises the question
whether the ne-grained temporal structure of neural activity in different brain
regions has additional explanatory power. Thus, we discussed whether more
dynamic mechanisms are required to allow the exibility, robustness, and
speed at which cognition operates in the performing brain.
Dynamic coordination is required when the results of computations achieved
in different processing cascades need to be recombined and associated in a
exible, nonstereotyped way. A paradigmatic case is working memory, where
ever-changing items have to be temporarily associated with each other. It is
unlikely that xed neuronal architectures would be sufcient to anticipate and
cope with the virtually innite number of possible constellations of associable
contents. For each possible constellation, one would need a devoted set of neurons receiving the appropriately selected converging inputs. Because of this
limitation, the mechanisms required would need to be capable of establishing
transitorily, and in a highly exible way, relations among signals originating
in different processing cascades. These mechanisms would need to be able to
select in a dynamic, task- or goal-directed way signals from different, spatially
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
195
segregated processing streams and assure selective interactions between, and
further joint processing of, these signals. In our deliberations, we considered
the possibility that if several such signals are to be coordinated for the rst
time, that is, before anything is known about their relatedness, a possible
mechanism for such relatedness (presumably due to common cause) might be
in the form of temporal correlation. This transient coordination may be necessary to establish new linking paths, although its subsequent recruitment might
not always be required.
Furthermore, we recognized that most behaviors, and relations between different processing streams, are not generated ex novo but are highly predictive
based on the history of activation and learning of the system. We rarely reach
for a visual object by allowing an arm to cross the midline; we rarely raise a
leg and an arm simultaneously unless we are dancing. These relationships cannot be coded just in the anatomy, but must be implemented in the statistical
properties of the intrinsic (not task-driven) functional connectivity, which may
also require mechanisms of temporal correlation. Dynamic coordination thus
does not arise from noise, but from a patterned baseline landscape that may
constrain to some extent its exibility.
The question then arises: How can the presence of dynamic coordination
be diagnosed? We agreed that dynamic coordination should be apparent in the
patterns of interactions between dened neuronal populations. For all coordinated processes it should be the case that more information can be retrieved
by considering the joint activity of neurons belonging to different processing
cascades than evaluating the activities of the respective neurons in isolation.
In nondynamically coordinated processes, the relations among the respective
ring sequences will be stereotyped across trials, if stimulus conditions remain
constant. The additional information contained in the relations between the respective ring patterns can therefore be retrieved in sequential recordings from
these neurons and with averaging across trials. Such analysis has been applied,
for example, to the motor cortex and has led to the discovery of population
codes for movement trajectories (see Georgopoulos et al. 1986; Frégnac et al.,
this volume). However, this approach may be insensitive to more ne-grained
temporal relations between the discharges of dynamically grouped neurons.
These relations can only be determined by simultaneously recording from the
neurons whose activity is suspected to be coordinated. What one might observe in these cases is that the individual responses change only little, if at all,
in terms of average rate, while measures of relations between the responses
change in a systematic way.
A further constraint is that, in dynamic coordination, the information containing relations must change in a context-, task-, and goal-dependent way. For
example, temporal structures may depend on whether a trial was successful or
an error trial, and they may change as a function of stimulus context or shifts
in attention or goal denitions. To understand the functions of dynamic coordination on the basis of extracellular recordings, it is imperative to perform
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
196
E. I. Moser et al.
multisite recordings and to search for real-time correspondences between
temporal structure and task demands. Such approaches may require analytical
methods that extract activity patterns in many cells at the same time and take
their trial-by-trial covariation into account. Note, however, that it is possible
to infer coordination-related processes from single site recordings by appropriate multiscale analysis of intracellular membrane recordings, which reect the
impact of long-distance feedback and lateral connectivity (see Frégnac et al.
2009; also Frégnac et al., this volume).
Experimental evidence for such conditions of dynamic coordination is
available. In neuronal recordings during binocular rivalry, for example, whether the considered neuron is part of the processing cascade leading to conscious
experience cannot be predicted from the rates of individual neurons in V1.
In contrast, a measure of synchrony, in this case precise synchronization of
periodic activity in the gamma frequency range, predicts correctly whether
or not a given pair of neurons participates in the processing cascade which,
in this very moment, conveys the activity that reaches consciousness and is
perceived (Fries et al. 1997). In inferotemporal cortex, this information can be
retrieved from the rates of individual neurons, suggesting that synchronization
(dynamic coordination) at early processing stages was used to select responses
for further joint processing by enhancing their saliency, which then facilitated
transmission of this synchronized activity to higher levels where it then evoked
increased rate responses.
Examples for highly specic inter-regional coordination can also be found
in human neuroimaging. Functional neuroimaging is particularly suitable to
reveal large-scale dynamic coordination processes that span the entire brain.
Several methods are available that allow researchers to investigate how cognitive processes change the interactions between remote brain regions. These
include psychophysiological interactions (Friston et al. 1997), dynamic causal
modeling (Friston et al. 2003), and Granger causality mapping (Roebroeck et
al. 2005). Haynes et al. (2005), for example, have investigated the effects of attention on the connectivity between representations of attended locations. They
let subjects attend to two out of four spiral stimuli and report whether they had
the same or different handedness. Thereafter they measured the functional connectivity between the individual representations of these stimuli in retinotopic
visual cortex. They found that functional connectivity was increased between
the retinotopic representations of jointly attended stimuli, both within regions
(i.e., V1–V1, V2–V2) and between regions (V1–V2, V2–V1).
A topic of our discussion was whether the denition of dynamic coordination as structured temporal relationships between neuronal populations excludes simultaneous changes in the individual populations to be synchronized.
A strict application of the criteria implied for dynamic coordination by Phillips
et al. (this volume) would suggest that local processes, such as the discharge
elds of V1 neurons, would remain unchanged under associative stimulation
protocols. The underlying assumption is that the relational information should
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
197
remain separable from the information carried initially by each stimulus component taken in isolation. It was agreed that a strict application of the denition
may not be valid as there are numbers of examples which suggest that changes
in dynamic coordination give rise to, or are associated with, changes in individual subcircuits. Three examples illustrate the diversity of interactions between coordinated processes and activity in single cells and single populations.
The rst example was taken from the stomatogastric ganglion of the lobster. This particular sensorimotor network constitutes a striking example of
assembly dynamic reconguration correlated with changes in behavior where,
during the coordination, the electrical input–output properties of individual
elements are not preserved. Coordination is controlled by “orchestra leaders”
(PS cell), whose activity triggers the widespread broadcast of neuromodulators. This neuromodulation impacts on the intrinsic reactivity of the other cells
by changing reversibly the expressed repertoire of membrane conductances.
Consequently, the individual excitability patterns of any given cell will change
depending on the context (before, during, or after the orchestra leader cell has
red). Note that in this paucineuronal biological network where all partners
are known, the coordinator is identied and the causal link between temporal
assembly motifs and the behavioral actions, as well as their functional signicance (food swallowing, crunching, and expulsion), are clearly dened
(Meyrand et al. 1991).
In a second example, the coordinating agent was part of the high-order statistical features present in the sensory input stream. Changing the statistical
regularities of the environment may produce drastic reorganization of ensemble activity patterns and their stimulus-locked reliability. For instance, it is well
known that repeated presentation of drifting luminance gratings in V1 receptive elds evokes strong but highly unreliable responses, both at the spiking
and subthreshold levels. In contrast, in the same cells, virtual eye-movement
animation of natural scenes evokes temporally precise sparse spike responses
and stimulus-locked membrane potential dynamics which are highly reproducible from one trial to the next (Frégnac et al. 2005). In this second example,
coordination is unrelated to the behavioral outcome or neuromodulation since
it is observed in the anesthetized and paralyzed preparation as well as in the
attentive-behaving monkey (Vinje and Gallant 2000). This self-organized process adapts the temporal precision of the sensory code to the statistics of the
input. However, in contrast to the rst example, this adaptive form of temporal
coordination is done in the absence of internal executive or supervision units.
As demonstrated in the rst example, the full eld “whole” condition will affect the functional identity of the recorded unit (i.e., the individual receptive
elds of the V1 cells).
These two examples illustrate conditions where properties of the individual
units of a circuit clearly change in parallel with coordinating processes; however, the literature also contains illustrations where the information that can
be stored or recalled on the basis of coordinated activity is separable from the
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
198
E. I. Moser et al.
rate responses of single neurons. This can be seen in recordings of responses
to long-duration single frequency tones from the auditory cortex, where highly
transient burst responses are detected at the onset or offset of the tone whereas
the mean activity is unchanged during the tonic phase of the stimulation. The
presence of the stimulus is here signaled by a dramatic and tonic elevation in
the correlation between cortical units coding for the sound frequency, without
any apparent change of ring rate (deCharms and Merzenich 1996).
It remains a challenge to dene a taxonomy of coordination where the underlying mechanisms of each phenomenological form can be clearly separated.
Nonetheless, we agreed that dynamic coordination is apparent in a number of
studies that show changes in the temporal structure of the joint activity of two
or more neuronal populations that differ in character to those taking place at
the single-population level.
A Possible Need for Fast-changing Neuronal Architectures
All coordination requires the denition of relations. In the nervous system,
relations are established by the anatomical connections among neurons, the
anatomical architecture of the networks, and the patterns of inter-regional
spontaneous activity, the baseline or intrinsic functional architecture of the
networks. Work in nonhuman and human primates indicates that the anatomical connectivity matrix has small world properties allowing for the coexistence of local processing and long-range integration (Kotter 2004; Hagmann
et al. 2008). This small world architecture also gives rise to space-time structures of coupling and time delays, which in the presence of noise denes a
dynamic framework for the emergence of spontaneous and task-driven cortical
dynamics at different temporal scales (minutes, seconds, hundreds of milliseconds) and could support both long- and short-term changes in functional
connectivity. To allow for dynamic coordination in behavior (task-dependent
selection of responses for joint processing, selective association of subsystems to be engaged, etc.), the functional architecture must be modiable at
the same rapid pace as cognitive and executive processes can change. This
requires fast changes in effective coupling among neurons; that is, the gain or
the efciency of a connection must be modiable. The brain is likely to have
a number of mechanisms for achieving such changes in coupling, operating at
different timescales.
Coordination by Gain Modulation
Dynamic coupling can to some extent be accomplished by well-characterized
gain-modulation mechanisms. Synaptic gain changes can be induced within
tens of milliseconds, they may be (but do not have to be) associative and can
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
199
last from a few tens of milliseconds (e.g., during frequency-dependent changes
in transmitter release) to many decades (e.g., when activity is stored in longterm memory). Effective coupling can also be changed by purely activity-dependent gating, such as when dendritic segments are switched off by shunting inhibition, or when the sequence of activated synapses along a dendrite
is changed so that excitatory postsynaptic potentials (EPSPs) either summate
effectively or shunt each other, or when the nonlinear amplifying effect of
NMDA receptors is enabled or vetoed by local or global adjustments of membrane potential. The question raised in our discussions was whether such gainmodulating mechanisms would be sufcient to account for the speed and exibility of cognitive operations.
Coordination by Synchronization of Oscillation Patterns
A candidate mechanism for effective change of the coupling among neurons
involves rhythmic modulation of discharge activity (neuronal oscillations).
Oscillating networks facilitate the establishment of synchrony because they
can capitalize on the effects of entrainment and resonance. Oscillators that are
tuned to similar frequencies have the tendency to engage in synchronous oscillations if reciprocally coupled. This is the case even if coupling is very weak
and even if their frequency tuning is broad and the preferred frequencies are
not identical.
An oscillatory modulation of membrane potential, such as occurs in oscillating cell assemblies, connes spiking to the rising slope of the depolarizing phase. Thus, spikes emitted by networks engaged in synchronous oscillations become synchronized. The temporal precision of this synchronization
increases with oscillation frequency. In the case of gamma oscillations, output
spikes can be synchronized with a precision in the range of a few milliseconds.
Because of the coincidence sensitivity of neurons, this synchronization greatly
increases the impact that the output of synchronized cell assemblies has on
subsequent target neurons.
Another virtue of oscillations is that they allow the exploitation of phase
(relative timing) for coding (see discussions on phase precession in the hippocampus in Mehta et al., this volume). In oscillating, synchronized cell populations, responses to strong excitatory inputs will occur earlier on the rising
phase of the oscillation than responses to weak inputs. Thus, intensity can be
encoded in the time of spiking relative to the oscillation phase. This is a convenient way of coding since the latency of rst spikes already contains all
information about the amplitude of the driving input. Early studies on retinal
coding by Kufer (1953) showed that relative intensities of visual stimuli can
readily be assessed from the relative latencies of the rst spikes of ganglion
cells. Later studies showed that image reconstruction from rst spike latencies
is as good as counting rates over several hundred milliseconds (VanRullen and
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
200
E. I. Moser et al.
Thorpe 2001). Thus, readout time for this temporal code is much faster than
for the rate code. In the case of the retina, these intensity-dependent differences
of spike latencies are of course caused by receptor kinetics. In central processing, the same conversion of an amplitude code into a temporal code can be
achieved, in principle, by oscillatory modulation of cell assemblies.
These considerations provide answers to the question: Under which circumstances are oscillations needed? They are needed or at least highly advantageous if (a) spikes have to be synchronized with high precision to support
their propagation in sparsely connected networks (see synre chains of Abeles
1991); (b) spike timing has to be adjusted with high precision for the denition
of relations in learning processes such as spike timing-dependent plasticity
(STDP); or (c) phase is used as coding space, i.e., if timing relations between
spikes or between spikes and the phase of a population oscillation convey information about input amplitude or the relatedness of distributed processes.
There was consensus in our group that several pieces of experimental data
are consistent with a role for oscillation-based synchronization in cognitive
processes. For example, researchers have shown that attention during visual
search correlates with increases in coherence between local eld potentials
(LFPs) from the frontal cortex and the parietal cortex (Buschman and Miller
2007, 2009). Around the time when monkeys nd and shift attention to a visual target, there is an increase in coherence in two different frequency bands:
an upper frequency band (35–55 Hz) for bottom-up attention (pop-out), and a
lower frequency band (22–34 Hz) for top-down attention (conjunction search).
During search for conjunctions, the monkeys shift the location of their attention every 40 ms. The attention-related shifts in frontal eye eld spiking activity were correlated with increased power in the lower frequency band, suggesting that the oscillations act as a “clocking” signal that controls when attention
is shifted (Fries 2009). The study suggests that serial covert shifts of attention
are directed by the frontal eye eld and that synchronization between cortical
systems may regulate the timing of cognitive processing. Task-induced changes in synchronization or coherence have been reported at the level of individual
regions during sensory integration (Roelfsema et al. 1997), selective attention
(Fries, Reynolds et al. 2001), working memory (Pesaran et al. 2002; Howard
et al. 2003), and motor control (Crone et al. 1998). Between distant cortical
regions they have been reported during object recognition (Varela et al. 2001),
working memory (Jones and Wilson 2005), long-term memory encoding (Fell
et al. 2001), visual attention (Gregoriou et al. 2009), and sensorimotor integration (Roelfsema et al. 1997).
Oscillations and Dynamic Routing
Oscillations may inuence routes of communication within structurally constrained brain networks. Consider two groups (A and B) of neurons that provide
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
201
converging synaptic input to a common target group (C) and compete for inuence on this target group. If there is rhythmic synchronization among the neurons in group A and among the neurons in group B, but not between A and B,
then C will most likely synchronize to either A or B, but not to both at the same
time (Börgers and Kopell 2008). The locking of C to either A or B implements
a winner-takes-all between the competing inputs of A and B and establishes
an exclusive communication link between the target C and the more strongly
synchronized input (Fries 2005; Fries et al. 2008).
The described constellation of two neuronal groups converging onto one
target group is a fundamental motif in cortex. While this motif renders the
postsynaptic neurons selective to diagnostic features of the learned input pattern, it also renders them nonselective or invariant to nondiagnostic accidental
features. This invariance is an advantage, because it might provide the basis
for object recognition in the face of changes to irrelevant stimulus aspects;
however it is also a curse, because a given stimulus will never cover the complete input space of a given neuron, leaving room for competing stimuli. It
would be benecial if the effective input of a given neuron at a given moment in time were limited to functional subsets corresponding to one actual
object. This selective efcacy of subsets of a neuron’s input might be implemented through the above mentioned exclusive communication link, possibly
by synchronization in the gamma frequency band. For this solution to work,
two conditions have to be met simultaneously: First, inputs driven by a given
stimulus need to be rhythmically synchronized to each other, but not to inputs
driven by other stimuli. This corresponds to the binding-by-synchronization
hypothesis (Singer and Gray 1995; von der Malsburg 1981/1994). Second,
one of the input segments has to be given a competitive advantage over the
other through an enhancement. This corresponds to the hypothesis of biased
competition through enhanced synchronization (Fries 2005). Thus, the way to
use structural convergence in order to harvest both selectivity and invariance
seems to lie in the interplay between structural neuronal connectivity and dynamic neuronal synchronization.
In proposing a role for oscillatory activity in dynamic coordination of neuronal populations, our group agreed that one should not forget that oscillatory
activity, which may certainly be considered a signature or a manifestation of
dynamic coordination, does not necessarily explain the causes or mechanisms
by which such coordination arises. Consider a simple dynamic routing example like the one described above, where information in a low-level sensory
processing area could be routed toward one of two possible targets in a higher
area, with the choice of direction being endogenous, i.e., not dictated by the
stimulus. When information is routed one way (choice A), some neurons may
oscillate in one manner; when routed the other way (choice B), the same or
other neurons may oscillate in a different manner. The mere existence of these
oscillations does not explain how the selection was implemented in spatially
specic synchronization patterns. The result, signature, or manifestation of
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
202
E. I. Moser et al.
choosing between A and B, and communicating it to lower areas, may be what
is expressed in patterns of oscillations in the functionally connected areas, but
how the decision is made and what mechanisms and pathways are employed to
communicate it remain pressing questions.
Oscillations and Phase Relationships
Coupling between cell populations is heavily dependent on the phase relationship of the cells in the groups to be linked. By adjusting phase angles, coupling
can be modied over the whole range from ineffective to maximal. Controlled
changes in ring phases can also be used for dynamic routing if sender and
receiver are oscillating at a similar frequency and phases are adjusted. Because
oscillations can occur over a wide frequency range, many routes can be specied at the same time without interference. Finally, because coupling in oscillatory networks depends on phase and because we observe the coexistence of
oscillations in different frequency bands (theta, beta, gamma), many different
and graded adjustments of couplings can be structured, providing opportunities for establishing dynamically graded and nested relations, which could be
advantageous for the encoding of compositionality.
Consistent with a role for oscillations in routing of information, experimental data suggest that the phase of the ongoing oscillation can establish preferential windows for information processing. Inputs that arrive in the “good phase”
of the ongoing oscillation will be processed preferentially, whereas those arriving at the “bad phase” will be suppressed. For a long time it has been known
that the ability to perceive weak signals uctuates slowly over several seconds
(streaking effect). A recent study showed that infra-slow (0.01–0.1 Hz) uctuations of ongoing brain activity correlate with this behavioral dynamics. In this
study (Monto et al. 2008), the probability of detecting a tactile target at threshold was 55% more likely in the rising phase of the uctuation, and strongly
correlated with the power amplitude of higher frequency (1–40 Hz) EEG uctuations. Support for the same hypothesis comes from two more recent studies
showing a relationship between visual detection and phase in the theta–alpha
range (Busch et al. 2009; Mathewson et al. 2009).
Finally, we discussed the potential impact of precise phase relationships on
learning mechanisms. This seems important because processing architectures
have to be adjusted to the requirements of mechanisms establishing durable
relations (e.g., in associative learning); that is, they have to transform the (semantic) relations dened during processing into permanent changes in coupling that represent these relations. If any of the mechanisms of associative
synaptic plasticity known to date (LTP, LTD, STDP) have anything to do with
learning, it would seem that processing architectures need to be capable of dening relations in the temporal domain and that they will have to do so by adjusting the timing of individual spikes with a precision of a few milliseconds.
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
203
In STDP, for example, it matters whether an EPSP arrives just a few milliseconds before or after a spike to increase or decrease the efcacy of a connection. It would seem, therefore, that in signal processing and in dynamical
coordination, relations should be specied with a similar temporal resolution
and precision. Synchronization and phase adjustments in the gamma frequency
range could provide the time frame for the precise adjustment of spike timing
required for STDP. One should note, however, that the evidence for STDP in
vivo is relatively scarce and that it is easier to demonstrate the negative (depression) than the positive (potentiation) parts of the STDP curve in the adult
cortex (Jacob et al. 2007).
A Wider Repertoire of Coordination Mechanisms
Oscillations cannot be the sole mechanism of dynamic coordination. The neuroscience literature contains a number of examples of dynamic coordination
where brain circuits communicate using precise temporal codes not expressed
as lasting synchronization in oscillatory patterns in the LFP. The diversity of
mechanisms can be illustrated by patterns of hippocampal–neocortical interactions in slow-wave sleep. Slow-wave associated transitions in excitability from
low ring rate (putative down-state) to high ring rate (putative up-state) exhibit a systematic timing relationship in which the neocortex leads the hippocampus. During the elevated ring rate period that follows that transition, the
hippocampus expresses a series of sharp wave-ripple burst events that replay
sequential spatial memory information in which the timing relationship is reversed, with the hippocampus thought to be leading the neocortex. The dialog
that may be reected in shifting timing relationships may reect the dynamic
coordination of oscillatory modes during memory processing.
The possibility of a wider repertoire of mechanisms was further illustrated
by discussion of the mechanisms for rapid object recognition in the visual cortex. A given scene may be analyzed in terms of complex arrays of relations by
focal attention visiting here and there, and even if the relations thus identied
are indeed represented in terms of correlated oscillations, a more permanent
trace of these relations must be left behind to be available at later visits by focal
attention. If, for instance, there is a number of objects and a number of persons
present in a scene, then sequential focal attention may discover which object
belongs to which person, one by one, as a result of some inference; when coming back to one of the persons or objects, this result should be available immediately without the necessity of going through the process of inference from
scratch. In addition, there is the necessity of maintaining ongoing relations
or links between different neuronal ensembles over longer timescales. Many
patterns of behavior are predictable, although not necessarily across individuals, repeated over and over with little variation, while at the same time novel
behaviors can be stabilized with learning.
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
204
E. I. Moser et al.
What could the mechanism of such short- and longer-term storage of relations be? It was proposed that in addition to the elementary symbols represented by neurons (or groups of neurons), there might be a large network of
dynamic links. These links correspond to permanent neural connections, which
can, however, be modied (e.g., made ineffective) temporarily. In this view
the brain’s representations would not have the form of vectors of activity, or
neural signals, as in classical conceptualizations, but they would have the form
of dynamical graphs. There are various mechanisms by which the efciency of
connections can be rapidly modied. There is synaptic plasticity on a continuous range of timescales, starting from a few milliseconds, and the effective
connectivity (Aertsen et al. 1989) of a network can be changed by a variety of
presynaptic and postsynaptic inuences.
Recent work in human brain imaging shows that spontaneous activity, as
measured by uctuations of the blood oxygenation level dependent (BOLD)
signal, is not random but organized in specic spatiotemporal patterns (Deco
and Corbetta 2010; Fox and Raichle 2007) that resemble functional networks
recruited during active behavior. These correlations occur at a very slow temporal scale (<0.1 Hz), which correspond to uctuations of slow cortical potential (0.1–4 Hz) and band limited power uctuations of the gamma band (He et
al. 2007). These patterns of spatiotemporal correlation at rest reect not only
the underlying anatomy, but are gated by their recruitment during tasks. The
leading hypothesis, supported by studies showing changes with learning and
lesions, is that these patterns of spontaneous activity code for relations in the
cortex that are related to the history of network activation and learning. They
may represent attractor states that constrain and potentially bias the recruitment of brain networks during active behavior.
If these views are correct, then neuroscience is currently ignoring a large
part of the representational machinery of the brain—very large indeed, as there
are many more connections than there are neurons in the brain.
If coordination is expressed largely by dynamic connections, then what is
the importance of signal correlations? We agreed that signal correlations are
likely to be indispensable when a set of neurons are to be coordinated for the
rst time; that is, when the downstream circuits have not yet encoded this
relatedness in their link structure. In this way, neural oscillations could play
a vanguard role, appearing only early in some learning task, disappearing as
soon as the coordination pattern is encoded in some connectivity structure.
Computer modeling work will be particularly useful in shaping our thoughts
about neural operations if the model can be related in a convincing way to neural operations, instead of just using the brute force of high-speed computers,
and if the performance of the model can be proven superior in public benchmark tests. Such tests are available for face recognition (e.g., FRVT 2002;
Messer et al. 2004; Phillips et al. 2005). The consistently winning systems
were all correspondence-based; that is, they are based on representations of
faces in terms of two-dimensional arrays of local features (mostly of Gabor
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
205
type, i.e., modeled after receptive eld types found in primary visual cortex)
and on nding correspondences between local feature sets in model and image. As a large number of such correspondences are to be found for a given
match, temporal coding is bound to be rather time-consuming. A model of correspondence-nding by temporal coding (Wiskott and von der Malsburg 1996)
would, if implemented with realistic neurons, have needed more than 10 s to
recognize a face, two orders of magnitude slower than human performance. If,
however, connectivity patterns were installed that allowed for the fast dynamic
installation of topographic ber maps with the help of control units (introduced
by Olshausen et al. 1993), high-performance face recognition by correspondence-nding within physiological times of 100 ms was feasible (Wolfrum et
al. 2008). In a related study (Bergmann and von der Malsburg, submitted), it is
shown that the necessary control unit circuitry can be developed on the basis
of synaptic plasticity controlled by synchrony coding.
Correspondence-based object recognition models have explicit representation of shape. It was argued, however, that pure feedforward models (such as
Serre et al. 2007; Poggio and Edelman 1990), which do not make use of dynamic coordination, are also able to represent shape. The Chorus of Prototypes
algorithm (Duvdevani-Bar and Edelman 1999) and the Chorus of Fragments
model (Edelman and Intrator 2003) may represent shape if endowed with a
mechanism for relating together the responses of the ensemble of neurons
that represent, in a distributed yet low-dimensional manner, the current input.
Temporal binding by synchrony may be just such a mechanism (as was proposed in Hummel and Biederman 1992).
The Mechanisms for Synchronization between Neural Populations
Although there was consensus that the brain has a wide repertoire of mechanisms for achieving dynamic coordination, we chose to discuss in more detail
coordination by synchronization of oscillatory activity across neuronal populations. This form of coordination has support in the experimental literature, as
suggested above, and there is now a considerable literature exploring mechanisms of synchronization at the level of cellular assemblies.
We began the discussion by reviewing models for synchronization between cell populations. Several models have been proposed as mechanisms
for achieving zero-phase lag between same-frequency oscillatory activity in
different populations, which by denition might be seen as the ultimate expression of synchronization. Evidence indicates that zero-phase lag synchronization is ubiquitous and can occur over surprisingly large distances, such as
between the hemispheres (Engel, König, Kreiter et al. 1991), despite the rather
considerable conduction delays of pathways connecting the synchronized assemblies. At rst it may seem that such synchronization between widely dispersed populations could be achieved only by common input from a central
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
206
E. I. Moser et al.
oscillator that slaves the respective synchronized assemblies. However, since
callosotomy abolishes interhemispheric synchrony, it may instead rely on interactions between the synchronized assemblies. Several models for such interactions were considered.
One class of models relies on spike doublets of inhibitory interneurons
(Traub et al. 1996; Ermentrout and Kopell 1998). When neurons synchronize
locally in the gamma band, there is a characteristic interaction between excitatory and inhibitory neurons: excitatory neurons spike rst and trigger inhibitory neuron spiking with short delays. The ensuing inhibition shuts down the
local network until inhibition fades and the cycle starts again with the ring
of excitatory neurons. Long-range synchronization between two such gamma
oscillatory groups can occur when excitatory neurons of group A excite interneurons of group B, even if this entails a conduction delay of a few milliseconds between A and B. Essentially, the excitatory input from A to B triggers a second inhibitory spike in B and thereby prolongs the inhibition inside
B by the conduction delay. The two interneuron spikes in rapid succession
gave this model the name “spike doublet mechanism.” One prediction of this
model is that local gamma band synchronizations decrease in frequency when
coupled across long distances and the frequency decrease is proportional to the
conduction delay.
Synchronization across long distances might also be supported by other
congurations of reciprocal interaction between the subcircuits. Evidence is
now available which shows that zero-phase synchrony can be established despite conduction delays in the coupling connections both from experiments
with coupled lasers (Fischer et al. 2006) and modeling of networks with spiking neurons (Vicente et al. 2008). As long as at least three reciprocally coupled
systems are allowed to interact (triangular congurations), zero-phase synchrony is easily established and very robust against scatter in conduction times
of coupling connections.
A useful mathematical perspective on the phenomenon of zero-phase
synchronization comes from the study of coupled map lattices and globally
coupled maps. These are systems of coupled nonlinear dynamical systems,
whose long-term (ergodic) behavior can show some universal properties under some simplifying assumptions (Tsuda 2001). One of these assumptions is
that the system is globally and symmetrically coupled with a single coupling
strength. Under these constraints, it can be shown that the states of every coupled dynamical system come eventually to occupy a synchronization manifold.
Crucially, because of the symmetry constraints on the dynamic equations, the
set of all solutions must obey the same symmetry. Zero-phase synchronization
represents a symmetrical solution. Due to spontaneous symmetry breaking,
however, individual solutions might violate symmetry (i.e., exhibit nonzerophase synchronization). A simple example is a ball sitting on top of a hill in
a completely symmetric state. However, as this state is unstable, the slightest
perturbation will cause the ball to roll downhill. This movement will not occur
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
207
symmetrically in all directions but in each case in some particular direction.
Each single solution (i.e., rolling in a specic direction) breaks the symmetry
of the initial problem. Only the set of all solutions and the probability for the
ball to roll in a specic direction is symmetric. This has been used in a model
investigating oscillatory interactions in primary visual cortex (Schillen and
König 1991). Here a specic type of excitatory tangential connection avoids
the trivial solution of global synchronization. With the additional assumption
of ergodicity (i.e., the system that evolves over a long timescale and visits
all regions of state space), individual solutions have to obey symmetry constraints. For systems with more than two coupled oscillators, this reduced the
solution for the entire system to global synchrony with zero-lag quasi periodic
or chaotic oscillations. At low but nonnegligible coupling strength, the synchronization manifold is “riddled” with unstable points that “eject” the trajectory away from the synchronization manifold to produce intermittent bursts of
localized activity (Breakspear et al. 2009).
The main message from these theoretical treatments is that there is nothing
mysterious about zero-phase lag synchronization among three or more populations. Indeed, under the constraints of the model, it is impossible to get any
synchronization other than zero lag. To get consistent (nonzero) phase coupling, one has to break the symmetry, in terms of the intrinsic parameters of
the system or its coupling parameters. This basic phenomenon has been illustrated using neuronally plausible simulations by Chawla et al. (2001), where
it proved difcult to break the symmetry provided by three or more neuronal
systems that are interconnected in a roughly symmetrical fashion.
Although zero-phase synchronization could serve as a useful guide for understanding the mechanisms underlying long-range synchronization between
neural circuits, it was argued that the current models for producing zero-phase
lag have relied on unrealistic architectures and that the physiological properties of the model neurons do not match those of the performing brain (e.g.,
neurons do not regularly re in doublets during gamma oscillations). It was
proposed that zero-phase lags may not even be desirable for synchronization
when information is communicated over long distances. It may often be advantageous to introduce systematic phase shifts to coordinate convergence of
distributed information from different sources or to enforce timing relationships that would establish specic patterns of dynamic routing. The actual
phase lags between oscillating populations in two regions may vary across
task conditions and network states. One example for modulating the efciency
of interareal coupling by systematic phase shifts between oscillatory activity is
cortico-tectal communication (Brecht et al. 1998). It was also recognized that
regulation of spike timing through systematic phase locking can be used to
encode temporal relationships (such as spatial behavioral sequence encoding
through theta phase precession).
An example of time-shifted synchronization across brain areas was recently
reported in a study of frontal eye eld (FEF)–V4 interactions in an attention
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
208
E. I. Moser et al.
task (Gregoriou et al. 2009). There is considerable evidence that FEF plays an
important role in the top-down control of attention in visual cortex, including
V4. In the Gregoriou et al. study, spikes and LFPs were recorded simultaneously from FEF and V4, in monkeys trained in a covert attention task. One
stimulus always appeared inside the shared receptive eld and two others appeared outside; the monkey was cued to attend to a different stimulus on each
trial. Spike-eld coherence in the gamma band increased with attention in V4
and FEF. The effect was particularly strong when cells in the two areas had
overlapping receptive elds. However, there was almost a 180° phase lag in
synchrony in the gamma frequency band between FEF and V4, corresponding
to a time delay of about 10 ms. The same 10 ms time shift was found in other
frequency bands of the V4–FEF synchronous activity, suggesting that there is
a constant 10 ms time shift between the time while cells spike in one area and
cells are maximally depolarized in the other. It was suggested that this time
shift may be accounted for by conduction and synaptic delays between the two
areas. If so, then spikes from one area would actually arrive in the connected
area at a time when the receiving cells were most prepared (depolarized) to
receive them, which is consistent with the strong effects of FEF activity on the
top-down attentional modulation of V4 responses. The study illustrates the potential role of time-shifted synchrony between areas as a common mechanism
for functional interactions between cortical areas and raises the possibility that
zero-lag synchronization may be implemented primarily in local circuits.
To add to the complexity, a neuronal population may have different phase
lags to different subsets of a population with which it interacts. The recent
description of traveling theta frequency waves in the hippocampus (Lubenov
and Siapas 2009) suggests that neurons in regions that communicate with the
hippocampus may be synchronized with a subset of the hippocampal population across a wide range of the oscillatory cycle, but the identity of the neurons
with which synchronization occurs may change with phase. These phase lags
may inuence the wider patterns of coherence between the hippocampus and
other structures, such as the striatum, for which phase angle changes with task
and with learning (Tort et al. 2008; DeCoteau et al. 2007).
There was general consensus that the mechanisms enabling synchronous
ring across widespread brain regions are poorly understood, especially for
the higher frequency (e.g., gamma), and that alternative solutions should be
considered. One possibility considered involves long-range axonal collaterals. Synchronization between two oscillating populations might be achieved
if the collaterals of gamma-modulated pyramidal cells in one location phase
reset the basket neurons in the other location. A fundamental problem with this
model is the limited axonal conduction velocities of pyramidal cells. An alternative fast-conducting conduit between distant sites may instead be provided
by axon collaterals of so-called “long-range” interneurons. The anatomical
“short cuts” provided by the long-range interneurons may offer the interarea
fast transmission that is required to phase-synchronize gamma oscillations
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
209
between distant cortical regions. Such far-projecting fast-conducting interneurons have been described in the hippocampus (Sik et al. 1994). The axons of
this interneuron family innervate multiple regions of the hippocampus and can
project to multiple external regions, including medial septum, subiculum, presubiculum, entorhinal cortex, induseum griseum, and possibly other cortical
regions. Similarly, GABAergic interneurons in the medial septum project to
the hippocampus, preferentially to GABAergic interneurons (Freund and Antal
1988), hippocampal GABAergic neurons provide long-range projection back
to the septum (Gulyás et al. 2003), the basal forebrain has GABAergic neurons
that project widely across the cortex (Sarter and Bruno 2002), and long-range
GABAergic interneurons are known to connect remote areas in the ipsilateral
and contralateral cortex (Buhl and Singer 1989; Gonchar et al. 1995; Kimura
and Baughman 1997; Tomioka et al. 2005). A common property of many of
the long-range interneurons is that their axon caliber is nearly twice as large
as that of parallel conduits from pyramidal cells connecting the same regions
and the diameter of the surrounding myelin is three times thicker (Jinno et al.
2007). The estimated volume of the total axon arbor of a long-range interneuron is several times larger than the volume occupied by the axon tree of pyramidal cells, suggesting that only few such neurons may be needed to establish
coherence between regions. There was consensus in the group about the need
for further investigation of the potential role of long-range fast-transmitting
inhibitory interneurons in fast inter-area cortical synchronization.
We also discussed the potential role of ascending neuromodulatory systems
in synchronization of activity across brain regions. The broad terminal elds
of axonal projections from monoaminergic and cholinergic cell groups generally speak against a role in controlling dynamic changes in specic subsets of
interacting cell clusters, as does the slowness of many receptors for such transmitters (e.g., dopamine) and the long time that it takes to clear the transmitter
from the synaptic cleft. These facts do not, however, exclude a key permissive function for ascending neuromodulatory systems in providing necessary
conditions for inducing oscillatory activity. The discharge patterns of cholinergic as well as monoaminergic cell groups change radically during transitions between brain states (e.g., when subjects switch between awake states
and sleep), and such changes are temporally correlated with massive changes
in the oscillatory properties of cortical networks. Observations suggest that,
although terminal elds are broad, subtypes of intermingled interneurons are
innervated by different neuromodulatory systems (e.g., 5-HT axons terminate
on CCK-expressing interneurons but not parvalbumin-expressing cells, whereas cholinergic projections primarily terminate on basket cells). The specicity of the neuromodulatory innervation, as well as the specic combinations
of receptor subtypes expressed by different classes of interneurons, and the
ability of neuromodulators to change the time constants of GABA receptor
potentials are likely to have signicant impact on the generation of oscillations
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
210
E. I. Moser et al.
and synchrony across brain regions, although the exact mechanisms remain to
be determined.
In our discussions, we briey straddled the issue as to whether areas in the
neocortex only exchange information once they have nished their respective computations and then transmit the result (discontinuous communication)
or whether they permanently interact (continuous processing) until they converge to a collective result. We felt that the latter scenario is more realistic,
although some ERP studies seem to suggest that information is transmitted in
discrete packages.
Steps into the Future
Computational Models
How might neuroscientists improve their understanding of the brain’s mechanisms for dynamic coordination? Our discussion of models and experiments
will be presented sequentially, although the consensus is that advances require
an integrated approach.
Models will play a critical role in interpreting the many disparate empirical ndings regarding coordination in neural systems. Cortical models for dynamic coordination across brain systems can be roughly categorized according
to whether they are focused on the role of large population inuences on single
neuron properties versus models centered on the nature of interactions across
two or more specic cortical structures or layers.
There are numerous examples of models that examine the effects of attentional feedback or task demands on single neuron properties. The feedback in
these models comes from unspecied sources, and in most cases the models
only consider the effects of feedback on average ring rates. In the eld of
attention, for example, biased competition (Desimone and Duncan 1995; now
described as normalization models, Reynolds and Heeger 2009), feature-similarity gain (Maunsell and Treue 2006), and response gain models all attempt
to explain how attentional feedback cause the enhancement of responses to
attended targets and the suppression of responses to unattended distracters.
Normalization models explain and predict the large majority of attentional effects that have been reported on single neuron properties.
In contrast to these attentional models, based on average ring rates, some
models also address the role of spike timing and synchrony in neural populations. It is claimed that only spiking neuron models that incorporate gamma
synchrony can explain the effects of attention on competing stimuli within the
same receptive eld (Börgers et al. 2008), although direct tests of competing
models on these data are missing. In the future, it will be critical to make differential predictions from models based on static ring rates versus synchrony
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
211
and population dynamics, which can then be tested empirically in neurophysiological studies.
Fewer quantitative models take on the daunting task of modeling the interactions among two or more cortical areas. Efforts are ongoing to collect
data on a large number of individuals (upward of 2,000 healthy subjects) to
characterize the anatomical, functional, and electrophysiological neuromatrix
of the human brain (The Human Connectome Project). The goal of this project
is to provide the neuroscience community with a public data set, which will
hopefully describe for the rst time the entire array of cortical areas, as well
as their anatomical and functional links. This will allow quantitative mathematical modeling of their properties and exploration of the range of dynamics
and interactions that are possible within these networks both in healthy and
damaged brains. Presently, more limited systems models are being considered.
In attention, Hamker (2005) proposed a model that incorporates interactions
among a large number of visual areas and the “attentional control” system
that provides feedback. Quantitative models of object recognition typically
incorporate the receptive eld properties of neurons located along the ventral
stream. Examples of these types of models are ones developed by Poggio,
Edelman and colleagues (Serre et al. 2007; Edelman and Duvdevani-Bar 1997;
Duvdevani-Bar and Edelman 1999). These models are strictly feedforward,
based on ndings that inferotemporal neurons show object-selective responses
at times so short that they seem to preclude multiple recursive cycling up and
down the visual pathways. When trained on a large database of images, these
models are able to achieve recognition performance of human observers who
classify images based on very brief stimulus presentation times. For more
complex, cluttered scenes that require more recognition time, the latest version
of the Poggio model incorporates attentional feedback (Chikkerur et al. 2009).
By contrast, the face recognition model of von der Malsburg incorporates feedback to visual cortex from neurons holding stored representations of faces (see
above). This feedback model achieves a high level of performance on published databases of faces. However, it was argued that in all of these system
models, only average ring rates are considered and the timescale of the feedback is still relatively slow. A critical goal for the future is to nd out whether
the proven success of object classication and face recognition models are only
rst steps and that models based on binding mechanisms can be expanded into
a broad range of functional models for dynamically coordinated perception.
It was agreed that an essential element for evaluating models is their performance on large, publically available image databases. Although some databases
exist, there is a need for more realistic conditions in the databases, including
the recognition of objects at different scales and embedded in complex scenes.
Furthermore, beyond simple recognition, there is a need for models that can
answer at least basic questions about the objects, such as shape, size, and location. The development of such models will help in understanding how and why
synchronous interactions may be important for perception and memory.
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
212
E. I. Moser et al.
Experiments
We considered a number of experimental approaches to the testing of the
role of dynamic coordination in cognitive performance. Because oscillations
and synchronization are currently the best-explored mechanistic paradigms,
our discussion focused on possible ways to test whether such phenomena
are necessary and sufcient for the cognitive functions performed by those
brain regions where synchrony is observed. There was consensus that such
experiments must monitor activity from two or more cell populations at the
same time; as discussed in the introductory section, changes in the joint activity of two or more cell ensembles can be seen as a dening criterion for
dynamic coordination.
We agreed that much of the current evidence linking synchronization of
oscillatory patterns to coordination functions is correlation-based—a concern
that is shared with most other elds of study in systems neuroscience—and
that results thus, in principle, might be explained by other models, including
those based solely on rate changes. However, the literature does contain some
interventional studies which at least partly address the question of whether
synchronization between cell populations is necessary for behavioral functions
relying on the synchronized assemblies. In a study with multisite recordings
from the frog retina, for example, activity was recorded from cells that respond
to changes in shadows on the retina. Interventions that disrupted the synchrony
of ring across the recording electrodes disrupted escape behavior elicited by
shadow stimuli under conditions that did not change the average rates of the
cells (Ishikane et al. 2005). Other experiments, performed in the hippocampus
of the rat, have shown that using cannabinoids or other approaches to disrupt
temporal order in hippocampal place cells, in a manner that does not change
the average ring rates of the neurons, is sufcient to disrupt navigational performance in a spatial memory task (e.g., Robbe et al. 2006). In awake-behaving monkeys and healthy human subjects, some experiments have modied
activity in visual cortex during stimulus detection by stimulating putative attention control regions in frontal cortex (Ruff et al. 2006; Ekstrom et al. 2008).
Interference with frontal or parietal regions by TMS has been shown to alter,
in behaviorally signicant ways, anticipatory alpha rhythms in occipital visual
cortex (Capotosto et al. 2009). The invention of optogenetic tools for selective stimulation or silencing of genetically dened cell populations is likely
to result in a number of experiments along these lines within the next few
years. It is clear that synchrony can be interrupted experimentally, and those
data that exist so far suggest that such interventions may disrupt the functions
performed by the affected cell populations.
Although interventional approaches represent the gold standard for studies of causal relations between coordination and brain function, we agreed
that the caveats of such studies should not be forgotten. Interventions such as
stimulation or inhibition of target cell populations may have additional effects
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Coordination in Brain Systems
213
on top of the intended ones; for example, disrupting synchronization between
brain areas may also affect the proximal activity of each subpopulation, such
as ring rates or precise local phase relationships. We also spent some time
discussing strategies for gaining insight about coordination mechanisms under
circumstances where physiological variables cannot be manipulated directly.
One possible approach exploits the fact that human subjects often confuse the
color and shape of different objects. Such “illusory conjunctions” can be used
as a diagnostic tool to investigate which neural mechanism breaks down during
binding errors. To investigate the role of response synchronization in feature
binding, one could ask patients with intracranial electrodes to report on the
color and shape of multiple objects in the visual eld under conditions that
lead to occasional misbindings (e.g., when stimuli are presented very briey).
One would need to record from cells that encode two distinct properties of two
different objects in the visual eld (e.g., one could choose color and motion
as features and then record from color-sensitive cells in V4 and from motionsensitive cells in MT). If synchronization is indeed the neural mechanism for
feature binding, one would expect that the action potentials of cells belonging
to the same object are synchronized when perception is successful, and that
synchronization reects illusory conjunctions when they occur. The same recordings could be used as well to test a different model, where the positional
information encoded in V4 and MT signals maps corresponding features together. In this case, the positional information might be disrupted or shifted
in either of these populations, thus providing a potential alternative account
for the misassignment of features and spatial positions. If intrinsic dynamic
connectivity turns out to be an important mechanism to code relations, especially for behaviors that are predictable or well-learned, then new investigations should be directed toward manipulating the ongoing intrinsic connectivity, either through behavioral paradigms or interventions like stimulation
or disruption, and then correlate these changes to behavioral performance or
task-driven activity.
We concluded that a variety of experimental approaches and systems are
available to explore the function of oscillation-based synchronization and other
possible mechanisms of dynamic coordination between neuronal populations.
A common factor of all experiments that aim to test these functions should be
the recording of activity from two or more brain regions at the same time; this
is the only way to study changes in inter-regional temporal structure that may
or may not be accompanied by activity changes in each of the areas locally. A
number of brain systems, each with their unique advantages, should be used
to extract the mechanisms of coordination. The study of temporal structure in
large dispersed neuronal populations is likely to require an arsenal of new analytical and statistical techniques. Finally, there is a strong need for interaction
between computational models and experimental testing; models should make
clear predictions about activity changes in realistic neuronal architectures, and
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
214
E. I. Moser et al.
experimental strategies should be developed to test specic predictions from
the models.
Chapter from "Dynamic Coordination in the Brain: From Neurons to Mind," edited by C. von der Malsburg, W. A. Phillips,
and W. Singer. Strüngmann Forum Report, vol. 5. ISBN 978-0-262-01471-7. Cambridge, MA: The MIT Press.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Related manuals

Download PDF

advertisement