Dynamical principles in neuroscience * Rabinovich

Dynamical principles in neuroscience * Rabinovich
REVIEWS OF MODERN PHYSICS, VOLUME 78, OCTOBER–DECEMBER 2006
Dynamical principles in neuroscience
Mikhail I. Rabinovich*
Institute for Nonlinear Science, University of California, San Diego,
9500 Gilman Drive 0402, La Jolla, California 92093-0402, USA
Pablo Varona
GNB, Departamento de Ingeniería Informática, Universidad Autónoma de Madrid,
28049 Madrid, Spain and Institute for Nonlinear Science, University of California,
San Diego, 9500 Gilman Drive 0402, La Jolla, California 92093-0402, USA
Allen I. Selverston
Institute for Nonlinear Science, University of California, San Diego,
9500 Gilman Drive 0402, La Jolla, California 92093-0402, USA
Henry D. I. Abarbanel
Department of Physics and Marine Physical Laboratory (Scripps Institution of
Oceanography) and Institute for Nonlinear Science, University of California,
San Diego, 9500 Gilman Drive 0402, La Jolla, California 92093-0402, USA
共Published 14 November 2006兲
Dynamical modeling of neural systems and brain functions has a history of success over the last half
century. This includes, for example, the explanation and prediction of some features of neural
rhythmic behaviors. Many interesting dynamical models of learning and memory based on
physiological experiments have been suggested over the last two decades. Dynamical models even of
consciousness now exist. Usually these models and results are based on traditional approaches and
paradigms of nonlinear dynamics including dynamical chaos. Neural systems are, however, an unusual
subject for nonlinear dynamics for several reasons: 共i兲 Even the simplest neural network, with only a
few neurons and synaptic connections, has an enormous number of variables and control parameters.
These make neural systems adaptive and flexible, and are critical to their biological function. 共ii兲 In
contrast to traditional physical systems described by well-known basic principles, first principles
governing the dynamics of neural systems are unknown. 共iii兲 Many different neural systems exhibit
similar dynamics despite having different architectures and different levels of complexity. 共iv兲 The
network architecture and connection strengths are usually not known in detail and therefore the
dynamical analysis must, in some sense, be probabilistic. 共v兲 Since nervous systems are able to
organize behavior based on sensory inputs, the dynamical modeling of these systems has to explain the
transformation of temporal information into combinatorial or combinatorial-temporal codes, and vice
versa, for memory and recognition. In this review these problems are discussed in the context of
addressing the stimulating questions: What can neuroscience learn from nonlinear dynamics, and what
can nonlinear dynamics learn from neuroscience?
DOI: 10.1103/RevModPhys.78.1213
PACS number共s兲: 87.19.La, 05.45.⫺a, 84.35.⫹i, 87.18.Sn
CONTENTS
3. Synaptic plasticity
1222
4. Examples of the cooperative dynamics of
I. What are the Principles?
A. Introduction
individual neurons and synapses
1214
B. Robustness and adaptability in small microcircuits
1214
C. Intercircuit coordination
D. Chaos and adaptability
III. Informational Neurodynamics
A. Time and neural codes
1. Temporal codes
2. Spatiotemporal codes
3. Coexistence of codes
4. Temporal-to-temporal information
transformation: Working memory
B. Information production and chaos
1. Stimulus-dependent motor dynamics
2. Chaos and information transmission
C. Synaptic dynamics and information processing
B. Classical nonlinear dynamics approach for neural
systems
C. New paradigms for contradictory issues
1215
1217
II. Dynamical Features of Microcircuits: Adaptability and
Robustness
1218
A. Dynamical properties of individual neurons and
synapses
1218
1. Neuron models
1218
2. Neuron adaptability and multistability
1219
*Electronic address: [email protected]
0034-6861/2006/78共4兲/1213共53兲
1213
1223
1224
1228
1229
1231
1231
1231
1232
1233
1234
1237
1237
1239
1240
©2006 The American Physical Society
1214
Rabinovich et al.: Dynamical principles in neuroscience
D. Binding and synchronization
IV. Transient Dynamics: Generation
and Processing of Sequences
A. Why sequences?
B. Spatially ordered networks
1. Stimulus-dependent modes
2. Localized synfire waves
C. Winnerless competition principle
1. Stimulus-dependent competition
2. Self-organized WLC networks
3. Stable heteroclinic sequence
4. Relation to experiments
D. Sequence learning
E. Sequences in complex systems with random
connections
F. Coordination of sequential activity
V. Conclusion
Acknowledgments
Glossary
References
1242
1244
1244
1244
1244
1247
1248
1248
1249
1250
1251
1252
1254
1256
1258
1259
1259
1260
“Will it ever happen that mathematicians will know
enough about the physiology of the brain, and neurophysiologists enough of mathematical discovery, for efficient cooperation to be possible?”
—Jacques Hadamard
I. WHAT ARE THE PRINCIPLES?
A. Introduction
Building dynamical models to study the neural basis
of behavior has a long tradition 共Ashby, 1960; Block,
1962; Rosenblatt, 1962; Freeman, 1972, 2000兲. The underlying idea governing neural control of behavior is the
three-step structure of nervous systems that have
evolved over billions of years, which can be stated in its
simplest form as follows: Specialized neurons transform
environmental stimuli into a neural code. This encoded
information travels along specific pathways to the brain
or central nervous system composed of billions of nerve
cells, where it is combined with other information. A
decision to act on the incoming information then requires the generation of a different motor instruction set
to produce the properly timed muscle activity we recognize as behavior. Success in these steps is the essence of
survival.
Given the present state of knowledge about the brain,
it is impossible to apply a rigorous mathematical analysis
to its functions such as one can apply to other physical
systems like electronic circuits, for example. We can,
however, construct mathematical models of the phenomena in which we are interested, taking account of what is
known about the nervous system and using this information to inform and constrain the model. Current knowledge allows us to make many assumptions and put them
into a mathematical form. A large part of this review
will discuss nonlinear dynamical modeling as a particularly appropriate and useful mathematical framework
that can be applied to these assumptions in order to
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 1. 共Color online兲 Illustration of the functional parts and
electrical properties of neurons. 共a兲 The neuron receives inputs
through synapses on its dendritic tree. These inputs may or
may not lead to the generation of a spike at the spike generation zone of the cell body that travels down the axon and triggers chemical transmitter release in the synapses of the axonal
tree. If there is a spike, it leads to transmitter release and
activates the synapses of a postsynaptic neuron and the process
is repeated. 共b兲 Simplified electrical circuit for a membrane
patch of a neuron. The nonlinear ionic conductances are voltage dependent and correspond to different ion channels. This
type of electrical circuit can be used to model isopotential
single neurons. Detailed models that describe the morphology
of the cells use several isopotential compartments implemented by these circuits coupled by a longitudinal resistance;
these are called compartmental models. 共c兲 A typical spike
event is of the order of 100 mV in amplitude and 1 – 2 ms in
duration, and is followed by a longer after-hyperpolarization
period during which the neuron is less likely to generate another spike; this is called a refractory period.
simulate the functioning of the different components of
the nervous system, to compare simulations with experimental results, and to show how they can be used for
predictive purposes.
Generally there are two main modeling approaches
taken in neuroscience: bottom-up and top-down models.
• Bottom-up dynamical models start from a description of individual neurons and their synaptic connections, that is, from acknowledged facts about the details resulting from experimental data that are
essentially reductionistic 共Fig. 1兲. Using these anatomical and physiological data, the particular pattern
of connectivity in a circuit is reconstructed, taking
into account the strength and polarity 共excitatory or
inhibitory兲 of the synaptic action. Using the wiring
diagram thus obtained along with the dynamical features of the neurons and synapses, bottom-up models
have been able to predict functional properties of
Rabinovich et al.: Dynamical principles in neuroscience
neural circuits and their role in animal behavior.
• Top-down dynamical models start with the analysis
of those aspects of an animal’s behavior that are robust, reproducible, and important for survival. The
top-down approach is a more speculative big-picture
view that has historically led to different levels of
analysis in brain research. While this hierarchical division has put the different levels on an equal footing, the uncertainty implicit in the top-down approach should not be minimized. The first step in
building such large-scale models is to determine the
type of stimuli that elicit specific behaviors; this
knowledge is then used to construct hypotheses
about the dynamical principles that might be responsible for their organization. The model should predict how the behavior evolves with a changing environment represented by changing stimuli.
It is possible to build a sufficiently realistic neural circuit model that expresses dynamical principles even
without knowledge of the details of the neuroanatomy
and neurophysiology of the corresponding neural system. The success of such models depends on the universality of the underlying dynamical principles. Fortunately, there is a surprisingly large amount of similarity
in the basic dynamical mechanisms used by neural systems, from sensory to central and motor processing.
Neural systems utilize phenomena such as synchronization, competition, intermittency, and resonance in
quite nontraditional ways with regard to classical nonlinear dynamics theory. One reason is that the nonlinear
dynamics of neural modules or microcircuits is usually
not autonomous. These circuits are continuously or sporadically forced by different kinds of signals, such as sensory inputs from the changing environment or signals
from other parts of the brain. This means that when we
deal with neural systems we have to consider stimulusdependent synchronization, stimulus-dependent competition, etc. This is a departure from the considerations of
classical nonlinear dynamics. Another very important
feature of neuronal dynamics is the coordination of neural activities with very different time scales, for example,
theta rhythms 共4 – 8 Hz兲 and gamma rhythms
共40– 80 Hz兲 in the brain.
One of our goals in this review is to understand why
neural systems are very specific from the nonlinear dynamics point of view and to discuss the importance of
such specificities for the functionality of neural circuits.
We will talk about the relationship between neuroscience and nonlinear dynamics using specific subjects as
examples. We do not intend to review here the methods
or the nonlinear dynamical tools that are important for
the analysis of neural systems as they have been discussed extensively in many reviews and books 共e.g.,
Guckenheimer and Holmes, 1986; Crawford, 1991;
Abarbanel et al., 1993; Ott, 1993; Kaplan and Glass,
1995; Abarbanel, 1997; Kuznetsov, 1998; Arnold et al.,
1999; Strogatz, 2001; Izhikevich, 2006兲.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1215
B. Classical nonlinear dynamics approach for neural systems
Let us say a few words about the role of classical dynamical theory. It might seem at first sight that the apparently infinite diversity of neural activity makes its dynamical description a hopeless, even meaningless, task.
However, here one can exploit the knowledge accumulated in classical dynamical theory, in particular, the
ideas put forth by Andronov in 1931 concerning the
structural stability of dynamical models and the investigation of their bifurcations 共Andronov, 1933; Andronov
and Pontryagin, 1937; Andronov et al., 1949兲. The essential points of these ideas can be traced back to Poincaré
共Poincaré, 1892; Goroff, 1992兲. In his book La Valeur de
la Science, Poincaré 共1905兲 wrote that “the main thing
for us to do with the equations of mathematical physics
is to investigate what may and should be changed in
them.” Andronov’s remarkable approach toward understanding dynamical systems contained three key points:
• Only models exhibiting activity that does not vary
with small changes of parameters can be regarded as
really suitable to describe experiments. He referred
to them as models or dynamical systems that are
structurally stable.
• To obtain insight into the dynamics of a system it is
necessary to characterize all its principal types of behavior under all possible initial conditions. This led
to Andronov’s fondness for the methods of phasespace 共state-space兲 analysis.
• Considering the behavior of the system as a whole
allows one to introduce the concept of topological
equivalence of dynamical systems and requires an
understanding of local and global changes of the dynamics, for example, bifurcations, as control parameters are varied.
Conserving the topology of a phase portrait for a dynamical system corresponds to a stable motion of the
system with small variation of the governing parameters.
Partitioning parameter space for the dynamical system
into regions with different phase-space behavior, i.e.,
finding the bifurcation boundaries, then furnishes a complete picture of the potential behaviors of a dynamical
model. Is it possible to apply such a beautiful approach
to biological neural network analysis? The answer is yes,
at least for small, autonomous neural systems. However,
even in these simple cases we face some important restrictions.
Neural dynamics is strongly dissipative. Energy derived from biochemical sources is used to drive neural
activity with substantial energy loss in action-potential
generation and propagation. Nearly all trajectories in
the phase space of a dissipative system are attracted by
some trajectories or sets of trajectories called attractors.
These can be fixed points 共corresponding to steady-state
activity兲, limit cycles 共periodic activity兲, or strange attractors 共chaotic dynamics兲. The behavior of dynamical
systems with attractors is usually structurally stable.
Strictly speaking a strange attractor is itself structurally
1216
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 2. Six examples of limit
cycle bifurcations observed in
living and model neural systems
关see Chay 共1985兲; Canavier et al.
共1990兲; Guckenheimer et al.
共1993兲; Huerta et al. 共1997兲; Crevier and Meister 共1998兲; Maeda
et al. 共1998兲; Coombes and Osbaldestin 共2000兲; Feudel et al.
共2000兲; Gavrilov and Shilnikov
共2000兲; Maeda and Makino
共2000兲; Mandelblat et al. 共2001兲;
Bondarenko et al. 共2003兲; Gu et
al. 共2003兲; Shilnikov and Cymbalyuk 共2005兲; Soto-Trevino et
al. 共2005兲兴.
unstable, but its existence in the system state space is a
structurally stable phenomenon. This is a very important
point for the implementation of Andronov’s ideas.
The study of bifurcations in neural models and in in
vitro experiments is a keystone for understanding the
dynamical origin of many single-neuron and circuit phenomena involved in neural information processing and
the organization of behavior. Figure 2 illustrates some
typical local bifurcations 关their support consists of an
equilibrium point or a periodic trajectory—see the detailed definition by Arnold et al. 共1999兲兴 and some global
bifurcations 共their support contains an infinite set of orbits兲 of periodic regimes observed in neural systems.
Many of these bifurcations are observed both in experiments and in models, in particular in the conductancebased Hodgkin-Huxley–type equations 共Hodgkin and
Huxley, 1952兲, considered the traditional framework for
modeling neurons, and in the analysis of network stability and plasticity.
The most striking results in neuroscience based on
classical dynamical system theory have come from
bottom-up models. These results include the description
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
of the diversity of dynamics in single neurons and
synapses 共Koch, 1999; Vogels et al., 2005兲, the spatiotemporal cooperative dynamics of small groups of neurons
with different types of connections 共Selverston et al.,
2000; Selverston, 2005兲, and the principles of synchronization in networks with dynamical synapses 共Loebel and
Tsodyks, 2002; Elhilali et al., 2004; Persi et al., 2004兲.
Some top-down models also have attempted a classical nonlinear dynamics approach. Many of these models
are related to the understanding and description of cognitive functions. Nearly half a century ago, Ashby hypothesized that cognition could be modeled as a dynamical process 共Ashby, 1960兲. Neuroscientists have
spent considerable effort implementing the dynamical
approach in a practical way. The most widely studied
examples of cognitive-type dynamical models are multiattractor networks: models of associative memory that
are based on the concept of an energy function or
Lyapunov function for a dynamical system with many
attractors 共Hopfield, 1982兲 关see also Cohen and Grossberg 共1983兲; Waugh et al. 共1990兲; Doboli et al. 共2000兲兴.
The dynamical process in such networks is often called
Rabinovich et al.: Dynamical principles in neuroscience
1217
“computation with attractors.” The idea is to design during the learning stage, in a memory network phase
space, a set of attractors, each of which corresponds to a
specific output. Neural computation with attractors involves the transformation of a given input stimulus,
which defines an initial state inside the basin of attraction of one attractor, leading to a fixed desired output.
The idea that computation or information processing
in neural systems is a dynamical process is broadly
accepted today. Many dynamical models of both
bottom-up and top-down type that address the encoding
and decoding of neural information as the inputdependent dynamics of a nonautonomous network have
been published in the last few years. However, there are
still huge gaps in our knowledge of the actual biological
processes underlying learning and memory, making accurate modeling of these mechanisms a distant goal. For
reviews see Arbib et al. 共1997兲 and Wilson 共1999兲.
Classical nonlinear dynamics has provided some basis
for the analysis of neural ensembles even with large
numbers of neurons in networks organized as layers of
nearly identical neurons. One of the elements of this
formulation is the discovery of stable low-dimensional
manifolds in a very high-dimensional phase space. These
manifolds are mathematical images of cooperative
modes of activity, for example, propagating waves in
nonequilibrium media 共Rinzel et al., 1998兲. Models of
this sort are also interesting for the analysis of spiral
waves in cortical activity as experimentally observed in
vivo and in vitro 共Huang et al., 2004兲. Many interesting
questions have been approached by using the phase portrait and bifurcation analysis of models and by considering attractors and other asymptotic solutions. Nevertheless, new directions may be required to address the
important complexity of nervous system functions.
C. New paradigms for contradictory issues
The human brain contains approximately 1011 neurons
and a typical neuron connects with ⬇104 other neurons.
Neurons show a wide diversity in terms of their morphology and physiology 共see Fig. 3兲. A wide variety of
intracellular and network mechanisms influence the activity of living neural circuits. If we take into account
that even a single neuron often behaves chaotically, we
might argue that such a complex system most likely behaves as if it were a turbulent hydrodynamic flow. However, this is not what is observed. Brain dynamics are
more or less regular and stable despite the presence of
intrinsic and external noise. What principles does nature
use to organize such behavior, and what mathematical
approaches can be utilized for their description? These
are the very difficult questions we need to address.
Several important features differentiate the nervous
system from traditional dynamical systems:
• The architecture of the system, the individual neural
units, the details of the dynamics of specific neurons,
as well as the connections among neurons are not
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 3. Examples of 共a兲 the anatomical diversity of neurons,
and 共b兲 the single-neuron membrane voltage activity associated with them. 共1兲 Lobster pyloric neuron; 共2兲 neuron in rat
midbrain; 共3兲 cat thalamocortical relay neuron; 共4兲 guinea pig
inferior olivary neuron; 共5兲 aplysia R15 neuron; 共6兲 cat thalamic reticular neuron; 共7兲 sepia giant axon; 共8兲 rat thalamic
reticular neuron; 共9兲 mouse neocortical pyramidal neuron; 共10兲
rat pituitary gonadotropin-releasing cell. In many cases, the
behavior depends on the level of current injected into the cell
as shown in 共b兲. Modified from Wang and Rinzel, 1995.
usually known in detail, so we can describe them
only in a probabilistic manner.
• Despite the fact that many units within a complex
neural system work in parallel, many of them have
different time scales and react differently to the same
nonstationary events from outside. However, for the
whole system, time is unified and coherent. This
means that the neural system is organized hierarchically, not only in space 共architecture兲 but also in time:
each behavioral event is the initial condition for the
next window of time. The most interesting phenomenon for a neural system is the presence not of at-
1218
Rabinovich et al.: Dynamical principles in neuroscience
tractor dynamics but of nonstationary behavior. Attractor dynamics assumes long-time evolution from
initial conditions; we must consider transient responses instead.
• The structure of neural circuits is—in principle—
genetically determined; however, it is nevertheless
not fixed and can change with experience 共learning兲
and through neuromodulation.
We could expand this list, but the facts mentioned already make the point that the nervous system is a very
special field for the application of classical nonlinear dynamics, and it is clear now why neurodynamics needs
new approaches and a fresh view.
We use the following arguments to support an optimistic view about finding dynamical principles in neuroscience:
• Complex neural systems are the result of evolution,
and thus their complexity is not arbitrary but follows
some universal rules. One such rule is that the organization of the central nervous system 共CNS兲 is hierarchical and based on neural modules.
• It is important to note that many modules are organized in a very similar way across different species.
Such units can be small, like central pattern generators 共CPGs兲, or much more complex, like sensory
systems. In particular, the structure of one of the oldest sensory systems, the olfactory system, is more or
less the same in invertebrates and vertebrates and
can be described by similar dynamical models.
• The possibility of considering the nervous system as
an ensemble of interconnected units is a result of the
high level of autonomy of its subsystems. The level
of autonomy depends on the degree of selfregulation. Self-regulation of neural units on each
level of the nervous system, including individual neurons, is a key principle determining hierarchical neural network dynamics.
• The following conjecture seems reasonable: Each
specific dynamical behavior of the network 共e.g.,
traveling waves兲 is controlled by only a few of the
many parameters of a system 共like neuromodulators,
for example兲, and these relevant parameters influence the specific cell or network dynamics
independently—at least in a first approximation. This
idea can be useful for the mathematical analysis of
network dynamics and can help to build an approximate bifurcation theory. The goal of this theory is to
predict the transformation of specific dynamics based
on bifurcation analysis in a low-dimensional control
subspace of parameters.
• For the understanding of the main principles of neurodynamics, phenomenological top-down models are
very useful because even different neural systems
with different architectures and different levels of
complexity demonstrate similar dynamics if they execute similar functions.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
In the main part of this review we discuss two critical
functional properties of neural systems that at first
glance appear incompatible: robustness and sensitivity.
Finding solutions to such apparent contradictions will
help us formulate some general dynamical principles of
biological neural network organization. We note two examples.
Many neural systems, especially sensory systems, must
be robust against noise and at the same time must be
very sensitive to incoming inputs. A new paradigm that
can deal with the existence of this fundamental contradiction is the winnerless competition 共WLC兲 principle
共Rabinovich et al., 2001兲. According to this principle, a
neural network with nonsymmetric inhibitory connections is able to exhibit structurally stable dynamics if the
stimulus is fixed, and qualitatively change its dynamics if
the stimulus is changed. This ability is based on different
features of the signal and the noise, and the different
ways they influence the dynamics of the system.
Another example is the remarkable reproducibility of
transient behavior. Because transient behavior, in contrast to the long-term stable stationary activity of attractors, depends on initial conditions, it is difficult to imagine how such behavior can be reproducible from
experiment to experiment. The solution to this paradox
is related to the special role of global and local inhibition, which sets up the initial conditions.
The logic of this review is related to the specificity of
neural systems from the dynamical point of view. In Sec.
II we consider the possible dynamical origin of robustness and sensitivity in neural microcircuits. The dynamics of information processing in neural systems is considered in Sec. III. In Sec. IV, together with other
dynamical concepts, we focus on a new paradigm of neurodynamics: the winnerless competition principle in the
context of sequence generation, sensory coding, and
learning.
II. DYNAMICAL FEATURES OF MICROCIRCUITS:
ADAPTABILITY AND ROBUSTNESS
A. Dynamical properties of individual neurons and synapses
1. Neuron models
Neurons receive patterned synaptic input and compute and communicate by transforming these synaptic
input patterns into an output sequence of spikes. Why
spikes? As spike wave forms are similar, information encoded in spike trains mainly relies on the interspike intervals. Relying on timing rather than on the details of
action-potential wave forms increases the reliability and
reproducibility in interneural communication. Dispersion and attenuation in transmission of neural signals
from one neuron to others changes the wave form of the
action potentials but preserves their timing information,
again allowing for reliability when depending on interspike intervals.
The nature of spike train generation and transformation depends crucially on the properties of many
voltage-gated ionic channels in neuron cell membranes.
Rabinovich et al.: Dynamical principles in neuroscience
The cell body 共or soma兲 of the neuron gives rise to two
kinds of processes: short dendrites and one or more
long, tubular axons. Dendrites branch out like trees and
receive incoming signals from other neurons. In some
cases the synaptic input sites are on dendritic spines,
thousands of which can cover the dendritic arbor. The
output process, the axon, transmits the signals generated
by the neuron to other neurons in the network or to an
effector organ. The spikes are rapid, transient, all-ornone 共binary兲 impulses, with a duration of about 1 ms
共see Fig. 1兲. In most cases, they are initiated at a specialized region at the origin of the axon and propagate
along the axon without distortion. Near its end, the tubular axon divides into branches that connect to other
neurons through synapses.
When the spike emitted by a presynaptic neuron
reaches the terminal of its axon, it triggers the emission
of chemical transmitters in the synaptic cleft 共the small
gap, of order a few tens of nanometers, separating the
two neurons at a synapse兲. These transmitters bind to
receptors in the postsynaptic neuron, causing a depolarization or hyperpolarization in its membrane, exciting or
inhibiting the postsynaptic neuron, respectively. These
changes in the polarization of the membrane relative to
the extracellular space spread passively from the synapses on the dendrites across the cell body. Their effects
are integrated, and, when there is a large enough depolarization, a new action potential is generated 共Kandel et
al., 2000兲. Other types of synapses called gap junctions
function as Ohmic electrical connections between the
membranes of two cells. A spike is typically followed by
a brief refractory period, during which no further spikes
can be fired by the same neuron.
Neurons are quite complex biophysical and biochemical entities. In order to understand the dynamics of neurons and neural networks, phenomenological models
have to be developed. The Hodgkin-Huxley model is
foremost among such phenomenological descriptions of
neural activity. There are several classes of neural models possessing various degrees of sophistication. We summarize the neural models most often considered in biological network development in Table I. For a more
detailed description of these models see, for example,
Koch 共1999兲, Gerstner and Kistler 共2002兲, and Izhikevich
共2004兲.
Detailed conductance-based neuron models take into
account ionic currents flowing across the membrane
共Koch, 1994兲. The neural membrane may contain several
types of voltage-dependent sodium, potassium, and calcium channels. The dynamics of these channels can also
depend on the concentration of specific ions. In addition, there is a leakage current of chloride ions. The flow
of these currents results in changes in the voltage across
the membrane. The probability that a type of ionic channel is open depends nonlinearly on the membrane voltage and the current state of the channel. These dependencies result in a set of several coupled nonlinear
differential equations describing the electrical activity of
the cell. The intrinsic membrane conductances can enable neurons to generate different spike patterns, inRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1219
cluding high-frequency bursts of different durations
which are commonly observed in a variety of motor neural circuits and brain regions 关see Fig. 3共b2兲兴. The biophysical mechanisms of spike generation enable individual neurons to encode different stimulus features into
distinct spike patterns. Spikes, and bursts of spikes of
different durations, code for different stimulus features,
which can be quantified without a priori assumptions
about those features 共Kepecs and Lisman, 2003兲.
How detailed does the description of neurons or synapses have to be to make a model of neural dynamics
biologically realistic while still remaining computationally tractable? It is reasonable to separate neuron models into two classes depending on the general goal of the
modeling. If we wish to understand, for example, how
the ratio of inhibitory to excitatory synapses in a neural
ensemble with random connections influences the activity of the whole network, it is reasonable to use a simple
model that keeps only the main features of neuron behavior. The existence of a spike threshold and the increase of the output spike rate with an increase in the
input may be sufficient. On the other hand, if our goal is
to explain the flexibility and adaptability of a small network like a CPG to a changing environment, the details
of the ionic channel dynamics can be of critical importance 共Prinz et al., 2004b兲. In many cases neural models
built on simplified paradigms lead to more detailed
conductance-based models based on the same dynamical
principles but implemented with more biophysically realistic mechanisms. A good indication that the level of
the description was chosen wisely comes if the model
can reproduce with the same parameters the main bifurcations observed in the experiments.
2. Neuron adaptability and multistability
Multistability in a dynamical system means the coexistence of multiple attractors separated in phase space at
the same value of the system’s parameters. In such a
system qualitative changes in dynamics can result from
changes in the initial conditions. A well-studied case is
the bistability associated with a subcritical AndronovHopf bifurcation 共Kuznetsov, 1998兲. Multistable modes
of oscillation can arise in delayed-feedback systems
when the delay is larger than the response time of the
system. In neural systems multistability could be a
mechanism for memory storage and temporal pattern
recognition in both artificial 共Sompolinsky and Kanter,
1986兲 and living 共Canavier et al., 1993兲 neural circuits. In
a biological nervous system recurrent loops involving
two or more neurons are found quite often and are particularly prevalent in cortical regions important for
memory 共Traub and Miles, 1991兲. Multistability emerges
easily in these loops. For example, the conditions under
which time-delayed recurrent loops of spiking neurons
exhibit multistability were derived by Foss et al. 共1996兲.
The study used both a simple integrate-and-fire neuron
and a Hodgkin-Huxley 共HH兲 neuron whose recurrent
inputs are delayed versions of their output spike trains.
The authors showed that two kinds of multistability with
1220
Rabinovich et al.: Dynamical principles in neuroscience
TABLE I. Summary of many frequently used neuronal models.
Model
Integrate-andfire neurons
Example
dv共t兲
dt
=
冦
−
v共t兲
+ Iext + Isyn共t兲, 0 ⬍ v共t兲 ⬍ ␪
␶
v共t−0 兲 = ␪
v共t+0 兲 = 0,
Isyn共t兲 = g
兺 f共t − t
spike兲
spikes
and
f共t兲=A关exp共−t / ␶1兲 − exp共−t / ␶2兲兴
冧
Variables
Remarks
References
v共t兲 is the neuron
membrane potential; ␪
is the threshold for
spike generation. Iext
is an external stimulus
current; Isyn is the
sum of the synaptic
currents; and ␶1 and
␶2 are time constants
characterizing the synaptic currents.
A spike occurs when
the neuron reaches
the threshold ␪ in v共t兲
after which the cell is
reset to the resting
state.
Lapicque,
1907
Rate models
ȧi共t兲 = Fi共ai共t兲兲关Gi„ai共t兲…
− 兺j␳ijQj„aj共t兲…兴
ai共t兲 ⬎ 0 is the spiking
rate of the ith neuron
or cluster; ␳ij is the
connection matrix;
and F , G , Q are
polynomial functions.
This is a generalization of the LotkaVolterra model 关see
Eq. 共9兲兴.
Fukai and
Tanaka, 1997;
Lotka, 1925;
Volterra, 1931
McCulloch and
Pitts
xi共n + 1兲 = ⌰共兺jgijxj共n兲 − ␪兲
1, x ⬎ 0
⌰共x兲 =
0, x 艋 0
再
␪ is the firing
threshold; xj共n兲 are
synaptic inputs at the
discrete “time” n; xi共n
+ 1兲 is the output.
Inputs and outputs
are binary 共one or
zero兲; the synaptic
connections gij are 1,
−1, or 0.
The first
computational model
for an artificial
neuron; it is also
known as a linear
threshold device
model. This model
neglects the relative
timing of neural
spikes.
McCulloch
and Pitts,
1943
v共t兲 is the membrane
potential, m共t兲, and
h共t兲, and n共t兲
represent empirical
variables describing
the activation and
inactivation of the
ionic conductances; I
is an external current.
The steady-state
values of the
conductance variables
m⬁ , h⬁ , n⬁ have a
nonlinear voltage
dependence, typically
through sigmoidal or
exponential functions.
These ODEs
represent point
neurons. There is a
large list of models
derived from this one,
and it has become the
principal tool in
computational
neuroscience. Other
ionic currents can be
added to the
right-hand side of the
voltage equation to
better reproduce the
dynamics and
bifurcations observed
in the experiments.
Hodgkin and
Huxley, 1952
x共t兲 is the membrane
potential, and y共t兲
describes the
dynamics of fast
currents; I is an
external current. The
parameter values a, b,
and c are constants
chosen to allow
spiking.
A reduced model
describing oscillatory
spiking neural
dynamics including
bistability.
FitzHugh,
1961;
Nagumo et
al., 1962
冎
˙ = g 关v − v共t兲兴
Hodgkin-Huxley Cv共t兲
L L
+ gNam共t兲3h共t兲关vNa − v共t兲兴
+ gKn共t兲4共vK兲 − v共t兲 + I,
˙ = m⬁„v共t兲… − m共t兲
m共t兲
␶m„v共t兲…
h⬁„v共t兲… − h共t兲
˙ =
h共t兲
␶h„v共t兲…
n⬁„v共t兲… − n共t兲
˙ =
n共t兲
␶n„v共t兲…
FitzHugh-Nagumo
ẋ ␮ x − cx3 − y + I,
ẏ = x + by − a
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1221
Rabinovich et al.: Dynamical principles in neuroscience
TABLE I. 共Continued.兲
Model
Variables
Remarks
兵E共x , t兲 , I共x , t兲其 are the
number density of
active excitatory and
inhibitory neurons at
location x of the
continuous neural
media. „wee共x兲 , wie共x兲 ,
wei共x兲 , wii共x兲… are
connectivity distributions among the populations of cells. 兵Le ,
Li其 are nonlinear responses reflecting different populations of
thresholds. The operator 丢 is a convolution involving the connectivity distributions.
The first “mean-field”
model. It is an
attempt to describe a
cluster of neurons, to
avoid the inherent
noisy dynamical
behavior of individual
neurons; by averaging
to a distribution noise
is reduced.
Wilson and
Cowan, 1973
v共t兲 is the membrane
potential; n共t兲
describes the recovery
activity of a calcium
current; I is an
external current.
Simplified model that
reduces the number of
dynamical variables of
the HH model. It
displays action
potential generation
when changing I leads
to a saddle-node
bifurcation to a limit
cycle.
Morris and
Lecar, 1981
˙ = y共t兲 + ax共t兲2 − bx共t兲3 − z共t兲 + I
Hindmarsh-Rose x共t兲
˙
y共t兲 = C − xx共t兲2 − y共t兲
˙ = rˆs关x共t兲 − x 兴 − z共t兲‰
z共t兲
x共t兲 is the membrane
potential; y共t兲
describes fast
currents; z共t兲 describes
slow currents; and I is
an external current.
Simplified model that
uses a polynomial
approximation to the
right-hand side of a
Hodgkin-Huxley
model. This model
fails to describe the
hyperpolarized
periods after spiking
of biological neurons.
Hindmarsh
and Rose,
1984
d␪i共t兲
=␻+
dt
␪共t兲 is the phase of
the ith neuron with
approximately
periodic behavior; and
Hij is the connectivity
function determining
how neuron i and j
interact.
First introduced for
chemical oscillators;
good for describing
strongly dissipative
oscillating systems in
which the neurons are
intrinsic periodic
oscillators.
Cohen et al.,
1982;
Ermentrout
and Kopell,
1984;
Kuramoto,
1984
xt represents the
spiking activity and yt
represents a slow
variable. A discrete
time map.
One of a class of
simplephenomenological models for spiking,
bursting neurons. This
kind of model can be
computationally very
fast, but has little biophysical foundation.
Cazelles et al.,
2001; Rulkov,
2002
Wilson-Cowan
Example
␮
␮
Morris-Lecar
⳵E共x , t兲
⳵t
⳵I共x , t兲
⳵t
= −E共x , t兲 + 关1 − rE共x , t兲兴
⫻ Le关E共x , t兲 丢 wee共x兲
− I共x , t兲 丢 wei共x兲 + Ie共x , t兲兴
= −I共x , t兲 + 关1 − rI共x , t兲兴
⫻ Li关E共x , t兲 丢 wie共x兲
− I共x , t兲 丢 wii共x兲 + Ii共x , t兲兴
˙ = g 关v − v共t兲兴 + n共t兲g
v共t兲
L L
n
⫻ 关vn − v共t兲兴
+ gmm⬁„v̇共t兲…关vm − v共t兲兴 + I,
˙ = ␭„v共t兲…关n „v共t兲… − n共t兲兴
n共t兲
⬁
1
v − vm
m⬁共v兲 = 1 + tanh 0
2
vm
1
v − vn
n⬁共v兲 = 1 + tanh 0
2
vn
v − vn
␭共v兲 = ␾n cosh
2vn0
共
共
兲
兲
0
Phase oscillator
models
Map models
xt+1共i兲 =
兺 H 共␪ 共t兲 − ␪ 共t兲兲
ij
i
j
j
␣
+ yt共i兲
1 + xt共i兲2
⑀
+
xt共j兲
N j
兺
yt+1共i兲 = yt共i兲 − ␴xt共i兲 − ␤
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
References
1222
Rabinovich et al.: Dynamical principles in neuroscience
respect to initial spiking functions exist, depending on
whether the neuron is excitable or repetitively firing in
the absence of feedback.
Following Hebb’s 共1949兲 ideas most studies of the
mechanisms underlying learning and memory focus on
changing synaptic efficacy. Learning is associated with
changing connectivity in a network. However, the network dynamics also depends on complex interactions
among intrinsic membrane properties, synaptic
strengths, and membrane-voltage time variation. Furthermore, neuronal activity itself modifies not only synaptic efficacy but also the intrinsic membrane properties
of neurons. Papers by Marder et al. 共1996兲 and Turrigiano et al. 共1996兲 present examples showing that
bistable neurons can provide short-term memory
mechanisms that rely solely on intrinsic neuronal properties. While not replacing synaptic plasticity as a powerful learning mechanism, these examples suggest that
memory in networks could result from an ongoing interplay between changes in synaptic efficacy and intrinsic
neuron properties.
To understand the biological basis for such computational properties we must examine both the dynamics of
the ionic currents and the geometry of neuronal morphology.
3. Synaptic plasticity
Synapses as well as neurons are dynamical nonlinear
devices. Although synapses throughout the CNS share
many features, they also have distinct properties. They
operate with the following sequences of events: A spike
is initiated in the axon near the cell body, it propagates
down the axon, and arrives at the presynaptic terminal,
where voltage-gated calcium channels admit calcium,
which triggers vesicle fusion and neurotransmitter release. The released neurotransmitter then binds to receptors on the postsynaptic neuron and changes their
conductance 共Nicholls et al., 1992; Kandel et al., 2000兲.
This series of events is regulated in many ways, making
synapses adaptive and plastic.
In particular, the strength of synaptic conductivity
changes in real time depending on their activity, as Katz
observed many years ago 共Fatt and Katz, 1952; Katz,
1969兲. A description of such plasticity was made in 1949
by Hebb 共1949兲. He proposed that “When an axon of
cell A is near enough to excite a cell B and repeatedly or
persistently takes part in firing it, some growth process
or metabolic change takes place in one or both cells such
that A’s efficiency, as one of the cells firing B, is increased.” This neurophysiological postulate has since
become a central concept in neuroscience through a series of classic experiments demonstrating Hebbian-like
synaptic plasticity. These experiments show that the efficacy of synaptic transmission in the nervous system is
activity dependent and continuously modified. Examples
of such modification are long-term potentiation and depression 共LTP and LTD兲, which involve increased or decreased conductivity, respectively, of synaptic connections between two neurons, leading to increased or
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
decreased activity over time. Long-term potentiation
and depression are presumed to produce learning by differentially facilitating the association between stimulus
and response. The role of LTP and LTD, if any, in producing more complex behaviors is less closely tied to
specific stimuli and more indicative of cognition, and is
not well understood.
Long-term potentiation was first reported in the hippocampal formation 共Bliss and Lomo, 1973兲. Changes
induced by LTP can last for many days. Long-term potentiation has long been regarded, along with its counterpart LTD, as a potential mechanism for short-termmemory formation and learning. In fact, the hypothesis
is widely accepted in learning and memory research that
activity-dependent synaptic plasticity is induced at appropriate synapses during memory formation and is
both necessary and sufficient for the information storage
underlying the type of memory mediated by the brain
area in which that plasticity is observed 关see for a review
Martin et al. 共2000兲兴. Hebb did not anticipate LTD in
1949, but along with LTP it is thought to play a critical
role in “rewiring” biological networks.
The notion of a coincidence requirement for Hebbian
plasticity has been supported by classic studies of LTP
and LTD using presynaptic stimulation coupled with
prolonged postsynaptic depolarization 关see, for example,
Malenka and Nicoll 共1999兲兴. However, coincidence there
was loosely defined with a temporal resolution of hundreds of milliseconds to tens of seconds, much larger
than the time scale of typical neuronal activity characterized by spikes that last for a couple of milliseconds. In
a natural setting, presynaptic and postsynaptic neurons
fire spikes as their functional outputs. How precisely
must such spiking activities coincide in order to induce
synaptic modifications? Experiments addressing this
critical issue led to the discovery of spike-timingdependent synaptic plasticity 共STDP兲. Spikes initiate a
sequence of complex biochemical processes in the
postsynaptic neuron during the short time window following synaptic activation. Identifying detailed molecular processes underlying LTP and LTD remains a complex and challenging problem. There is good evidence
that it consists of a competition between processes removing 共LTD兲 and processes placing 共LTP兲 phosphate
groups from on postsynaptic receptors, or increasing
共LTP兲 or decreasing 共LTD兲 the number of such receptors
in a dendritic spine. It is also widely accepted that
N-methyl-D-aspartate 共NMDA兲 receptors are crucial for
the development of LTP or LTD and that it is calcium
influx onto the postsynaptic cell that is critical for both
LTP and LTD.
Experiments on synaptic modifications of excitatory
synapses between hippocampal glutamatergic neurons in
culture 共Bi and Poo, 1998, 2001兲 共see Fig. 4兲 indicate that
if a presynaptic spike arrives at time tpre and a postsynaptic spike is observed or induced at tpost, then when
␶ = tpost − tpre is positive the incremental percentage increase in synaptic strength behaves as
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 4. Spike-timing-dependent synaptic plasticity observed in
hippocampal neurons. Each data point represents the relative
change in the amplitude of evoked postsynaptic current after
repetitive application of presynaptic and postsynaptic spiking
pairs 共1 Hz for 60 s兲 with fixed spike timing ⌬t, which is defined as the time interval between postsynaptic and presynaptic spiking within each pair. Long-term potentiation 共LTP兲 and
depression 共LTD兲 windows are each fitted with an exponential
function. Modified from Bi, 2002.
⌬g
⬇ a Pe −␤P␶ ,
g
共1兲
with ␤P ⬇ 1 / 16.8 ms. When ␶ ⬍ 0, the percentage decrease in synaptic strength behaves as
⌬g
⬇ − a De ␤D␶ ,
g
共2兲
with ␤D ⬇ 1 / 33.7 ms. aP and aD are constants. This is
illustrated in Fig. 4.
Many biochemical factors contribute differently to
LTP and LTD in different synapses. Here we discuss a
phenomenological dynamical model of synaptic plasticity 共Abarbanel et al., 2002兲 which is very useful for modeling neural plasticity; its predictions agree with several
experimental results. The model introduces two dynamical variables P共t兲 and D共t兲 that do not have a direct relationship with the concentration of any biochemical
components. Nonlinear competition between these variables imitates the known competition in the postsynaptic
cell. These variables satisfy the following simple firstorder kinetic equations:
dP共t兲
= f„Vpre共t兲…关1 − P共t兲兴 − ␤PP共t兲,
dt
dD共t兲
= g„Vpost共t兲…关1 − D共t兲兴 − ␤DD共t兲,
dt
共3兲
where the functions f共V兲 and g共V兲 are typical logistic or
sigmoidal functions which rise from zero to the order of
unity when their argument exceeds some threshold.
These driving or input functions are a simplification of
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1223
the detailed way in which each dynamical process is
forced. The P共t兲 process is associated with a particular
time constant 1 / ␤P while the D共t兲 process is associated
with a different time constant 1 / ␤D. Experiments show
that ␤P ⫽ ␤D, and this is the primary embodiment of the
two different time scales seen in many observations. The
two time constants are a coarse-grained representation
of the diffusion and leakage processes which dampen
and terminate activities. Presynaptic voltage activity
serves to release neurotransmitters in the usual manner
and this in turn induces the postsynaptic action of P共t兲,
which has a time course determined by the time constant
␤P−1. Similarly, the postsynaptic voltage, constant or time
varying, can be associated with the induction of the D共t兲
process.
P共t兲 and D共t兲 compete to produce a change in synaptic
strength ⌬g共t兲 as
d⌬g共t兲
= ␥关P共t兲D␩共t兲 − D共t兲P␩共t兲兴,
dt
共4兲
where ␩ ⬎ 1 and ␥ ⬎ 0. This dynamical model reproduces
some of the key STDP experimental results like, for example, those shown in Fig. 4. It also accounts for the
case where the postsynaptic cell is depolarized while a
presynaptic spike train is presented to it.
4. Examples of the cooperative dynamics of individual neurons
and synapses
To illustrate the dynamical significance of plastic synapses we consider the synchronization of two neurons: a
living neuron and an electronic model neuron coupled
through a STDP or inverse STDP electronic synapse.
Using hybrid circuits of model electronic neurons and
biological neurons is a powerful method for analyzing
neural dynamics 共Pinto et al., 2000; Szücs et al., 2000;
LeMasson et al., 2002; Prinz et al., 2004a兲. The representation of synaptic input to a cell using a computer to
calculate the response of the synapse to specified
presynaptic input goes under the name “dynamic clamp”
共Robinson and Kawai, 1993; Sharp et al., 1993兲. It has
been shown in modeling and in experiments 共Nowotny,
Zhigulin, et al., 2003; Zhigulin et al., 2003兲 that coupling
through plastic electronic synapses leads to neural synchronization or, more correctly, entrainment that is more
rapid, more flexible, and much more robust against
noise than synchronization mediated by connections of
constant strength. In these experiments the neural circuit consists of a specified presynaptic signal, a simulated
synapse 共via the dynamic clamp兲, and a postsynaptic biological neuron from the Aplysia abdominal ganglion.
The presynaptic neuron is a spike generator producing
spikes of predetermined form at predetermined times.
The synapse and its plasticity are simulated by dynamic
clamp software 共Nowotny, 2003兲. In each update cycle of
⬃100 ␮s the presynaptic voltage is acquired, the spike
generator voltage is updated, the synaptic strength is determined according to the learning rule, and the resulting synaptic current is calculated and injected into the
living neuron through a current injection electrode. As
1224
Rabinovich et al.: Dynamical principles in neuroscience
one presents the presynaptic signal many times, the synaptic conductance changes from one fixed value to another depending on the properties of the presynaptic
signal.
The calculated synaptic current is a function of the
presynaptic and postsynaptic potentials of the spike generator Vpre共t兲 and the biological neuron Vpost共t兲, respectively. It is calculated according to the following model.
The synaptic current depends linearly on the difference
between the postsynaptic potential Vpost and its reversal
potential Vrev, on an activation variable S共t兲, and on its
maximal conductance g共t兲:
Isyn共t兲 = g共t兲S共t兲关Vpost共t兲 − Vrev兴.
共5兲
The activation variable S共t兲 is a nonlinear function of the
presynaptic membrane potential Vpre and represents the
percentage of neurotransmitter docked on the postsynaptic cell relative to the maximum that can dock. It has
two time scales: a docking time and an undocking time.
We take it to satisfy the dynamical equation
S⬁„Vpre共t兲… − S共t兲
dS共t兲
=
.
dt
␶syn关S1 − S⬁„V1共t兲…兴
共6兲
S⬁共V兲 is a sigmoid function which we take to be
S⬁共V兲 =
再
tanh关共V − Vth兲/Vslope兴 for V ⬎ Vth
0
otherwise.
冎
共7兲
The time scale is ␶syn共S1 − 1兲 for neurotransmitter docking and ␶synS1 for undocking. For AMPA excitatory receptors, the docking time is about 0.5 ms, and the undocking time is about 1.5 ms. The maximal conductance
g共t兲 is determined by the learning rule discussed below.
In the experiments, the synaptic current is updated at
⬃10 kHz.
To determine the maximal synaptic conductance g共t兲
of the simulated STDP synapse, an additive STDP learning rule was used. This is accurate if the time between
presented spike pairs is long compared to the time between spikes in the pair. To avoid runaway behavior, the
additive rule was applied to an intermediate graw that
was then filtered through a sigmoid function. In particular, the change ⌬graw in synaptic strength is given by
⌬graw共⌬t兲 =
冦
⌬t − ␶0 −共⌬t−␶ 兲/␶
0 +
for ⌬t ⬎ ␶0
e
␶+
⌬t − ␶0 共⌬t−␶ 兲/␶
0 −
for ⌬t ⬍ ␶0 ,
A−
e
␶−
A+
冧
共8兲
where ⌬t = tpost − tpre is the difference between postsynaptic and presynaptic spike times. The parameters ␶+ and
␶− determine the widths of the learning windows for potentiation and depression, respectively, and the amplitudes A+ and A− determine the magnitude of synaptic
change per spike pair. The shift ␶0 reflects the finite time
of information transport through the synapse.
As one can see in Fig. 5, the postsynaptic neuron
quickly synchronizes to the presynaptic spike generator
which presents spikes with an interspike interval 共ISI兲 of
255 ms 共top panel兲. The synaptic strength continuously
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 5. Example of a synchronization experiment. Top: The
interspike intervals 共ISIs兲 of the postsynaptic biological neuron. Bottom: The synaptic strength g共t兲. Presynaptic spikes
with ISI of 255 ms were presented to a postsynaptic neuron
with periodic oscillations at an ISI of 330 ms. Before coupling
with the presynaptic spike generator, the biological neuron
spikes tonically at its intrinsic ISI of 330 ms. Coupling was
switched on with g共t = 0兲 = 15 nS at time 6100 s. As one can see
the postsynaptic neuron quickly synchronizes to the presynaptic spike generator 共top panel, dashed line兲. The synaptic
strength continuously adapts to the state of the postsynaptic
neuron, effectively counteracting adaptation and other modulations of the system. This leads to a very precise and robust
synchronization at a nonzero phase lag. The precision of the
synchronization manifests itself in small fluctuations of the
postsynaptic ISIs in the synchronized state. Robustness and
phase lag cannot be seen directly. Modified from Nowotny,
Zhigulin, et al., 2003.
adapts to the state of the postsynaptic neuron, effectively counteracting adaptation and other modulations
of the system 共bottom panel兲. This leads to a very precise and robust synchronization at a nonzero phase lag.
The precision of the synchronization manifests itself in
small fluctuations of the postsynaptic ISIs in the synchronized state. Robustness and phase lag cannot be
seen directly in Fig. 5. Spike-timing-dependent plasticity
is a mechanism that enables synchronization of neurons
with significantly different intrinsic frequencies as one
can see in Fig. 6. The significant increase in the regime
of synchronization associated with synaptic plasticity is a
welcome, perhaps surprising, result and addresses the
issue raised above about robustness of synchronization
in neural circuits.
B. Robustness and adaptability in small microcircuits
The precise relationship between the dynamics of individual neurons and the mammalian brain as a whole
remains extremely complex and obscure. An important
reason for this is a lack of knowledge on the detailed
cell-to-cell connectivity patterns as well as a lack of
knowledge on the properties of the individual cells. Although large-scale modeling of this situation is attempted frequently, parameters such as the number and
kind of synaptic connections can only be estimated. By
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 6. 共Color online兲 The presynaptic signal generator presents a periodic spike train with ISI of T1 to a postsynaptic
neuron with ISI of T02, before coupling. When neurons are
coupled, T02 → T2. We plot the ratio of these periods after coupling as a function of the ratio before coupling 共a兲, for a synapse with constant g and 共b兲 for a synaptic connection g共t兲
following the rule in the text. The enlarged domain of one-toone synchronization in the latter case is quite clear and, as
shown by the change in the error bar sizes, the synchronization
is much better. This result persists when noise is added to the
presynaptic signal and to the synaptic action 共not shown兲.
Modified from Nowotny, Zhigulin, et al., 2003.
using the less complex microcircuits 共MCs兲 of invertebrates, a more detailed understanding of neural circuit
dynamics is possible.
Central pattern generators are small MCs that can
produce stereotyped cyclic outputs without rhythmic
sensory or central input 共Marder and Calabrese, 1996;
1225
Stein et al., 1997兲. Thus CPGs are oscillators, and the
image of their activity in the corresponding system state
space is a limit cycle when oscillations are periodic and a
strange attractor in more complex cases. Central pattern
generators underlie the production of most motor commands for muscles that execute rhythmic animal activity
such as locomotion, breathing, heartbeat, etc. The CPG
output is a spatiotemporal pattern with specific phase
lags between the temporal sequences corresponding to
the different motor units 共see below兲.
The network architecture and the main features of
CPG neurons and synapses are known much better than
any other brain circuits. Examples of typical invertebrate CPG networks are shown in Fig. 7. Common to
many CPG circuits are electrical and inhibitory connections and the spiking-bursting activity of their neurons.
The characteristics of the spatiotemporal patterns generated by the CPG, such as burst frequency, phase, length,
etc., are determined by the intrinsic properties of each
individual neuron, the properties of the synapses, and
the architecture of the circuit.
The motor patterns produced by CPGs fall into two
categories: those that operate continuously such as respiration 共Ramirez et al., 2004兲 or heartbeat 共Cymbalyuk
et al., 2002兲, and those that are produced intermittently
such as locomotion 共Getting, 1989兲 or chewing 共Selverston, 2005兲. Although CPGs autonomously establish correct rhythmic firing patterns, they are under constant
supervision by descending fibers from higher centers and
by local reflex pathways. These inputs allow the animal
to constantly adapt its behavior to the immediate environment, which suggests that there is considerable flexibility in the dynamics of motor systems. In addition
there is now a considerable body of information showing
that anatomically defined small neural circuits can be
reconfigured in a more general way by neuromodulatory
substances in the blood, or released synaptically so that
they are functionally altered to produce different stable
spatiotemporal patterns, which must also be flexible in
response to sensory inputs on a cycle-by-cycle basis; see
FIG. 7. Examples of invertebrate CPG microcircuits from arthropod, mollusk, and annelid
preparations. All produce rhythmic spatiotemporal motor patterns when activated by
nonpatterned input. The black dots represent
chemical inhibitory synapses. Resistors represent electrical connections. Triangles are
chemical excitatory synapses, and diodes are
rectifying synapses 共electrical synapses in
which the current flows only in one direction兲.
Individual neurons are identifiable from one
preparation to another.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1226
Rabinovich et al.: Dynamical principles in neuroscience
Simmers and Moulins 共1988兲, for example.
Central pattern generators have similarity with neural
MCs in the brain 共Silberberg et al., 2005; Solis and Perkel, 2005; Yuste et al., 2005兲 and are often studied as
models of neural network function. In particular, there
are important similarities between vertebrate spinal
cord CPGs and neocortical microcircuits which have
been emphasized by Yuste et al. 共2005兲: 共i兲 CPG interactions, which are fundamentally inhibitory, dynamically
regulate the oscillations. Furthermore, subthresholdactivated voltage-dependent cellular conductances that
promote bistability and oscillations also promote synchronization with specific phase lags. The same cellular
properties are also present in neocortical neurons, and
underlie the observed oscillatory synchronization in the
cortex. 共ii兲 Neurons in spinal cord CPGs show bistable
membrane dynamics, which are commonly referred to as
plateau potentials. A correlate of bistable membrane behavior, in this case termed “up” and “down” states, has
also been described in the striatum and neocortex both
in vivo and in vitro 共Sanchez-Vives and McCormick,
2000; Cossart et al., 2003兲. It is still unclear whether this
bistability arises from intrinsic or circuit mechanisms or
a combination of the two 共Egorov et al., 2002; Shu et al.,
2003兲. 共iii兲 Both CPGs and cortical microcircuits demonstrate attractor dynamics and transient dynamics 关see,
for example, Abeles et al. 共1993兲; Ikegaya et al. 共2004兲兴.
共iv兲 Modulations by sensory inputs and neuromodulators
are also a common characteristic that is shared between
CPGs and cortical circuits. Examples in CPGs include
the modulation of oscillatory frequency, of temporal coordination among different populations of neurons, of
the amplitude of network activity, and of the gating of
CPG input and output 共Grillner, 2003兲. 共v兲 Switching between different states of CPG operation 共for example,
switching coordinated motor patterns for different
modes of locomotion兲 is under sensory afferent and neurochemical modulatory control. This makes CPGs multifunctional and dynamically plastic. Switching between
cortical activity states is also under modulatory control,
as shown, for example, by the role of the neurotransmitter dopamine in working memory in monkeys
共Goldman-Rakic, 1995兲. Thus modulation reconfigures
microcircuit dynamics and transforms activity states to
modify behavior.
The CPG concept was built around the idea that behaviorally relevant spatiotemporal cyclic patterns are
generated by groups of nerve cells without the need for
rhythmic inputs from higher centers or feedback from
structures that are moving. If activated, isolated invertebrate preparations can generate such rhythms for many
hours and as a result have been extremely important in
trying to understand how simultaneous cooperative interactions between many cellular and synaptic parameters can produce robust and stable spatiotemporal patterns 关see Fig. 8共d兲兴. An example of a three-neuron CPG
phase portrait is shown in Figs. 8共a兲–8共c兲. The effect of a
hyperpolarizing current leads to changes in the pattern
as reflected by the phase portrait in Figs. 8共b兲 and 8共c兲.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 8. 共Color兲 Phase portrait of typical CPG output. The data
were recorded in the pyloric CPG of the lobster stomatogastric
ganglion. Each axis represents the firing rate of one of three
pyloric neurons: LP, VD, and PD 共see Fig. 7兲. 共a兲 The orbit of
the oscillating pyloric system is shown in blue and the average
orbit is shown in red; 共b兲 the same but with a hyperpolarizing
dc current injected into the PD; 共c兲 the difference between the
averaged orbits; 共d兲 time series of the membrane potentials of
the three neurons. Figure provided by T. Nowotny, R. Levi,
and A. Szücs.
Neural oscillations arise either through interactions
among neurons 共network-based mechanism兲 or through
interactions among currents in individual neurons 共pacemaker mechanism兲. Some CPGs use both mechanisms.
In the simplest case, one or more neurons with intrinsic
bursting activity acts as the pacemaker for the entire
CPG circuit. The intrinsic currents may be constitutively
active or they may require activation by neuromodulators, so-called conditional bursters. Synaptic connections
act to determine the pattern by exciting or inhibiting
other neurons at the appropriate time. Such networks
are extremely robust and have generally been thought to
be present in systems in which the rhythmic activity is
active all or most of the time. In the second case, it is the
synaptic interactions between nonbursty neurons that
generate the rhythmic activity and many schemes for the
types of connections necessary to do this have been proposed. Usually reciprocal inhibition serves as the basis
for generating bursts in antagonistic neurons and there
are many examples of cells in pattern-generating microcircuits connected in this way 共see Fig. 7兲. Circuits of this
type are usually found for behaviors that are intermittent in nature and require a greater degree of flexibility
than those based on pacemaker cells.
Physiologists know that reciprocal inhibitory connections between oscillatory neurons can produce, as a result of the competition, sequential activity of neurons
and rhythmic spatiotemporal patterns 共Szekely, 1965;
Stent and Friesen, 1977; Rubin and Terman, 2004兲. However, even for a rather simple MC, consisting of just
three neurons, there is no quantitative description. If the
connections are symmetric, the MC can reach an attrac-
1227
Rabinovich et al.: Dynamical principles in neuroscience
tor. It is reasonable to hypothesize that asymmetric inhibitory connections are necessary to preserve the order
of patterns with more than two phases per cycle. The
contradiction, noted earlier, between robustness and
flexibility can then be resolved because external signals
can modify the effective topology of connections so one
can have functionally different networks for different
stimuli.
Theoretical analysis and computer experiments with
MCs based on the winnerless competition principle 共discussed in Sec. IV.C兲 show that sufficient conditions for
the generation of sequential activity do exist and the
range of allowed nonsymmetric inhibitory connections is
quite wide 共Rabinovich et al., 2001; Varona, Rabinovich,
et al., 2002; Afraimovich, Rabinovich, et al., 2004兲 We
illustrate this using a Lotka-Volterra rate description of
neuron activity:
冉
N
冊
dai共t兲
= ai共t兲 1 − 兺 ␳ij共Si兲aj共t兲 + Si,
dt
j=1
i = 1, . . . ,N,
共9兲
where the rate of each N neuron is ai共t兲, the connection
matrix is ␳ij, and the stimuli Si are constants here. This
model can be justified as a rate neural model as follows
共Fukai and Tanaka, 1997兲. The firing rate ai共t兲 and membrane potential vi共t兲 of the ith neuron can be described
by
ai共t兲 = G共vi − ␪兲,
共10兲
dvi共t兲
= − ␭vi共t兲 + Ii共t兲,
dt
共11兲
where G共vi − ␪兲 is a gain function, ␪ and ␭ are constants,
and the input current Ii共t兲 to neuron i is generated by the
rates aj共t兲 of the other neurons:
N
Ii共t兲 = Si − 兺 ␳ijaj共t兲.
共12兲
j
Here Si is the input and ␳ij is the strength of the inhibitory synapse from neuron j to neuron i. We suppose that
G共x兲 is a sigmoidal function:
G共x兲 = G0/关1 + exp共− ␤x兲兴.
共13兲
Let us then make two assumptions: 共i兲 the firing rate is
always much smaller than its maximum value G0; and 共ii兲
the system is strongly dissipative 共this is reasonable because we are considering inhibitory networks兲. Based on
these assumptions, after combining and rescaling Eqs.
共10兲–共13兲, we obtain the Lotka-Volterra rate description
共9兲 with an additional positive term on the right side that
can be replaced by a constant 关see Fukai and Tanaka
共1997兲 for details兴.
The tests of whether WLC is operating in a reduced
pyloric CPG circuit are shown in Fig. 9. This study used
estimates of the synaptic strengths shown in Fig. 9共a兲.
Some of the key questions here are these. 共i兲 What is the
minimal strength for the inhibitory synapse from the pyRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 9. 共Color online兲 Competition without winner in a model
of the pyloric CPG. 共a兲 Schematic diagram of the three-neuron
network used for rate modeling. Black dots represent chemical
inhibitory synapses with strengths given in nanoseconds 共X
⬎ 160兲. 共b兲 Phase portrait of the model: The limit cycle corresponding to the rhythmic activity is in the 2D simplex 共Zeeman
and Zeeman, 2002兲. 共c兲 Robustness in the presence of noise:
Noise introduced into the model shows no effect on the order
of activation for each type of neuron. Figure provided by R.
Huerta.
loric dilator 共PD兲 neuron or AB group to the VD neuron such that WLC exists? 共ii兲 Does the connectivity
obtained from the competition without winner condition
produce the order of activation observed in the pyloric
CPG? 共iii兲 Is this dynamics robust against noise, in the
sense that strong perturbations of the system do not alter the sequence? If the strengths of ␳ij are taken as
冢
冣
1
1.25
0
1
1.25 ,
␳ij = 0.875
X/80 0.625 1
the WLC formulas imply that the sufficient conditions
for a reliable and robust cyclic sequence are satisfied if
X ⬎ 160. The activation sequence of the rate model with
noise shown in Fig. 9共c兲 is similar to that observed experimentally in the pyloric CPG. When additive Gaussian noise is introduced into the rate equations, the activation order of neurons is not broken, but the period of
the limit cycle depends on the level of perturbation.
Therefore the cyclic competitive sequence is robust and
can be related to the synaptic connectivity seen in real
MCs. If individual neurons in a MC are not oscillating,
one can consider small subgroups of neurons that may
form oscillatory units and apply the WLC principle to
these units.
An important question about modeling the rhythmic
activity of small inhibitory circuits is how the specific
dynamics of individual neurons influences the network
rhythm generation. Figure 10 represents the threedimensional 共3D兲 projection of the many-dimensional
phase portrait of a circuit with the same architecture as
1228
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 10. 共Color online兲 Three-dimensional projection of the
many-dimensional phase portrait of a circuit with the same
architecture as the one shown in Fig. 9, using Hodgkin-Huxley
spiking-bursting neuron models.
shown in Fig. 9共a兲 but using Hodgkin-Huxley spikingbursting neuron models. The switching dynamics seen in
the rate model is shown in Fig. 9共c兲, and this circuit is
robust when noise is added to it.
Pairs of neurons can interact via inhibitory, excitatory,
or electrical 共gap junction兲 synapses to produce basic
forms of neural activity which can serve as the foundation for MC dynamics. Perhaps the most common 共and
well-studied兲 CPG interaction consists of reciprocal inhibition, an arrangement that generates a rhythmic
bursting pattern in which neurons fire approximately out
of phase with each other 共Wang and Rinzel, 1995兲. This
is called a half-center oscillator. It occurs when there is
some form of excitation to the two neurons sufficient to
cause their firing and some form of decay mechanism to
slow high firing frequencies. The dynamical range of the
bursting activity varies with the synapse strength and in
some instances can actually produce in-phase bursting.
Usually reciprocal excitatory connections 共unstable if
too large兲 or reciprocal excitatory-inhibitory connections
are able to reduce the intrinsic irregularity of neurons
共Varona, Torres, Abarbanel, et al., 2001兲.
Modeling studies with electrically coupled neurons
have also produced nonintuitive results 共Abarbanel et
al., 1996兲. While electrical coupling is generally thought
to provide synchrony between neurons, under certain
conditions the two neurons can burst out of phase with
each other 共Sherman and Rinzel, 1992; Elson et al., 1998,
2002兲; see Fig. 11 and also Chow and Kopell 共2000兲 and
Lewis and Rinzel 共2003兲. An interesting modeling study
of three neurons 共Soto-Trevino et al., 2001兲 with synapses that are activity dependent found that the synaptic
strengths self-adjusted in different combinations to produce the same three-phase rhythm. There are many examples of vertebrate MCs in which a collection of neurons can be conceptually isolated to perform a particular
function or to represent the canonical or modular circuit
for a particular brain region 关see Shepherd 共1998兲兴.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 11. Artificial electrical coupling between two living chaotic PD cells of the stomatogastric ganglion of a crustacean can
demonstrate both synchronous and asynchronous regimes of
activity. In this case the artificial electrical synapse was built on
top of the existing natural coupling between two PD cells. This
shows different synchronization levels 共a兲–共d兲 as a function of
the artificial coupling ga and a dc current I injected in one of
the cells. 共a兲 With their natural coupling ga = 0 the two cells are
synchronized and display irregular spiking-bursting activity. 共b兲
With an artificial electrical coupling that changes the sign of
the current ga = −200 nS, and thus compensates the natural
coupling, the two neurons behave independently. 共c兲 Increasing
the negative conductance leads to a regularized antiphase spiking activity 共by mimicking mutual inhibitory synapses兲. 共d兲
With no artificial coupling but adding a dc current two neurons
are synchronized, displaying tonic spiking activity. Modified
from Elson et al., 1988.
C. Intercircuit coordination
It is often the case that more or less independent MCs
must synchronize in order to perform some coordinated
function. There is a growing literature suggesting that
large groups of neurons in the brain synchronize oscillatory activity in order to achieve coherence. This may be
a mechanism for binding disparate aspects of cognitive
function into a whole 共Singer, 2001兲, as we will discuss in
Sec. III.D. However, it is more persuasive to examine
intercircuit coordination in motor circuits where the
phases of different segments or limbs actually control
movements. For example, the pyloric and gastric circuits
can be coordinated in the crustacean stomatogastric system by a higher-level modulatory neuron that channels
the faster pyloric rhythm to a key cell in the gastric mill
rhythm 共Bartos and Nushbaum, 1997; Bartos et al., 1999兲
共Fig. 12兲. In crab stomatogastric MCs, the gastric mill
cycle has a period of approximately 10 s while the pyloric period is approximately 1 s. When an identified
modulatory projection neuron 共MCN1兲 关Fig. 12共a兲兴 is activated, the gastric mill pattern is largely controlled by
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 12. 共a兲 Schematic circuit diagram underlying MCN1 activation of the gastric mill rhythm of a crustacean. The circuit
represents two phases of the rhythm, retraction 共left兲 and protraction 共right兲. Lighter lines represent inactive connections.
LG, Int1, and DG are members of the gastric CPG and AB
and PD are members of the pyloric CPG. Arrows represent
functional transmission pathways from the MCN1 neuron.
Bars are excitatory and dots are inhibitory. 共b兲 The gastric mill
cycle period; the timing of each cycle is a function of the pyloric rhythm frequency. With the pyloric rhythm turned off, the
gastric rhythm cycles slowly 共LG兲. Replacing the AB inhibition
of Int1 with current into LG using a dynamic clamp reduces
the gastric mill cycle period. Modified from Barots et al., 1999.
interactions between MCN1 and gastric neurons LG and
Int 1 共Bartos et al., 1999兲. When Int 1 is stimulated, the
AB to LG synapse 关see Fig. 12共b兲兴 plays a major role in
determining the gastric cycle period and coordination
between the two rhythms. The two rhythms become coordinated because LG burst onset occurs with a constant
latency after the onset of the triggering pyloric input.
These results suggest that intercircuit synapses can enable an oscillatory circuit to control the speed of a
slower oscillatory circuit as well as provide a mechanism
for intercircuit coordination 共Bartos et al., 1999兲.
Another type of intercircuit coupling occurs among
segmental CPGs. In the crayfish, abdominal appendages
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1229
called swimmerets beat in a metachronal rhythm from
posterior to anterior with a frequency-independent
phase lag of about 90°. Like most rhythms of this kind,
the phase lag must remain constant over different frequencies. In theoretical and experimental studies by
Jones et al. 共2003兲, it was shown that such phase constancy could be achieved by ascending and descending
excitatory and inhibitory synapses, if the right connections were made. It appears realistic to look at rhythmic
MCs as recurrent networks with many intrinsic feedback
connections so that the information on a complete spatiotemporal pattern is contained in the long-term activity of just a few neurons in the circuit. The number of
intercircuit connections necessary for coordination of
the rhythms is therefore much smaller than the total
number of neurons in the MC.
To investigate coordinating two elements of a population of neurons, one may investigate how various couplings, implemented in a dynamical clamp, might operate in the cooperative behavior of two pyloric CPGs.
This is a hybrid and simplified model of the more complex interplay between brain areas whose coordinated
activity might be used to achieve various functions. We
now describe such a set of experiments.
Artificially connecting neurons from the pyloric CPG
of two different animals using a dynamic clamp could
lead to different kinds of coordination depending on
which neurons are connected and what kind of synapses
are used 共Szücs et al., 2004兲. Connecting the pacemaker
group with electrical synapses could achieve same-phase
synchrony; connecting them with inhibitory synapses
provided much better coordination but out of phase.
The two pyloric circuits 共Fig. 13兲 are representative of
circuits driven by coupled pacemaker neurons that communicate with each other via both graded and conventional chemical interactions. But while the unit CPG
pattern is formed in this way, coordinating fibers must
use spike-mediated postsynaptic potentials only. It
therefore becomes important to know where in the circuit to input these connections in order to achieve maximum effectiveness in terms of coordinating the entire
circuit and ensuring phase constancy at different frequencies. Simply coupling the PDs together electrically
is rather ineffective although the bursts 共not spikes兲 do
synchronize completely even at high coupling strengths.
The fact that the two PDs are usually running at slightly
different frequencies leads to bouts of chaos in the two
neurons, i.e., a reduction in regularity. More effective
synchronization occurs when the pacemaker groups are
linked together with moderately strong reciprocal inhibitory synapses in the classic half-center configuration.
Bursts in two CPGs are of course 180° out of phase, but
the frequencies are virtually identical. The best in-phase
synchronization is obtained when the LPs are coupled to
the contralateral PDs with inhibitory synapses 共Fig. 13兲.
D. Chaos and adaptability
Over the past decades there have been many reports
of the observation of chaos in the analysis of various
1230
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 13. Coupling of two biological pyloric CPGs Pyl 1 and
Pyl 2 by means of dynamic clamp artificial inhibitory synapses.
The dynamic clamp is indicated by DCL. Reciprocal inhibitory
coupling between the pacemaker groups AB and PD leads to
antiphase synchronization while nonreciprocal coupling from
the LPs produces in-phase synchronization. Figure provided by
A. Szücs.
time courses of data from a variety of neural systems
ranging from the simple to the complex 共Glass, 1995;
Korn and Faure, 2003兲. Perhaps the outstanding feature
of these observations is not the presence of chaos but
the appearance of low-dimensional dynamical systems
as the origin of spectrally broadband, nonperiodic signals observed in many instances 共Rabinovich and Abarbanel, 1998兲. All chaotic oscillations occur in a bounded
state-space region of the system. This state space is captured by the multivariate time course of the vector of
dynamical degrees of freedom associated with neural
spike generation. These degrees of freedom are comprised of the membrane voltage and the characteristics
of the various ion currents in the cell. Using nonlinear
dynamical tools one can reconstruct a mathematically
faithful proxy state space for the neuron by using the
membrane voltage and its time-delayed values as coordinates for the state space 共see Fig. 14兲.
Chaos seems to be almost unavoidable in natural systems comprised of numerous simple or slightly complex
subsystems. As long as there are three or more dimensions, chaotic motions are generic in the broad mathematical sense. So neurons are dealt a chaotic hand by
nature and may have little choice but to work with it.
Accepting that chaos is more or less the only choice, we
can ask what benefits accrue to the robustness and
adaptability of neural activity.
Chaos itself may not be essential for living systems.
However, the multitude of regular regimes of operation
that can be accomplished in dynamical systems composed of elements which themselves can be chaotic gives
rise to a basic principle that nature may use for the orRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 14. Chaotic spiking-bursting activity of isolated CPG
neutrons. Top panel: Chaotic membrane potential time series
of a synaptically isolated LP neuron from the pyloric CPG.
Bottom panel: State-space attractor reconstructed from the
voltage measurements of the LP neuron shown in the top
panel using delayed coordinates 关x共t兲 , y共t兲 , z共t兲兴 = 关V共t兲 , V共t
− T兲 , V共t − 2T兲兴. This attractor is characterized by two positive
Lyapunov exponents. Modified from Rabinovich and Abarbanel, 1998.
ganization of neural assemblies. In other words, chaos is
not responsible for the work of various neural structures, but rather for the fact that those structures function at the edge of instability, and often beyond it. By
recognizing chaotic motions in a system state space as
unstable, but bounded, this geometric notion gives credence to the otherwise unappealing idea of system instability. The instability inherent in chaotic motions, or
more precisely in nonlinear dynamics of systems with
chaos, facilitates the extraordinary ability of neural systems to adapt, make transitions from one pattern of behavior to another when the environment is altered, and
consequently create a rich variety of patterns. Thus
chaos gives a means to explore the opportunities available to the system when the environment changes, and
acts as a precursor to adaptive, reliable, and robust behavior for living systems.
Throughout evolution neural systems have developed
different methods of self-control or self-organization.
On the one hand, such methods preserve all advantages
of the complex behavior of individual neurons, such as
allowing regulation of the time period of transitions between operating regimes, as well as regulation of the
operation frequency in any given regime. They also preserve the possibility of a rich variety of periodic and
nonperiodic regimes of behavior; see Fig. 11 and Elson
et al. 共1988兲 and Varona, Torres, Huerta, et al. 共2001兲. On
the other hand, these control or organizational techniques provide the needed predictability of behavioral
patterns in neural assemblies.
Organizing chaotic neurons through appropriate wiring associated with electrical, inhibitory, and excitatory
connections appears to allow for essentially regular operation of such an assembly 共Huerta et al., 2001兲. As an
example we mention the dynamics of an artificial micro-
Rabinovich et al.: Dynamical principles in neuroscience
1231
III. INFORMATIONAL NEURODYNAMICS
FIG. 15. Average bursting period of the model heartbeat CPG
activity as a function of the inhibitory coupling ⑀. Modified
from Malkov et al., 1996.
circuit that mimics the leech heartbeat CPG 共Calabrese
et al., 1995兲. This CPG model consists of six chaotic neurons implemented with Hindmarsh-Rose equations reciprocally coupled to their neighbors through inhibitory
synapses. The modeling showed that in spite of chaotic
oscillations of individual neurons the cooperative dynamics is regular and, most importantly, the period of
bursting of the cooperative dynamics sensitively depends on the values of the inhibitory connections
共Malkov et al., 1996兲 共see Fig. 15兲. This example shows
the high level of adaptability of a network consisting of
chaotic elements.
Chaotic signals have many of the traditional characteristics attributed to noise. In the present context we
recognize that both chaos and noise are able to organize
the irregular behavior of individual neurons or neural
assemblies, but the principal difference is that dynamical
chaos is a controllable irregularity, possessing structure
in state space, while noise is an uncontrollable action of
dynamical systems. This distinction is extremely important for information processing as discussed below 共see
Sec. III.B.2 and its final remarks兲.
There are several possible functions for noise 共Lindner et al., 2004兲, even seen as high-dimensional essentially unpredictable chaotic motion, in neural network
studies. In high-dimensional systems composed here of
many coupled nonlinear oscillators, there may be small
basins of attraction where, in principle, the system could
become trapped. Noise will blur the basin boundaries
and remove the possibility that the main attractors could
accidentally be missed and highly functional synchronized states lost to neuronal activity. Some noise may
persist in the dynamics of neurons to smooth out the
actions of the chaotic dynamics active in creating robust,
adaptable networks.
Chaos should not be mistaken for noise, as the former
has phase-space structure which can be utilized for synchronization, transmission of information, and regularization of the network for performance of critical functions. In the next section we discuss the role of chaos in
information processing and information creation.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
The flow of information in the brain goes from sensory systems, where it is captured and encoded, to central nervous systems, where it is further processed to
generate response signals. In the central nervous system
command signals are generated and transported to the
muscles to produce motor behavior. At all these stages
learning and memory processes that need specific representations take place. Thus it is not surprising that the
nervous system has to use different coding strategies at
different levels of the transport, storage, and use of information. Different transformations of codes have been
proposed for the analysis of spiking activity in the brain.
The details depend on the particular system under study
but some generalization is possible in the framework of
analyzing the spatial, temporal, and spatiotemporal
codes. There are many unknown factors related to the
cooperation between these different forms of information coding. Some key questions are as follows: 共i兲 How
can neural signals be transformed from one coding space
to another without loss of information? 共ii兲 What dynamical mechanisms are responsible for storing time in
memory? 共iii兲 Can neural systems generate new information based on their sensory inputs? In this section, we
discuss some important experimental results and new
paradigms that can help to address these questions.
A. Time and neural codes
Information from sensory systems arrives at sensory
neurons as analog changes in light intensity or temperature, or chemical concentration of an odorant, or skin
pressure, etc. These analog data are represented in internal neural circuit dynamics and computations by
action-potential sequences passed from sensory receivers to higher-order brain processes. Neural codes guarantee the efficiency, reliability, and robustness of the required neural computations 共Machens, Gollisch, et al.,
2005兲.
1. Temporal codes
Two of the central questions in understanding the dynamics of information processing in the nervous system
are how information is encoded and how the coding
space depends on time-dependent stimuli.
A code in the biophysical context of the nervous system is a specific representation of the information operated on or created by neurons. A code requires a unit of
information. However, this is already a controversial issue since, as we have previously discussed, information
is conveyed through chemical and electrical synapses,
neuromodulators, hormones, etc., which makes it difficult to point out a single universal unit of information. A
classical assumption at the cellular level, valid for many
neural systems, is that a spike is an all-or-nothing event
and thus a good candidate for a unit of information, at
least in a computational sense. This is not the only sim-
1232
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 16. 共Color online兲 Two possible codes for the activity of a
single neuron. In a rate code, different inputs 共A–D兲 are transformed into different output spiking rates. In a timing code,
different inputs are transformed into different spiking sequences with precise timing.
plification needed to analyze neural codes for a first approach. A coding scheme needs to determine a coding
space and take into account time.
A common hypothesis is to consider a universal time
for all neural elements. Although this is the approach we
discuss here, we remind the reader that this is also an
arguable issue, since neurons can sense time in many
different ways: by their intrinsic activity 共subcellular dynamics兲 or by external input 共synaptic and network dynamics兲. Internal and external 共network兲 clocks are not
necessarily synchronized and can have different degrees
of precision, time scales, and absolute references. Some
dynamical mechanisms can contribute to make neural
time unified and coherent.
On the one hand, when we consider just a single neuron, a spike as the unit of information, and a universal
time, we can talk about two different types of encoding:
the frequency of firing can encode information about the
stimulus in a rate code; on the other hand, the exact
temporal occurrence of spikes can encode the stimulus
and its response in a precise timing code. The two coding alternatives are schematically represented in Fig. 16.
In this context, a precise timing or temporal code is a
code in which relative spike timings 共rather than spike
counts兲 are essential for information processing. Several
experimental recordings have shown the presence of
both types of single-cell coding in the nervous system
共Adrian and Zotterman, 1926; Barlow, 1972; Abeles,
1991; McClurkin et al., 1991; Softky, 1995; Shadlen and
Newsome, 1998兲. In particular, fine temporal precision
and reliability of spike dynamics are reported in many
cell types 共Segundo and Perkel, 1969; Mainen and
Sejnowski, 1995; deCharms and Merzenich, 1996; de
Ryter van Steveninck et al., 1997; Segundo et al., 1998;
Mehta et al., 2002; Reinagel and Reid, 2002兲. Single neurons can display these two codes in different situations.
2. Spatiotemporal codes
A population of coupled neurons can have a coding
scheme different from the sum of the individual coding
mechanisms. Interactions among neurons through their
synaptic connections, i.e., their cooperative dynamics, allow for more complex coding paradigms. There is much
experimental evidence which shows the existence of socalled population codes that collectively express a complex stimulus better than the individual neurons 关see,
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
e.g., Georgopoulus et al. 共1986兲; Wilson and McNaughton 共1993兲; Fitzpatrick et al. 共1997兲; Pouget et al. 共2000兲兴.
The efficacy of population coding has been assessed
mainly using measures of mutual information in modeling efforts 共Seung and Sompolinsky, 1993; Panzeri et al.,
1999; Sompolinsky et al., 2001兲.
Two elements can be used to build population codes:
neuronal identity 共i.e., neuronal space兲 and the time occurrence of neural events 共i.e., the spikes兲. Accordingly,
information about the physical world can be encoded in
temporal or spatial 共combinatorial兲 codes, or combinations of these two: spike time can represent physical
time 共a pure temporal code兲, spike time can represent
physical space, neuronal space can represent physical
time 共a pure spatial code兲, and neuronal space can represent physical space 共Nádasdy, 2000兲. When we consider a population of neurons, information codes can be
spatial, temporal, or spatiotemporal.
Population coding can also be characterized as independent or correlated 共deCharms and Christopher,
1998兲. In an independent code, each neuron represents a
separate signal: all information that is obtainable from a
single neuron can be obtained from that neuron alone,
without reference to the activities of other neurons. For
a correlated or coordinated coding messages are carried
at least in part by the relative timing of the signals from
a population of neurons.
The presence of network coding, i.e., a spatiotemporal
dynamical representation of incoming messages, has
been confirmed in several experiments. As an example,
we discuss here the spatiotemporal representation of
episodic experiences in the hippocampus 共Lin et al.,
2005兲. Individual hippocampal neurons respond to a
wide variety of external stimuli 共Wilson and McNaughton, 1994; Dragoi et al., 2003兲. The response variability
at the level of individual neurons poses an obstacle to
the understanding of how the brain achieves its robust
real-time neural coding of the stimulus 共Lestienne,
2001兲. Reliable encoding of sensory or other network
inputs by spatiotemporal patterns resulting from the dynamical interaction of many neurons under the action of
the stimulus can solve this problem 共Hamilton and
Kauer, 1985; Laurent, 1996; Vaadia et al., 1999兲.
Lin et al. 共2005兲 showed that mnemonic short-time
episodes 共a form of one-trial learning兲 can trigger firing
changes in a set of CA1 hippocampal neurons with specific spatiotemporal relationships. To find such representations in the central nervous system of an animal is an
extremely difficult experimental and computational
problem. Because the individual neurons that participate in the representation of a specific stimulus and form
a temporal neural cluster in different trials can be different, it is necessary to measure simultaneously the activity of a large number of neurons. In addition, because of
the variability in the individual neuron responses, the
spatiotemporal patterns of different trials may also look
different. Thus, to show the functional importance of the
spatiotemporal representation of the stimulus, the
reader has to use sophisticated methods of data analysis.
Lin et al. 共2005兲 developed a 96-channel array to record
Rabinovich et al.: Dynamical principles in neuroscience
1233
FIG. 17. 共Color online兲 Temporal dynamics of individual CA1
neurons of the hippocampus in
response to “startling” events.
Spike raster plots 关共a兲–共d兲 upper, seven repetitions each兴 and
corresponding perievent histogram 关共a兲–共d兲 lower, bin width
500 ms兴 for units exhibiting the
four major types of firing
changes observed: 共a兲 transient
increase, 共b兲 prolonged increase, 共c兲 transient decrease,
共d兲 and prolonged decrease.
From Lin et al., 2005.
simultaneously the activity patterns of as many as 260
individual neurons in the mouse hippocampus during
various startling episodes 共air blow, elevator drop, and
earthquake shake兲. They used multiple-discriminant
analysis 共Duda et al., 2001兲 and showed that, even
though individual neurons express different temporal
patterns in different trials 共see Fig. 17兲, it is possible to
identify functional encoding units in the CA1 neuron
assembly 共see Fig. 18兲.
The representation of nonstationary sensory information, say, a visual stimulus, can use the transformation of
a temporal to a spatial code. The recognition of a specific neural feature can be implemented through the
transformation of a spatial code into a temporal one
through coincidence detection of spikes. A spatial representation can be transformed into a spatiotemporal
one to provide the system with higher capacity and robustness and sensitivity at the same time. Finally, a spatiotemporal code can be transformed into a spatial code
in processes related to learning and memory. These possibilities are summarized in Fig. 19.
Morphological constraints of neural connections in
some cases impose a particular spatial or temporal code.
For example, projection neurons transfer information
between areas of the brain along parallel pathways by
preserving the input topography as neuronal specificity
at the output. In many cases the input topography is
transformed to a different topography that is preserved;
for example, the retinotopic map of the primary visual
areas and somatotopic maps of the somatosensory and
motor areas. Other transformations do not preserve topology. These include transformations in place cells in
the hippocampus, and the tonotopic representation in
the auditory cortex. There is a high degree of convergence and divergence of projections in some of these
transformations that can be a computationally optimal
design 共Garcia-Sanchez and Huerta, 2003兲. In most of
these transformations, the temporal dimension of the
stimulus is encoded by spike timing or by the onset of
firing-rate transients.
An example of transforming a spatiotemporal code to
a pure spatial code was found in the olfactory system of
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
locusts, and has been modeled by Nowotny, Rabinovich,
et al. 共2003兲 and Nowotny et al. 共2005兲. Figure 20 gives a
graphical explanation of the connections involved. The
complex spatiotemporal code of sequences of transiently
synchronized groups of projection neurons in the antennal lobe 共Laurent et al., 2001兲 is sliced into temporal
snapshots of activity by feedforward inhibition and coincidence detection in the next processing layer, the mushroom body 共Perez-Orive et al., 2002兲. This snapshot code
is presumably integrated over time in the next stages of
the mushroom lobes, completing the transformation of
the spatiotemporal code in the antennal lobe to a purely
spatial code. It was shown in simulations that the temporal information on the sequence of activity in the antennal lobe that could be lost in downstream temporal
integration can be restored through slow lateral excitation in the mushroom body 共Nowotny, Rabinovich, et al.,
2003兲. This has been reported experimentally 共Leitch
and Laurent, 1996兲. With this extra feature the transformation from a spatiotemporal code to a pure spatial
code becomes free of information loss.
3. Coexistence of codes
Different stages of neural information processing are
difficult to study in isolation. In many cases it is hard to
distinguish between what is an encoding of an input and
what is a static or dynamic, perhaps nonlinear, response
to that input. This is a crucial observation that is often
missed. Encoding and decoding may or may not be part
of a dynamical process. However, the creation of information 共discussed in the next section兲 and the transformation of spatial codes to temporal or spatiotemporal
codes are always dynamical processes.
Another, but less frequently addressed, issue about
coding is the presence of multiple encodings in singlecell signals 共Latorre et al., 2006兲. This may occur since
multifunctional networks may need multiple coexisting
codes. The neural signatures in interspike intervals of
CPG neurons provide an example 共Szücs et al., 2003兲.
Individual fingerprints characteristic of the activity of
each neuron coexist with the encoding of information in
1234
Rabinovich et al.: Dynamical principles in neuroscience
formulation of information theory 共Fano, 1961; Gallager,
1968兲. For example, cochlear afferents in birds bifurcate
to two different areas of the brain with different decoding properties. One area extracts information about
relative timing from a spike train, whereas the other extracts the average firing rate 共Konishi, 1990兲.
4. Temporal-to-temporal information transformation: Working
memory
FIG. 18. 共Color online兲 Classification, visualization, and dynamical decoding of CA1 ensemble representations of startle
episodes by multiple-discriminant analysis 共MDA兲 methods.
共a兲 Firing patterns during rest, air blow, drop, and shake epochs are shown after being projected to a three-dimensional
space obtained by using MDA for mouse A; MDA1–MDA3
denote the discriminant axes. Both training 共dark symbols兲 and
test data are shown. After the identification of startle types, a
subsequent MDA is further used to resolve contexts 共full vs
empty symbols兲 in which the startle occurred for air-blow context 共b兲 and for elevator drop 共c兲. 共d兲 Dynamical monitoring of
ensemble activity and the spontaneous reactivation of startle
representations. Three-dimensional subspace trajectories of
the population activity in the two minutes after an air-blow
startle in mouse A are shown. The initial response to an air
blow 共black line兲 is followed by two large spontaneous excursions 共blue/dark and red/light lines兲, characterized by coplanar,
geometrically similar lower-amplitude trajectories 共directionality indicated by arrows兲. 共e兲 The same trajectories as in 共a兲
from a different 3D angle. 共f兲 The timing 共t1 = 31.6 s and t2
= 54.8 s兲 of the two reactivations 共marked in blue/dark and red/
light, respectively兲 after the actual startle 共in black兲 共t = 0 s兲.
The vertical axis indicates the air-blow classification probability. From Lin et al., 2005.
the frequency and phase of the spiking-bursting
rhythms. This is an example that shows that codes can
be nonexclusive. In bursting activity, coding can exist in
slow waves, but also, and simultaneously, in the spiking
activity.
In the brain, specific neural populations often send
messages through projections to several information
“users.” It is difficult to imagine that all of them decode
the incoming signals in the same way. In neuroscience
the relationship between the encoder and decoder is not
a one-to-one map but can be many simultaneous maps
from the senders to different receivers, based on different dynamics. This departs from Shannon’s classical
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
There is another important code transformation of interest here: the transformation of a finite amount of
temporal information to a slow temporal code lasting for
seconds, minutes, or hours. We are able to remember a
phone number from someone who just called us. Persistent dynamics is one of the mechanisms for this phenomenon, which is usually named short-term memory 共STM兲
or working memory; it is a basic function of the brain.
Working memory, in contrast to long-term memory
which most likely requires molecular 共membrane兲 or
structural 共connection兲 changes in neural circuits, is a
dynamical process. The dynamical origins of working
memory can vary.
One plausible idea is that STMs are the result of active reverberation in interconnected neural clusters that
fire persistently. Since its conceptualization 共de Nó, 1938;
Hebb, 1949兲, reverberating activity in microcircuits has
been explored in many modeling papers 共Grossberg,
1973; Amit and Brunel, 1997a; Durstewitz et al., 2000;
Seung et al., 2000; Wang, 2001兲. Experiments with cultured neuronal networks show that reverberatory activity can be evoked in circuits that have no preexisting
anatomical specialization 共Lau and Bi, 2005兲. The reverberation is primarily driven by recurrent synaptic excitation rather than complex individual neuron dynamics
such as bistability. The circuitry necessary for reverberating activity can be a result of network selforganization. Persistent reverberatory activity can exist
even in the simplest circuit, i.e., an excitatory neuron
with inhibitory self-feedback 共Connors, 2002; Egorov
et al., 2002兲. In this case, reverberation depends on asynchronous transmitter release and intracellular calcium
stores as shown in Fig. 21.
Nature seems to use different dynamical mechanisms
for persistent microcircuit activity: cooperation of many
interconnected neurons, persistent dynamics of individual neurons, or both. These mechanisms each have
distinct advantages. For example, network mechanisms
can be turned on and off quickly 共McCormick et al.,
2003兲 关see also Brunel and Wang 共2001兲兴. Most dynamical models with persistent activity are related to the
analysis of microcircuits with local feedback excitation
between principal neurons controlled by disynaptic
feedback inhibition. Such basic circuits spontaneously
generate two different modes: relative quiescence and
persistent activity. The triggering between modes is controlled by incoming signals. The review by Brunel 共2003兲
considers several basic models of persistent dynamics,
including bistable networks with excitation only and
multistable models for working memory of a discrete set
1235
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 19. 共Color online兲 Summary of possible scenarios for
the transformation of codes,
their functional implications,
and the dynamical mechanism
involved.
of pictures with structured excitation and global inhibition.
Working memory is used for tasks such as planning,
organizing, rehearsing, and movement preparation. Experiments with functional magnetic resonance imaging
reveal some aspects of the dynamics of working memory
关see, for example, Diwadkar et al. 共2000兲 and Nystrom et
al. 共2000兲兴. It is important to note that working memory
has a limited capacity of around four to seven items
共Cowan, 2001; Vogel and Michizawa, 2004兲. An essential
feature attributed to working memory is the labile and
transient nature of its representations. Because such
representations involve many coupled neurons from cortical areas 共Curts and D’Esposito, 2003兲, it is natural to
model working memory as the spatiotemporal dynamics
of large neural networks.
A popular idea is to model working memory with attractors. Representation of items in working memory by
attractors may guarantee its robustness. Although robustness is an important requisite for a working-memory
system, its transient properties are also important. Consider a foraging task in which an animal uses visual input
to catch prey 共Nakahara and Doya, 1998兲. It is helpful to
store the location of the prey in the animal’s working
memory if the prey goes behind a bush and the sensory
cue becomes temporarily unavailable. However, the
memory should not be retained forever because the prey
may have actually gone away or may have been eaten by
another animal. Furthermore, if more prey appears near
the animal, the animal should quickly load the location
of the new prey into its working memory without being
disturbed by the old memory.
This example illustrates that there are more requirements for a working-memory system than solely robust
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
maintenance. First, the activity should be maintained
but not for too long. Second, the activity should be reset
quickly when there is a novel sensory cue that needs to
be stored. In other words, the neural dynamics involved
in working memory for goal-directed behaviors should
have the properties of long-term maintenance and quick
switching. A corresponding model based on “nearsaddle-node” bifurcation dynamics has been suggested
by Nakahara and Doya 共1998兲. The authors have analyzed the dynamics of a network of model neural units
that are described by the following map 共see Fig. 22兲:
冉
冊
yi共tn+1兲 = F ayi共tn兲 + b + 兺 ␳ijyi共tn兲 + ␥iIi共tn兲 ,
j⫽i
共14兲
where yi共tn兲 is the firing rate of the ith unit at time tn,
F共z兲 = 1 / 关1 + exp共−z兲兴 is a sigmoid function, a is the selfconnection weight, ␳ij are the lateral connection weights,
Ii共t兲 are external inputs, b is the bias, and ␥i are constants used to scale the inputs Ii共t兲. As the bias b is increased, the number of fixed points changes sequentially
from one to two, three, two, and then back to one. A
saddle-node bifurcation occurs when the stable transition curve y共tn+1兲 = F共z兲 is tangent to the fixed point
y共tn+1兲 = y共tn兲 共see Fig. 22兲. Just near the saddle-node bifurcation the system shows persistent activity. This
means that it spends a long time in the narrow channel
between the bisectrix and the sigmoid activation curve
and then goes to the fixed point quickly. Such dynamical
behavior reminds one of the well-known intermittency
phenomenon in physics 共Landau and Lifshitz, 1987兲. Because the effect of the sum of the lateral and external
inputs in Eq. 共14兲 is equivalent to a change in the bias,
the mechanism may satisfy the requirements of the dy-
1236
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 20. 共Color兲 Illustration of the transformation of temporal
into spatial information. If a coincidence detection occurs, the
local excitatory connections activate the neighbors of the active neuron 共yellow neurons兲. Coincidence detection of input is
now more probable in these activated neighborhoods than in
other Kenyon cells 共KCs兲. Which of the neighbors might fire a
spike, however, depends on the activity of the projection neurons 共PNs兲 in the next cycle. It might be a different neuron for
active group B of PNs 共upper branch兲 than for active group C
共lower branch兲. In this way local sequences of active KCs form.
These depend on the identity of active PNs 共coincidence detection兲 as well as on the temporal order of their activity 共activated neighborhoods兲. Modified from Nowotny, Rabinovich,
et al., 2003.
namics of working memory for goal-directed behavior:
long-term maintenance and quick switching.
Another reasonable model for working memory consists of competitive networks with stimulus-dependent
inhibitory connections 关as in Eq. 共9兲兴. One of the advantages of such a model is the ability to have both working
memory and stimulus discrimination. This idea was proposed by Machens, Romo, et al. 共2005兲 in relation to the
frontal-lobe neural architecture. The network first perceives the stimulus, then holds it in the working memory,
and finally makes a decision by comparing that stimulus
with another one. The model integrates both working
memory and decision making since the number of stable
fixed points and the size of the basins of attractors are
controlled by the connection matrix ␳ij共S兲 which depends on the stimuli S. The working-memory phase corresponds to the bifurcation boundary, i.e., ␳ij = ␳ji = ␳ii. In
the state space of the dynamical model, this phase is
represented by a stable manifold called a “continuous
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 21. 共Color online兲 Reverberation can be the dynamical
origin for working memory in minimal circuits. 共a兲 Most neurons respond to excitatory stimuli 关upward steps in the line
below 共c兲兴 by spiking only as long as each stimulus lasts. 共b兲
Very rare neurons are bistable: brief excitation leads to persistent spiking, always at the same rate; brief inhibition 关downward steps in the line below 共c兲兴 can turn it off. 共c兲 Multistable
neurons persistently increase or decrease their spiking across a
range of rates in response to repeated brief stimuli. 共d兲 In the
reverberatory network model of short-term memory discussed
in the text, an excitatory stimulus 共left arrow兲 leads to recursive activity in interconnected neurons. Inhibitory stimuli 共bar
on the right兲 can halt the activity. 共e兲 Egorov et al. 共2002兲 suggest that graded persistent activity in single neurons 关as in 共c兲兴
might be triggered by a pulse of internal Ca2+ ions that enter
through voltage-gated channels; Ca2+ then activates calciumdependent nonspecific cation 共CAN兲 channels, through which
an inward current 共largely comprising Na+ ions兲 enters, persistently exciting the neuron. The positive feedback loop 共broken
arrows兲 may include the activity of many ionic channels. Modified from Connors, 2002.
attractor.” This is an attractor that consists of continuous
sets of fixed points 关see Amari 共1977兲 and Seung 共1998兲兴.
Thus the stimulus creates a specific fixed point and, at
the next stage, the working memory 共a continuous attractor兲 maintains it. During the comparison and decision phase, the second stimulus is mapped onto the same
state space as another attractor. The criterion of the decision maker is reflected in the positions of the separatrices that separate the basins of attraction of different
FIG. 22. Temporal responses of self-recurrent units: Nearsaddle-node bifurcation with a = 11.11, b = −7.9 共center panels兲.
Increased bias, b = −3.0 共left panels兲. Decreased bias
b = −9.0 共right panels兲. Modified from Nakahara and Doya,
1998.
Rabinovich et al.: Dynamical principles in neuroscience
1237
FIG. 23. Hallucinations generated by LSD are an example of a
dynamical representation of the internal activity of the visual
cortex without an input stimulus. Figure shows examples of 共a兲
funnel and 共b兲 spiral hallucinations. Modified from Bressloff, et
al. 2001.
stimuli, i.e., fixed points 共see an alternative approach in
Rabinovich and Huerta, 2006兲.
We think that the intersection of the mechanisms responsible for persistent activity of single neurons with
the activity of a network with local or nonlocal recurrence provides robustness against noise and perturbations, and at the same time makes working memory
more flexible.
B. Information production and chaos
Information processing in the nervous system involves
more than the encoding, transduction, and transformation of incoming information to generate a corresponding response. In many cases, neural information is created by the joint action of the stimulus and the
individual neuron and network dynamics. A creative activity like improvisation on the piano or writing a new
poem results in part from the production of new information. This information is generated by neural circuits
in the brain and does not directly depend on the environment.
Time-dependent visual hallucinations are one example of information produced by neural systems, in
this case the visual cortex, themselves. Such hallucinations consist in seeing something that is not in the visual
field. There are interesting models, beginning from the
pioneering paper of Ermentrout and Cowan 共1979兲, that
explain how the intrinsic circuitry of the brain’s visual
cortex can generate the patterns of activity that underlie
hallucinations. These hallucination patterns usually take
the form of checkerboards, honeycombs, tunnels, spirals, and cobwebs 共see two examples in Fig. 23兲. Because
the visual cortex is an excitable medium it is possible to
use spatiotemporal amplitude equations to describe the
dynamics of these patterns 共see the next section兲. These
models are based on advances in brain anatomy and
physiology that have revealed strong short-range connections and weaker long-range connections between
neurons in the visual cortex. Hallucination patterns can
be quasistatic, periodically repeatable, or chaotically repeatable as in low-dimensional convective turbulence;
see for a review Rabinovich et al. 共2000兲. Unpredictability of the specific pattern in the hallucination sequences
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 24. 共Color online兲 Dual sensory network dynamics. Top
panels: Schematic representation of the dual role of a single
statocyst, the gravity sensory organ of the mollusk Clione. During normal swimming, a stonelike structure, the statolith, hits
the mechanoreceptor neurons that react to this excitation. In
Clione’s hunting behavior, the statocyst receptors receive additional excitation from the cerebral hunting neuron 共H兲 which
generates a winnerless competition among them. Bottom panels: Chaotic sequential switching displayed by the activity of
the statocyst during hunting mode in a model of a six-receptor
network. This panel displays the time intervals in which each
neuron is active 共ai ⬎ 0.03兲. Each neuron is represented by a
different color. The dotted rectangles indicate the activationsequence locks among units that are active at a given time
interval within each network for time windows in which all six
neurons are active.
共movie兲 means the generation of information that in
principle can be characterized by the value of the
Kolmogorov-Sinai entropy 共Scott, 2004兲.
The creation or production of new information is a
theme that has been neglected in theoretical neuroscience, but it is a provocative and challenging point that
we discuss in this section. As mentioned before, information production or creation must be a dynamical process. Below we discuss an example that emphasizes the
ability of neural systems to produce information-rich
output from information-poor input.
1. Stimulus-dependent motor dynamics
A simple network with which we can discuss the creation of new information is the gravity-sensing neural
network of the marine mollusk Clione limacina. Clione
is a blind planktonic animal, negatively buoyant, that
has to maintain continuous motor activity in order to
keep its preferred head-up orientation. Its motor activity
is controlled by wing CPGs and tail motor neurons that
use signals from its gravity-sensing organs, the statocysts
共Panchin et al., 1995兲. A six-receptor neural network
model with synaptic inhibition has been built to describe
a single statocyst 共Varona, Rabinovich, et al., 2002兲 共see
Fig. 24兲. This is a small sphere in which a statolith, a
stonelike structure, moves according to the gravitational
1238
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 25. 共Color online兲 Clione swimming trajectories in different situations. 共a兲 Three-dimensional trajectory of routine
swimming. Here and in the following figures, different colors
共gray tones兲 are used to emphasize the three-dimensional perception of the trajectories and change according to the x axis.
The indicated time t is the duration of the trajectory. 共b兲 Swimming trajectory of Clione with the statocysts surgically removed. 共c兲 Trajectory of swimming during hunting behavior
evoked by the contact with the prey. 共d兲 Trajectory of swimming after immersion of Clione in a solution that pharmacologically evokes hunting. Modified from Levi et al., 2004.
field. The statolith excites the neuroreceptors by pressing down on them. When excited, the receptors send
signals to the neural systems responsible for wing beating and tail orientation.
The statocysts have a dual role 共Levi et al., 2004,
2005兲. During normal swimming only neurons that are
excited by the statolith are active, and this leads to a
winner-take-all dynamical mode as a result of inhibitory
connections in the network. 共Winner-take-all dynamics is
essentially the same as the attractor-based computational ideas discussed earlier.兲 However, when Clione is
searching for its food, a cerebral hunting neuron excites
each neuron of the statocyst 共see Fig. 24兲. This triggers a
competition between all statocyst neurons whose signals
participate in the generation of a complex motion that
the animal uses to scan the immediate space until it finds
its prey 共Levi et al., 2004, 2005兲 共see Fig. 25兲. The following Lotka-Volterra-type dynamics can be used to describe the activity of this network:
冉
N
冊
dai共t兲
= ai共t兲 ␴共H,S兲 − 兺 ␳ijaj共t兲 + Hi共t兲 + Si共t兲,
dt
j=1
共15兲
where ai共t兲 艌 0 represents the instantaneous spiking rate
of the statocyst neurons, Hi共t兲 represents the excitatory
stimulus from the cerebral hunting interneuron to neuRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 26. 共Color online兲 Irregular switching in a network of six
statocyst receptors. Traces represent the instantaneous spiking
rate of each neuron ai 关neurons 1,2,3 are shown in 共a兲, neurons
4,5,6 in 共b兲兴. Note that after a neuron is silent for a while, its
activity reappears with the same sequence relative to the others 共see arrows, and Fig. 24兲. 共c兲 A projection of the phase
portrait of the strange attractor in 3D space; see model 共15兲.
ron i, Si共t兲 represents the action of the statolith on the
receptor that it is pressing, and ␳ij is the nonsymmetric
statocyst connection matrix. When there is no stimulus
from the hunting neuron, Hi = 0, or the statolith, Si = 0,
then ␴共H , S兲 = −1 and all neurons are silent. When the
hunting neuron is active Hi ⫽ 0 and/or the statolith is
pressing one of the receptors, Si ⫽ 0, ␴共H , S兲 = + 1.
During hunting Hi ⫽ 0, and we assume that the action
of the hunting neuron overrides the effect of the statolith and thus Si ⬇ 0. As a result of the competition, the
receptors display a highly irregular, in fact chaotic,
switching activity. The phase-space image of the chaotic
dynamics of the statocyst model in this behavioral mode
is a strange attractor 关the heteroclinic loops in the phase
space of Eq. 共15兲 become unstable; see Sec. IV.C兴. For
six receptors we have shown 共Varona, Rabinovich, et al.,
2002兲 that the observed dynamical chaos is characterized
by two positive Lyapunov exponents.
The bottom panel in Fig. 24 is an illustration of the
nonsteady switching activity of the receptors. An interesting phenomenon can be seen in this figure and is also
pointed out in Fig. 26. Although the timing of each activity is irregular, the sequence of switching among the
statocyst receptors is the same for those neurons that are
active at a given time window. Dotted rectangles in Fig.
24 point out this fact. The activation-sequence lock
Rabinovich et al.: Dynamical principles in neuroscience
among the statocyst receptor neurons emerges in spite
of the highly irregular timing of the switching dynamics
and is a feature that can be used for motor coordination
共Venaille et al., 2005兲.
In this example the winnerless competition is triggered by a constant excitation to all statocyst receptors
关Hi = ci; see details by Varona, Rabinovich, et al. 共2002兲兴.
Thus the stimulus has low information content. Nonetheless, the network of statocyst receptors can use this
activity to generate an information-rich signal with positive Kolmogorov-Sinai entropy. This entropy is equal to
the value of the new information encoded in the dynamical motion. The statocyst sensory network is thus
multifunctional and can generate a complex spatiotemporal pattern useful for motor coordination even when
its dynamics are not evoked by gravity, as during hunting.
2. Chaos and information transmission
To illustrate the role of chaos in information transmission, we use as an example the inferior olive 共IO兲, which
is an input system to the cerebellum. Neurons of the IO
may chaotically recode the high-frequency information
carried by its inputs into chaotic, low-rate output
共Schweighofer et al., 2004兲. The IO has been proposed as
a system that controls and coordinates different rhythms
through the intrinsic oscillatory properties of its neurons
and the nature of their electrical interconnections
共Llinás and Welsh, 1993; de Zeeuw et al., 1998兲. It has
also been implicated in motor learning 共Ito, 1982兲 and in
comparing tasks of intended and achieved movements as
a generator of error signals 共Oscarsson, 1980兲.
Experimental recordings show that IO cells are electrically coupled and display subthreshold oscillations
and spiking activity. Subthreshold oscillations have a relevant role for information processing in the context of a
system with extensive electrical coupling. In such systems the spiking activity can be propagated through the
network, and, in addition, small differences in hyperpolarized membrane potentials propagate among neighboring cells.
A modeling study suggests that electrical coupling in
IO neurons may induce chaos, which would allow
information-rich, but low-firing-rate, error signals to
reach individual Purkinje cells in the cerebellar cortex.
This would provide the cerebellar cortex with essential
information for efficient learning without disturbing ongoing motor control. The chaotic firing leads to the generation of IO spikes with different timing. Because the
IO has a low firing rate, an accurate error signal will be
available for individual Purkinje cells only after repeated trials. Electrical coupling can provide the source
of disorder that induces a chaotic resonance in the IO
network 共Schweighofer et al., 2004兲. This resonance
leads to an increase in information transmission by distributing the high-frequency components of the error inputs over the sporadic, irregular, and non-phase-locked
spikes.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1239
The IO single-neuron model consists of two compartments that include a low-threshold calcium current
共ICal兲, an anomalous inward rectifier current 共Ih兲, a
Hodgkin-Huxley-type sodium current 共INa兲, and a delayed rectifier potassium current 共IKd兲 in the somatic
compartment 共see Table I兲. The dendritic compartment
contains a calcium-activated potassium current 共IKCa兲
and a high-threshold calcium current 共ICah兲. This compartment also receives electrical connections from
neighboring neurons. Fast ionic channels are located in
the soma, and slow channels are located in the dendritic
compartment. Some of the channel conductances depend on the calcium concentration. The equations for
each compartment of a single neuron can be summarized as
CM
dV共t兲
= − 共Iion + Il + Iinj + Icomp兲,
dt
共16兲
where CM is the membrane capacitance, Il is a leak current, Iinj is the injected stimulus current, Icomp connects
the compartments, and Iion is the sum of the currents
above for each compartment. In addition, the dendritic
compartment has the electrical coupling current Iec
= gc兺i关V共t兲 − Vi共t兲兴, where the index i runs over the neighbors of each neuron, and gc is the electrical coupling
conductance.
Each IO neuron is represented by a system of ordinary differential equations 共ODEs兲, and the network is a
set of these systems coupled through the electrical coupling currents Iec. The networks examined consisted of
2 ⫻ 2, 3 ⫻ 3, and 9 ⫻ 3 neurons, where cells are connected
to their two, three, or four neighbors depending on their
positions in the grid.
This is a complex network, even when it is only 2 ⫻ 2,
and one must select an important feature of the dynamics to characterize its behavior. The largest Lyapunov
exponent of the network is a good choice as it is independent of initial conditions and tells us about information flow in the network. Figure 27 displays the largest
Lyapunov exponent for each network as a function of
the electric coupling conductance gc. We also see in Fig.
27 that the gc producing the largest Lyapunov exponent
yields the largest information transfer through the network, evaluated as the average mutual information per
spike.
In a more general framework than the IO, it is remarkable that the chaotic activity of individual neurons
unexpectedly underlies higher flexibility and, at the
same time, greater accuracy and precision in their neural
dynamics. The origin of this phenomenon is the potential ability of coupled neurons with chaotic behavior to
synchronize their activities and generate rhythms whose
period depends on the strength of the coupling or other
network parameters 关for a review see Rabinovich and
Abarbanel 共1998兲 and Aihara 共2002兲兴. Networks with
many chaotic neurons can generate interesting transient
dynamics, i.e., chaotic itinerancy 共CI兲 共Tsuda, 1991;
Rowe, 2002兲. Chaotic itinerancy results from weak instabilities in the attractors, i.e., attractor sets in whose
1240
Rabinovich et al.: Dynamical principles in neuroscience
tally hinder the power of data analyses: 共i兲 finite statistical fluctuations, 共ii兲 external noise, and 共iii兲 nonstationarity of the neural circuit activity 关see, for example, Lai
et al. 共2003兲兴.
C. Synaptic dynamics and information processing
FIG. 27. 共Color online兲 Chaotic dynamics increases information transmission in IO models. Top panel: Largest Lyapunov
exponent as a function of the electrical coupling strength gc for
different IO networks of nonidentical cells. Bottom panel: Network average mutual information per spike as a function of gc.
Modified from Schweighofer et al., 2004.
neighborhood there are trajectories that do not go to the
attractors 共Milnor-type attractors兲. A developed CI motion needs both many neurons and a very high level of
interconnections. This is in contrast to the traditional
concept of computation with attractors 共Hopfield, 1982兲.
Chaotic itinerancy yields computations with transient
trajectories; in particular, there can be motion along
separatrices as in winnerless competition dynamics 共Sec.
IV.C兲. Although CI is an interesting phenomenon, applying it to explain and predict the activity of sensory
systems 共Kay, 2003兲, and to any nonautonomous neural
circuit dynamics, poses a question that has not been answered yet: How can CI be reproducible and robust
against noise and at the same time sensitive to a stimulus?
To conclude this section it is necessary to emphasize
that the answer to the question of the functional role of
chaos in real neural systems is still unclear. In spite of
the attractiveness of such ideas as 共i兲 chaos makes neural
circuits more flexible and adaptive, 共ii兲 chaotic dynamics
create information and can help to store it 共see above兲,
and 共iii兲 the nonlinear dynamical analyses of physiological data 共e.g., electroencephalogram time series兲 can be
important for the prediction or control of pathological
neural states, it is extremely difficult to confirm these
ideas directly in in vivo or even in vitro experiments. In
particular, there are three obstacles that can fundamenRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
Synaptic transmission in many networks of the nervous system is dynamical, meaning that the magnitude
of postsynaptic responses depends on the history of
presynaptic activity 共Thompson and Deuchars, 1994;
Fuhrmann et al., 2002兲. This phenomenon is independent of 共or in addition to兲 the plasticity mechanisms of
the synapses 共discussed in Sec. II.A.3兲. The role of synapses is often considered to be the simple notification to
the postsynaptic neuron of presynaptic cell activity.
However, electrophysiological recordings show that synaptic transmission can imply activity-dependent changes
in response to presynaptic spike trains. The magnitude
of postsynaptic potentials can change rapidly from one
spike to another, depending on the particular temporal
distribution of the presynaptic signals. Thus each single
postsynaptic response can encode information about the
temporal properties of the presynaptic signals.
The magnitude of the postsynaptic response is determined by the interspike intervals of the presynaptic activity and by the probabilistic nature of neurotransmitter
release. In depressing synapses a short interval between
presynaptic spikes is followed by small postsynaptic responses, while long presynaptic interspike intervals are
followed by a large postsynaptic response. Facilitating
synapses tend to generate responses that grow with successive presynaptic spikes. In this context, several theoretical efforts have tried to explore the capacity of single
responses of dynamical synapses to encode temporal information about the timing of presynaptic events.
Theoretical models for dynamical synapses are based
on the time variation of the fraction of neurotransmitter
released from the presynaptic terminal R共t兲, 0 艋 R共t兲
艋 1. When a presynaptic spike occurs at time tsp, the
fraction U of available neurotransmitters and the recovery time constant ␶rec determine the rate of return of
resources R共t兲 to the available presynaptic pool. In a
depressing synapse, U and ␶rec are constant. A simple
model describes the fraction of synaptic resources available for transmission as 共Fuhrmann et al., 2002兲
dR共t兲 1 − R共t兲
− UR共t兲␦共t − tsp兲,
=
dt
␶rec
共17兲
and the amplitude of the postsynaptic response at time
tsp is proportional to R共tsp兲.
For a facilitating synapse, U becomes a function of
time U共t兲 increasing at each presynaptic spike and decaying to the baseline level when there is no presynaptic
activity:
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 28. Dynamical synapses imply that synaptic transmission
depends on previous presynaptic activity. This shows the average postsynaptic activity generated in response to a presynaptic spike train 共bottom trace兲 in a pyramidal neuron 共top trace兲
and in a model of a depressing synapse 共middle trace兲. Postsynaptic potential in the model is computed using a passive membrane mechanism ␶m共dV / dt兲 = −V + RiIsyn共t兲, where Ri is the input resistance. Modified from Tsodyks and Markram, 1997.
dU共t兲
U共t兲
+ U1关1 − U共t兲兴␦共t − tsp兲,
=−
dt
␶facil
共18兲
where U1 is a constant determining the step increase in
U共t兲 and ␶facil is the relaxation time constant of the facilitation.
Other approaches to modeling dynamical synapses include probabilistic models to account for fluctuations in
presynaptic release of neurotransmitters. At a synaptic
connection with N release sites we can assume that at
each site there can be, at most, one vesicle available for
release, and that the release at each site is an independent event. When a presynaptic spike is produced at
time tsp, each site containing a vesicle will release it with
the same probability U共t兲. Once a release occurs, the site
can be refilled during a time interval dt with probability
dt / ␶rec. The probabilistic release and recovery can be
described by the probability Pv共t兲 for a vesicle to be
available for release at any time t:
dPv共t兲 1 − Pv共t兲
− U共t兲Pv共t兲␦共t − tsp兲.
=
dt
␶rec
共19兲
Figure 28 shows how this formulation permits an accurate description of a depressing synapse in response to a
specified presynaptic spike train.
The transmission of sensory information from the environment to decision centers through neural communication channels requires a high degree of reliability and
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1241
sensitivity from networks of heterogeneous, inaccurate,
and sometimes unreliable components. The properties
of the channel itself, assuming the sensor is accurate,
must be richer than conventional channels studied in engineering applications. Those channels are passive and,
when of high quality, can relay inputs accurately to a
receiver. Neural communication channels are composed
of dynamically active elements capable of complex autonomous oscillations. Individually, chaotic neurons can
create information in a way similar to the study of nonlinear systems with unstable trajectories: Two states of
the system, indistinguishable because only finiteresolution observations can occur, may through the action of the instabilities of the nonlinear dynamics find
themselves in the future widely separated in state space,
and thus distinguishable. Information about different
states that was unavailable at one time may become
available at a later time.
Biological neural communication pathways are able to
recover information from a hidden coding space and to
transfer information from one time scale to another because of the intrinsic nonlinear dynamics of synapses. As
an example, we discuss a very simple neural information
channel composed of sensory input in the form of a
spike train that arrives at a model neuron and then
moves through a realistic dynamical synapse to a second
neuron where the information in the initial sensory signal is read 共Eguia et al., 2000兲. The model neurons are
four-dimensional generalizations of the HindmarshRose neuron, and a model of chemical synapse derived
from first-order kinetics is used. The four-dimensional
model neuron has a rich variety of dynamical behaviors,
including periodic bursting, chaotic bursting, continuous
spiking, and multistability. For many of these regimes,
the parameters of the chemical synapse can be tuned so
that the information about the stimulus, which is unreadable to the first neuron in the path, can be recovered by the dynamical activity of the synapse, and the
second neuron can read it 共see Fig. 29兲.
The quantitative description of this unexpected phenomenon was done by calculating the average mutual
information I共S , N1兲 between the stimulus S and the response of the first neuron N1, and I共S , N2兲 between the
stimulus and the response of the second neuron N2. The
result in the example shown in Fig. 29 is I共S , N2兲
⬎ I共S , N1兲. This result indicates how nonlinear synapses
and neurons acting as input and output systems along a
communication channel can recover information apparently hidden in earlier synaptic connections in the pathway. Here the measure of information transmission used
is the average mutual information between elements,
and because the channel is active and nonlinear, the average mutual information between the sensory source
and the final neuron may be greater than the average
mutual information found in an intermediate neuron in
the channel 共but not greater than the original information兲.
Another form of synaptic dynamics involved in information processing and especially in learning is STDP
共already discussed in Sec. II.A.3兲. Information transduc-
1242
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 29. Example of recovery of hidden information in neural
channels. A presynaptic cell receives specified input and connects to a postsynaptic cell through a dynamical synapse. Top
panel: the time series of synaptic input to the presynaptic cell
J1共t兲; middle panel: the membrane potential of the first bursting neuron X1共t兲; bottom panel: the membrane potential of the
second bursting neuron X2共t兲. Note that features of the input
hidden in the response X1共t兲 are recovered in the response
following a dynamical synapse X2共t兲 共note hyperpolarization
regions for X2兲. Modified from Eguia et al., 2000.
tion is influenced by STDP 共Chechik, 2003; Hopfield and
Brody, 2004兲, which also plays an important role in binding and synchronization.
D. Binding and synchronization
We have discussed the diversity of neuron types and
the variability of neural activity. Neural processing requires the fast interaction of many neurons in different
neural subsystems. There are several dynamical mechanisms contributing to the complex integration of information that neural systems perform. Among them, the
synchronization of neural activity is the one that has
captured the most attention. Synchronization of neural
activity is also one of the proposed solutions to a widely
discussed question in neuroscience: the binding problem, which we describe briefly in this section.
The binding problem was originally formulated as a
theoretical problem by von der Malsburg in 1981 关see a
review by von der Malsburg 共1999兲, and Roskies 共1999兲;
Singer 共1999兲兴. However, examples of binding had already been proposed by Rosenblatt 共1962兲 for the visual
system 关for a review of the binding problem in vision see
Singer 共1999兲, and Wolfe and Cave 共1999兲兴. The binding
problem is formulated as the need for a coherent representation of an object provided by the association of all
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
its features 共shape, color, location, speed, etc.兲. The association of all features or binding allows a unified perception of the object. The binding problem is a generalized task of the nervous system as it seeks to reconstruct
any total perception from its components. There are also
cognitive binding problems related to cognitive identification and memory. No doubt the binding problem, like
many other problems in biology, has multiple solutions.
These solutions are most likely implemented through
the use of dynamical mechanisms for the control of neural activity.
The most widely studied mechanism proposed to
solve the binding problem is temporal synchrony 共or
temporal correlation兲 共Singer and Gray, 1995兲. It has
been suggested by von der Malsburg and Schneider
共1986兲 that synchronization is the basis for perceptual
binding. However, there is still criticism of the temporal
binding hypothesis 共Ghose and Maunsell, 1999; Riesenhuber and Poggio, 1999兲. Obviously, neural oscillations
and synchronous signals are ubiquitous in the brain, and
neural systems can make use of these phenomena to
encode, learn, and create effective outputs. There are
several lines of experimental evidence that reveal the
use of synchronization and activity correlation for binding tasks. Figure 30 shows an example of how neural
synchronization correlates with the perceptual segmentation of a complex visual pattern into distinct, spatially
overlapping surfaces 共Castelo-Branco et al., 2000兲 共see
the figure caption for details兲. Indeed, modeling studies
show that involving time in these processes can lead to
the binding of different features. The idea is to use the
coincidence of certain events in the dynamics of different neural units for binding. Usually such dynamical
binding is represented by synchronous neurons or neurons that are in phase with an external field. However,
dynamical events such as phase or frequency variations
usually are not very reproducible and robust. As discussed in the next section, it is reasonable to hypothesize
that brain circuits displaying sequential switching of neural activity use the coincidence of this switching to
implement dynamical binding of different WLC networks.
Any spatiotemporal coding needs the temporal coordination of neural activity among different populations
of neurons to provide 共i兲 better recognition of specific
features, 共ii兲 faster processing, 共iii兲 higher information
capacity, and 共iv兲 feature binding. Neural synchronization has been observed throughout the nervous system,
particularly in sensory systems, for example, in the olfactory system 共Laurent and Davidowitz, 1994兲 and the visual system 共Gray et al., 1989兲. From the point of view of
dynamical system theory, transient synchronization is an
ideal mechanism for binding neurons into assemblies for
several reasons: 共i兲 the synchronized neurons do not necessarily have to be neighbors; 共ii兲 a synchronization
event depends on the state of the neuron and the stimulus and can be very selective, that is, neurons from the
same network can be temporal members of different cell
assemblies at different instants of time; 共iii兲 basic brain
rhythms are able to synchronize neurons responsible for
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 30. An example of binding showing dependence of synchrony on transparency conditions and receptive field 共RF兲
configuration in the cat visual cortex. 共a兲 Stimulus configuration. 共b兲 Synchronization between neurons with nonoverlapping RFs and similar directional preferences recorded from
areas A18 and PMLS of the cat visual cortex. Left, RF constellation and tuning curves; right, cross correlograms for responses to a nontransparent 共left兲 and transparent plaid 共right兲
moving in the cells’ preferred direction. Grating luminance
was asymmetric to enhance perceptual transparency. Small
dark correlograms are shift predictors. 共c兲 Synchronization between neurons with different direction preferences recorded
from A18 共polar and RF plots, left兲. Top, correlograms of responses evoked by a nontransparent 共left兲 and a transparent
共right兲 plaid moving in a direction intermediate to the cells’
preferences. Bottom, correlograms of responses evoked by a
nontransparent plaid with reversed contrast conditions 共left兲,
and by a surface defined by coherent motion of intersections
共right兲. Scale on polar plots: discharge rate in spikes per second. Scale on correlograms: abscissa, shift interval in ms, bin
width 1 ms; ordinate, number of coincidences per trial, normalized. Modified from Castelo-Branco et al., 2000.
the processing of information from different sensory inputs; and 共iv兲 the synchronization is possible even between neural oscillators with strongly different frequencies 共Rabinovich et al., 2006兲.
In early visual processing neurons that encode features of a complex visual percept are associated in funcRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1243
tional assemblies through gamma-frequency synchronization 共Engel et al., 2001兲. When sensory stimuli are
perceptually or attentionally selected, and the respective
neurons are bound together to raise their saliency, then
gamma-frequency synchronization among these neurons
is also enhanced. Gamma-mediated coupling and its
modulation by attention are not limited to the visual
system: they are also found in the auditory 共Tiitinen et
al., 1993兲 and somatosensory domains 共Desmedt and
Tomberg, 1994兲. Gamma oscillations allow visiomotor
binding between posterior and central brain regions
共Rodriguez et al., 1999兲 and are involved in short-term
memory. As a means for dynamically binding neurons
into assemblies, gamma-frequency synchronization appears to be the prime mechanism for stabilizing cortical
connections among members of a neural assembly over
time. On the other hand, neurons can increase or decrease the strength of their synaptic connections depending on the precise coincidence of their activation
共STDP兲, and gamma-frequency synchronization provides the required temporal precision.
Hatsopoulos et al. 共2003兲 and Jackson et al. 共2003兲 revealed the functional significance of neural synchronization and correlations within the motor system. Preeminent among brain actions must be the aggregation of
disparate spiking patterns to form spatially and temporally coherent neural codes that then drive alpha motor
neurons and their associated muscles. Essentially, motor
binding seems to describe exactly what motor structures
of the mammalian brain do: provide high-level coordination of simple and complex voluntary movements.
Neurons with similar functional output have an increased likelihood of exhibiting neural synchronization.
In contrast to classical synchronization 共Pikovsky
et al., 2001兲, synchronization in the CNS is always transient. The phase-space image of transient synchronization can be a saddle limit cycle in the vicinity of which
the system spends finite time. Alternatively, it can be a
limit cycle whose basin of attraction decreases in time.
In both cases the system is able to leave the synchronization region after a specific stage of processing is completed and proceed with the next task. This is a broad
area where the issues and approaches are not settled,
and thus it provides an opportunity for innovative ideas
to explain the phenomenon.
To conclude this section, we note that the functional
role of synchronization in the CNS and the importance
of spike-timing coding in general are still a subject of
debate. On the one hand, it is possible to build models
that use dynamical patterns of spikes for neural computations, e.g., representation, recognition, and decision
making. Examples of such spike-timing-based computational models have been discussed by Hopfield and
Brody 共Hopfield and Brody, 2001; Brody and Hopfield,
2003兲. In this work the authors showed, in particular,
that spike synchronization across many neurons can be
achieved in the absence of direct synaptic interactions
between neurons through phase locking to a common
underlying oscillatory potential 共like gamma oscillation;
see above兲. On the other hand, the real connections of
1244
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 31. 共Color online兲 Spontaneous spatiotemporal patterns
observed in the neocortex in vitro under the action of carbachol. Images composed of optical signals recorded by eight
detectors arranged horizontally. The optical signal from each
detector was normalized to the maximum on that detector during that period and normalized values were assigned colors
according to a linear color scale 共at the top right兲. The traces
above images 2 and 5 are optical signals from two optical detectors labeled with this color scale. The x direction of the
images represents time 共12 s兲 and the y direction of each image
represents 2.6 mm of space in cortical tissue. Note also that the
first spike had a high amplitude but propagated more slowly in
the tissue. Modified from Bao and Wu, 2003.
such theoretical models with experiments in vivo are not
established 关see also Fell et al. 共2003兲 and O’Reilly et al.
共2003兲兴.
IV. TRANSIENT DYNAMICS: GENERATION
AND PROCESSING OF SEQUENCES
A. Why sequences?
The generation and control of sequences is of crucial
importance in many aspects of animal life. Working
memory, bird songs, finding food in a labyrinth, jumping
from one stone to another on the shore—all these are
the results of sequential activity generated by the nervous system. Lashley called the problem of coordination
of constituent actions into organized sequential spatiotemporal patterns the action syntax problem 共Lashley,
1960兲. The generation of sequences is also important for
intermediate information processing as we discuss below.
The sequences can be cyclic, like many brain rhythms
and spatiotemporal patterns generated by CPGs. They
can also be irregular, like neocortical theta oscillations
共4 – 10 Hz兲 generated spontaneously in cortical networks
共Bao and Wu, 2003兲 共see Fig. 31兲. The sequences can be
finite in time like those generated by a neural circuit
under the action of external input as in sensory systems.
From a physicist’s point of view, any reproducible finite
sequence that is functionally meaningful results from the
cooperative transient dynamics of the corresponding
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
neural ensemble or individual neurons. Even brain
rhythms demonstrate transient dynamics because the
circuit’s periodic activity is modulated by nonstationary
sensory inputs or signals from the periphery. It is important to emphasize the fundamental role of inhibition in
the generation and control of sequences in the nervous
system.
In this section we concentrate on the origin of sequence generation and the mechanisms of reproducibility, sensitivity, and functional reorganization of MCs. In
the standard study of nonlinear dynamical systems, attention is focused on the long-time behavior of a system.
This is typically not the relevant question in neuroscience. Here we must address the transient responses to
a stimulus external to the neural system and must consider the short-term binding of a collection of responses,
perhaps from different sensory inputs, to facilitate action commands directed to the motor system. If you attempt to swat a fly, it cannot ask you to perform this
action many times so that it can average over your actions, allowing it to perform some standard optimal response. Few flies wanting this repetition would survive.
B. Spatially ordered networks
1. Stimulus-dependent modes
Many neural ensembles are anatomically organized as
slightly inhomogeneous excitable media. Examples of
such media are retina 共Tohya et al., 2003兲, IO network
共Leznik and Llinas, 2002兲, cortex 共Ichinohe et al., 2003兲,
and thalamocortical layers 共Contreras et al., 1996兲. All
these are neuronal lattices with chemical or electrical
connections occurring primarily between neighbors.
There are some general dynamical mechanisms of sequence generation in such spatially ordered networks.
These mechanisms are usually related to the existence of
wave modes such as those shown in Fig. 31 that are
modulated by external inputs or stimuli.
Many significant observational and modeling results
for this subject are found in the visual system. Visual
systems are organized differently for different classes of
animals. For example, the mammalian visual cortex has
several topographically organized representations of the
visual field and neurons at adjacent points in the cortex
are excited by stimuli presented at adjacent regions of
the visual field. This indicates there is a continuous mapping of the coordinates of the visual field to the coordinates of the cortex 共van Essen, 1979兲. In contrast to such
a mapping connections from the visual field to the visual
cortex in the turtle, for example, are more complex: A
local spot in the visual field activates many neurons in
the cortex but in an ordered way. As a result the excitation of the turtle visual cortex is distributed and not localized, and this suggests the temporal dynamics of several interacting membrane modes 共see Fig. 32兲. In the
mammalian cortex a moving stimulus evokes a localized
wave or wave front, while in the turtle visual cortex a
differentially moving stimulus modulates temporal interactions of the cortical modes differently and is represented by different sequential switchings between them.
1245
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 32. 共Color online兲 Sequential changing of cortical modes
in the turtle visual cortex. Comparison between the spatial organization of the cortical activity in the turtle visual system and
the normal modes of a rectangular membrane 共drum兲. From
Senseman and Robbins, 1999.
To understand the dynamics of the wave modes, i.e.,
stability, sensitivity to stimuli, dependence on neuromodulators, etc., one has to build a model that is based
on the experimental information about the possibility of
these modes maintaining the topological space structure
observed in experiments. In many similar situations one
can introduce cooperative or population variables that
can be interpreted as the amplitude of such modes depending on time. The corresponding amplitude equations are essentially the widely studied evolution equations of the dynamical theory of pattern formation
共Cross and Hohenberg, 1993; Rabinovich et al., 2000兲.
For an analysis of the wave mode dynamics of the
turtle visual cortex Senseman and Robbins 共1999兲 used
the Karhunen-Loeve decomposition and a snapshot of a
spatiotemporal pattern at time t = t0 could be represented as a weighted sum of basic modes Mi共x , y兲 with
coordinates 共x , y兲 on the image:
N
u共x,y,t 兲 = 兺 ai共t0兲Mi共x,y兲,
0
共20兲
i
where u共x , y , t兲 represents the cooperative dynamics of
these modes. The presentation of different visual
stimuli, such as spots of light at different points in the
visual field, produced spatiotemporal patterns represented by different trajectories in the phase space
a1共t兲 , a2共t兲 , . . . , an共t兲. Du et al. 共2005兲 showed that it is
possible to make a reduction in the dimensionality of the
wave modes by a second Karhunen-Loeve decomposition, which maps in some time window the trajectory in
共ai兲 space into a point in a low-dimensional space 共see
Fig. 33兲. The observed transient dynamics is similar to
the experimental results on the representation of different odors in the invertebrate olfactory system 关see Fig.
46 and Galan et al. 共2004兲兴. Nenadic et al. 共2002兲 used a
large-scale computer model of turtle visual cortex to reproduce qualitatively the features of the cortical mode
dynamics seen in these experiments.
It is remarkable that not only do spatiotemporal patterns evoked by a direct stimulus look like wave modes,
but even spontaneous activity in the sensory cortex is
well organized and very different from turbulent flow
共Arieli et al., 1996兲. This means that the common assumption about the stochastic and uncorrelated spontaRev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 33. 共Color兲 Space representation of cortical responses in
the turtle visual cortex to left, center, and right stimuli. From
Du et al., 2005.
neous activity of neighboring neurons in neural
networks 关see, for example, van Vreeswijk and Sompolinsky 共1996兲; Amit and Brunel 共1997b兲兴 is not always
correct. Local field potentials and recordings from single
neurons indicate the presence of highly synchronous ongoing activity patterns or wave modes 共see Fig. 34兲. The
spontaneous activity of a single neuron connected with
others, in principle, can be reconstructed using the
evoked patterns of network activity 共Tsodyks et al.,
1999兲.
There are some illustrative models of wave modes
that we note here. In 1977 Amari 共1977兲 found spatially
localized regions of high neural activity 共“bumps”兲 in
network models consisting of a single layer of coupled
excitatory and inhibitory rate neurons. Laing et al.
共2002兲 extended Amari’s results to a nonmonotonic connection function 共“Mexican hat” with oscillating tails兲
共shown in Fig. 35兲 and a neural layer in two spatial dimensions:
⳵u共x,y,t兲
= − u共x,y,t兲
⳵t
+
冕冕
⍀
␻共x − q,y − p兲f„u共q,p,t兲…dq dp,
共21兲
2
f共u兲 = 2e−␶/共u − th兲 ⌰共u − th兲,
␻共x,y兲 = e−b
冑x2+y2
共22兲
关b sin共冑x2 + y2兲 + cos共冑x2 + y2兲兴.
共23兲
An example of a typical localized mode in such neural
media with local excitation and long-range inhibition is
represented in Fig. 36. Different modes 共with different
numbers of bumps兲 can be switched from one to another
by transient external stimuli. Multiple items can be
stored in this model because of the oscillating tails of the
1246
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 34. 共Color online兲 Relation between the spiking activity of a single neuron and the population state of cortical networks. 共a兲
From bottom to top: stimulus time course; correlation coefficient of the instantaneous snapshot of population activity with the
spatial pattern obtained by averaging over all patterns observed at the times corresponding to spikes evoked by the optimal
orientation of the stimulus called the neuron’s preferred cortical state 共PCS兲 pattern; observed spike train of evoked activity with
the optimal orientation for that neuron; reconstructed spike train. The similarity between the reconstructed and observed spike
trains is evident. Also, strong upswings in the values of correlation coefficients are evident each time the neuron emits bursts of
action potentials. Every strong burst is followed by a marked downswing in the values of the correlation coefficients. 共b兲 The same
as 共a兲, but for a spontaneous activity recording session from the same neuron 共eyes closed兲. 共c兲 The neuron’s PCS, calculated during
evoked activity and used to obtain both 共a兲 and 共b兲. 共d兲 The cortical state corresponding to spontaneous action potentials. The two
patterns are nearly identical 共correlation coefficient 0.81兲. 共e兲 and 共f兲 Another example of the similarity between the neuron’s PCS
共e兲 and the cortical state corresponding to spontaneous activity 共f兲 from a different cat obtained with the high-resolution imaging
system 共correlation coefficient 0.74兲. Modified from Tsodyks et al., 1999.
effective connection strength. This is the result of the
common activity of the excitatory and inhibitory connections between neurons. Inhibition plays a crucial role for
the stability of localized modes 共Laing et al., 2002兲.
Localized modes with different numbers of bumps remind one of complex localized patterns in a dissipative
nonequilibrium media 共Rabinovich et al., 2000兲. Based
on this analogy, it is reasonable to hypothesize that different modes may coexist in a neural layer and their
interaction and annihilation can explain the sequential
effectiveness of the different events. This suggests they
could be a model of sequential working memory 共see
below兲.
Many rhythms of the brain can take the form of
waves: spindle waves 共7 – 14 Hz兲 seen at the onset of
sleep 共Kim et al., 1995兲, slower delta rhythms of deeper
sleep, the synchronous discharge during an epileptic sei-
FIG. 35. Connection function ␻共x , y兲, centered at the center of
the domain. Modified from Laing et al., 2002.
FIG. 36. Six-bump stable solution of the model 共21兲–共23兲: b
= 0.45, ␶ = 0.1, th= 1.5. Modified from Laing et al., 2002.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
Rabinovich et al.: Dynamical principles in neuroscience
zure 共Connors and Amitai, 1997兲, waves of excitation
associated with sensory processing, 40-Hz oscillations,
and others. In thalamocortical networks the same clusters of neurons are responsible for different modes of
rhythmic activity. What is the dynamical origin of such
multifunctionality? There is no unique answer to this
question, and there are several different mechanisms
that can be responsible for it 共we have already discussed
this for small invertebrate networks; see Sec. II.B兲. Terman et al. 共1996兲 studied the transition between spindling and delta sleep rhythms. The authors showed that
these two rhythms make different uses of the fast inhibition and slow inhibition generated by thalamic reticularis cells. These two types of inhibition are mediated in
the cortex by GABA共A兲 and GABA共B兲 receptors, respectively 共Schutter, 2002; Tams et al., 2003兲.
The wave mode equation discussed above is familiar
to physicists and can be written both when interactions
between neuron populations are homogeneous and isotropic 共Ermentrout, 1998兲 and when the neural layer is
partitioned into domains or hypercolumns like the primary visual cortex 共V1兲 of cats and primates, which has a
crystallinelike structure at the millimeter length scale
共Bressloff, 2002; Bressloff and Cowan, 2002兲.
In the next section we discuss the propagation of patterns of synchronous activity along spatially ordered
neural networks.
2. Localized synfire waves
Auditory and visual sensory systems have a very high
temporal resolution. For example, the retina is able to
resolve sequential temporal patterns with a precision in
the millisecond range. Does the transmission of sensory
information from the periphery to the cortex maintain
such high resolution? If the answer is yes, what are the
dynamical mechanisms responsible for this? These questions are still open.
There are several neurophysiological experiments that
show the ability of neural systems to transmit temporarily modulated responses of sensory networks with
high precision over several processing levels. For example, cross correlations between simultaneously recorded responses of retinal cells relay neurons within
the thalamus, and cortical neurons show that the oscillatory patterning is reliably transmitted to the cortex with
a resolution in the millisecond range 关see for reviews
Singer 共1999兲 and Nase et al. 共2003兲兴. A similar phenomenon was observed by Kimpo et al. 共2003兲 who showed
evidence for the preserved timing of spiking activity
through multiple steps of a neural control loop in the
bird brain. The dynamical origin of such precise message
propagation, independent of the rate fluctuation, is often
attributed to synchronization of the many neurons in the
overall circuit 共Abeles, 1991; Diemann et al., 1999兲.
We now discuss briefly the dynamics of waves of synchronous neural firing, i.e., synfire waves. One modeling
study 共Diesmann et al., 1999兲 has shown that the stable
propagation of localized synfire waves, short-lasting synchronous spiking activity, is possible along a sequence of
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1247
FIG. 37. Sequence of pools of excitatory neurons, connected in
a feedforward way by so-called divergent and convergent connections. The network is called a synfire chain if it supports the
propagation of synchronous spike patterns. Modified from Gewaltig et al., 2001.
layers or pools of neurons in a feedforward cortical network such as the one shown in Fig. 37, a synfire chain
共Abeles, 1991兲. The degree of temporal accuracy of
spike times among the pools’ members determines
whether subsequent pools can reproduce 共or even improve兲 this accuracy 关Fig. 38共a兲兴, or whether synchronous
excitation disperses and eventually dies out as in Fig.
38共b兲 for a smaller number of spikes in the volley. Thus
in the context of synfire network function the quality of
timing is judged on whether synchronous spiking is sustained or whether it dies out.
Diesmann et al. 共1999兲, Cateau and Fukai 共2001兲, Kistler and de Zeeuw 共2002兲, and Nowotny and Huerta
共2003兲 have shown that if the pool size is more than a
critical value determined by the connectivity between
layers, the wave activity initiated at the first pool propagates from one pool to the next, forming a synfire wave.
Nowotny and Huerta 共2003兲 have theoretically proven
that no other states exist beyond synchronized or unsynchronized volleys as shown in the experiments by Reyes
共2003兲.
The synfire feedforward chain 共Fig. 37兲 is an oversimplified model for analyzing synfire waves because in reality any network with synfire chains is embedded in a
larger cortical network that also has inhibitory neurons
FIG. 38. Propagation of firing activity in synfire chains. 共a兲
Stable and 共b兲 unstable propagation of synchronous spiking in
a model of cortical networks. Raster displays of propagating
spike volley along fully connected synfire chain. Panels show
the spikes in ten successive groups of 100 neurons each 共synaptic delays arbitrarily set to 5 ms兲. Initial spike volley 共not
shown兲 was fully synchronized, containing 共a兲 50 or 共b兲 48
spikes. Modified from Diesmann et al., 1999.
1248
Rabinovich et al.: Dynamical principles in neuroscience
and many recurrent connections. This problem is discussed in detail by Aviel et al. 共2003兲.
C. Winnerless competition principle
1. Stimulus-dependent competition
Here we consider a paradigm of sequence generation
that does not depend on the geometrical structure of the
neural ensemble in physical space. It can, for example,
be a two-dimensional layer with connections between
neighbors or a three-dimensional network with sparse
random connections. This paradigm can be helpful for
the explanation and prediction of many dynamical phenomena in neural networks with excitatory and inhibitory synaptic connections. The paradigm is called the
winnerless competition principle. We have touched on
aspects of WLC networks earlier, and here we expand
on their properties and their possible use in neuroscience.
“Survival of the fittest” is a cliché that is often associated with the term competition. However, competition is
not merely a means of determining the winner, as in a
winner-take-all network. It is also a multifunctional instrument that nature uses at all levels of the neuronal
hierarchy. Competition is also a mechanism that maintains the highest level of variability and stability of neural dynamics, even if it is a transient behavior.
Over two hundred years ago the mathematicians
Borda and de Condorcet were interested in the process
of plurality elections at the French Royal Academy of
Sciences. They considered voting dynamics in a case of
three candidates A, B, and C. If A beats B and B beats
C in a head-to-head competition, we might reasonably
expect A to beat C. Thus predicting the results of the
election is easy. However, this is not always the case. It
may happen that C beats A, resulting in a so-called Condorcet triangle, and there is no real winner in such a
competitive process 共Borda, 1781; Saari, 1995兲. This example is also called a “voting paradox.” The dynamical
image of this phenomenon is a robust heteroclinic cycle
共see Fig. 39兲. In some specific cases the heteroclinic cycle
is even structurally stable 共Guckenheimer and Holmes,
1988; Krupa, 1997; Stone and Armbruster, 1999; Ashwin
et al., 2003; Postlethwaite and Dawes, 2005兲.
The competition without a winner is also known in
hydrodynamics: Busse and Heikes discovered that convective roll patterns in a rotating plane layer exhibit sequential changes of the roll’s direction as a result of the
competition between patterns with different roll orientations. No pattern becomes a winner and the system
exhibits periodic or chaotic switching dynamics 共Busse
and Heikes, 1980兲. For review see Rabinovich et al.
共2000兲. The same phenomenon has been discovered in a
genetic system, i.e., in experiments with a synthetic network of three transcriptional regulators 共Elowitz and
Leibler, 2000兲. Specifically, these authors described three
repressor genes A, B, and C organized in a closed chain
with unidirectional inhibitory connections such that A,
B, and C beat each other. This network behaves like a
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 39. 共Color online兲 Illustration of WLC dynamics. Top
panel: Phase portrait corresponding to the autonomous WLC
dynamics of a three-dimensional case. Bottom panel: Projection of a nine-dimensional heteroclinic orbit of three inhibitory
coupled FitzHugh-Nagumo spiking neurons in a threedimensional space 共the variables ␰1, ␰3, ␰3 are linear combinations of the actual phase variables of the system兲. From
Rabinovich et al., 2001.
clock: it periodically induces synthesis of green fluorescent proteins as an indicator of the state of individual
cells on a time scale of hours.
In neural systems such clock competitive dynamics
can result from the inhibitory connections among neurons. For example, Jefferys et al. 共1996兲 showed that hippocampal and neocortical networks of mutually inhibitory interneurons generate collective 40-Hz rhythms
共gamma oscillations兲 when excited tonically. Another example of neural competition without a winner was discussed by Ermentrout 共1992兲. The author studied the
dynamics of a single inhibitory neuron connected to a
small cluster of loosely coupled excitatory cells and observed the emergence of a limit cycle through a heteroclinic cycle. For autonomous dynamical systems competition without a winner is a well-known phenomenon.
We use the term WLC principle for the nonautonomous transient dynamics of neural systems receiving external stimuli and exhibiting sequential switching among
temporal winners. The main point of the WLC principle
is the transformation of incoming inputs into spatiotemporal outputs based on the intrinsic switching dynamics
of the neuronal ensemble 共see Fig. 40兲. In the phase
space of the network, such switching dynamics are represented by a heteroclinic sequence whose architecture
depends on the stimulus. Such a sequence consists of
1249
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 40. Transformation of the identity spatial input into spatiotemporal output based on the intrinsic sequential dynamics
of a neural ensemble with WLC.
many saddle equilibria or saddle cycles and many heteroclinic orbits connecting them, i.e., many separatrices.
The sequence can serve as an attracting set if every
semistable set has only one unstable direction 关see also
Ashwin and Timme 共2005兲兴.
The key points on which WLC networks are based are
the following: 共i兲 the stimulus-dependent heteroclinic sequence corresponding to a specific order of switching
has a large basin of attraction, i.e., the sequence is robust; and 共ii兲 the topology of the heteroclinic sequence
sensitively depends on the incoming signals, i.e., WLC
dynamics have a high resolution.
In this manner stimulus-dependent sequential switching of neurons or groups of neurons 共clusters兲 is able to
resolve the fundamental contradiction between sensitivity and robustness in sensory recognition. Any kind of
sequential activity can be programmed, in principle, by a
network with stimulus-dependent nonsymmetric inhibitory connections. It can be the creation of spatiotemporal patterns of motor activity, the transformation of the
spatial information into spatiotemporal information for
successful recognition 共see Fig. 40兲, and many other
computations.
The generation of sequences in inhibitory networks
has already been discussed when we analyzed the dynamics of CPGs 共see Sec. II.B兲 focusing on rhythmic
activity. The mathematical image in phase space of the
rhythmic sequential switching shown in Figs. 8 and 9 is a
limit cycle in the vicinity of the heteroclinic contour 关cf.
Fig. 39共a兲兴.
WLC dynamics can be described in the framework of
neural models at different levels. These could be rate
models, Hodgkin-Huxley-type models, or even simple
map models 共see Table I兲. For spiking neurons or groups
of synchronized spiking neurons in a network with nonsymmetrical lateral inhibition WLC may lead to switching between active and inactive states. The mathematical image of such switching activity is also a heteroclinic
loop, but in this case the separatrices do not connect
saddle equilibrium points 关Fig. 39共a兲兴 but saddle limit
cycles as shown in Fig. 39共b兲. The WLC dynamics in a
model network of nine spiking neurons with inhibitory
connections is shown in Fig. 41. Similar results based on
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 41. Spatiotemporal patterns generated by a network of
nine FitzHugh-Nagumo neurons with inhibitory connections.
The left and right panels correspond to two different stimuli.
From Rabinovich et al., 2001.
a map model of neurons have been reported by Casado
共2003兲.
An important advantage of WLC networks is that
they can produce different spatiotemporal patterns in
response to different stimuli, and, remarkably, neurons
spontaneously form synchronized clusters despite the
absence of excitatory synaptic connections. For a discussion of synchronization with inhibition see also van
Vreeswijk et al. 共1994兲 and Elson et al. 共2002兲.
Finally WLC networks also possess a strikingly different capacity or ability to represent in a distinguishable
manner a number of different patterns. In an attractor
computation network of the Hopfield variety, a network
with N attractors has been shown to have a capacity of
approximately N / 7. In a simple WLC network with N
nodes, this capacity has been shown 共Rabinovich et al.,
2001兲 to be of order e共N − 1兲!, which is a remarkable gain
in capacity.
2. Self-organized WLC networks
It is generally accepted that there is insufficient genetic information available to account for all the synaptic connectivity in the brain. How then can the functional architecture of WLC circuits be generated in the
process of development?
One possible answer has been found by Huerta and
Rabinovich. Starting with a model circuit consisting of
100 rate model neurons connected randomly with weak
inhibitory synapses, new synaptic strengths are computed for the connections using Hebb learning rules in
the presence of weak noise. The neuron rates ai共t兲 satisfy
a Lotka-Volterra model familiar from our earlier discussion. In this case the matrix ␳ij共t兲 is a dynamical variable:
冉
冊
dai共t兲
= ai共t兲 ␴共S兲 − 兺 ␳ij共t兲aj共t兲 + ␰i共t兲.
dt
j
共24兲
␴共S兲 is a function dependent on the stimulus S, ␳ij共t兲 are
the strengths of the inhibitory connections determined
1250
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 42. 共Color兲 Result of simulating a network of 100 neurons
subject to the learning rule g共ai , aj兲 = aiaj关10 tanh共aj − ai兲 + 1兴.
The activity of representative neurons in this network is shown
in different colors. The system starts from random initial conditions for the connections. The noise level is ␩ = 0.01. For simplicity, the switching activity of only four of the 100 neurons is
shown.
by some learning rules, and ␰i共t兲 is Gaussian noise with
具␰i共t兲␰j共t⬘兲典 = ␩␦ij␦共t − t⬘兲. The learning is described by the
equations
d␳ij共t兲
= ␳ij共t兲g„ai共t兲,aj共t兲,S… − 关␳ij共t兲 − ␥兴,
dt
共25兲
where g共ai , aj , S兲 represents the strengthening of interactions from neuron i to neuron j as a function of the
external stimulus S. The parameter ␥ represents the
lower bound of the coupling strengths among neurons.
Figure 42 shows the activity of representative neurons in
a network built with this model. After the selforganization phase, this network displays WLC switching dynamics.
Winnerless competition dynamics can also be the result of local self-organization in networks of HH model
neurons that display STDP with inhibitory synaptic connections as shown in Fig. 43. Such mechanisms of selforganization, as shown by Nowotny and Rabinovich, can
be appropriate for networks that generate not only
rhythmic activity but also transient heteroclinic sequences.
3. Stable heteroclinic sequence
The phase-space image of nonrhythmic WLC dynamics is a trajectory in the vicinity of a stable heteroclinic
sequence 共SHS兲 in the state space of the system. Such a
sequence 共see Fig. 44兲 is an open chain of saddle fixed
points connected by one-dimensional separatrices which
retain nearby trajectories in its vicinity. The flexibility of
WLC dynamics is provided by their dependence on the
identity of participating neural clusters of stimuli. Sequence generation in chainlike or layerlike networks of
neurons may result from a feedforward wavelike propagation of spikes like waves in synfire chains 共see above兲.
In contrast, WLC dynamics does not need a specific spatial organization of the network. However, the image of
a wave is a useful one, because in the case of WLC a
wave of neural activity propagates in state space along
the SHS. Such a wave is initiated by a stimulus. The
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
FIG. 43. 共Color online兲 Example of WLC dynamics entrained
in a network by a local learning rule. In isolation, the four HH
neurons in the network are rebound bursters, i.e., they fire a
brief burst of spikes after being strongly inhibited. The all-toall inhibitory synapses in the small network are governed by a
STDP learning rule which strengthens the synapse for positive
time delays between postsynaptic and presynaptic activity and
weakens it otherwise. Such STDP of inhibitory synapses has
been observed in the entorhinal cortex of rats 共Haas et al.,
2006兲. 共a兲 Before entrainment the neurons just follow the input
signal of periodic current pulses. 共b兲 The resulting bursts
strengthen the forward synapses corresponding to the input
sequence making them eventually strong enough to cause rebound bursts. 共c兲 After entrainment activating any one of the
neurons leads to an infinite repetition of the trained sequence
carried by the successive rebound bursts of the neurons.
speed of the sequential switching depends on the noise
level ␩. Noise controls the distance between trajectories
realized by the system and the SHS. For trajectories that
get closer to the SHS the time that the system spends
near semistable states 共saddles兲, i.e., the interval between switching, becomes longer 共see Fig. 44兲.
The mechanism of reproducing transient sequential
neural activity has been analyzed by Aframovich, Zhigulin, et al. 共2004兲 共see Fig. 44兲. It is quite general and does
not depend on the details of the neuronal model. Saddle
points in the phase space of the neural network can be
replaced by saddle limit cycles or even chaotic sets that
describe neural activity in more detail, as in typical spiking or spiking-bursting models. This feature is important
for neural modeling because it may help to build a
bridge between the concepts of neural competition and
synchronization of spikes.
We can formulate the necessary conditions for the
connectivity of a WLC network that must be satisfied in
order for the network to exhibit reproducible sequential
dynamics along the heteroclinic chain. As before, we
base our discussion on the rate model
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 44. A stable open heteroclinic sequence in a neural circuit with WLC. Wsi is a stable manifold of the ith saddle fixed
point 共heavy dots兲. The trajectories in the vicinity of the SHS
represent sequences with different timings. The time intervals
between switches is proportional to T ⬃ 兩ln ␩兩 / ␭u, where ␭u is a
positive Lyapunov exponent that characterizes the onedimensional unstable separatrices of the saddle points 共Stone
and Holmes, 1990兲. Modified from Afraimovich, Zhigulin,
et al., 2004.
冉
冊
Nl
ai共t兲
= ai共t兲 ␴i共Sជ l兲 − 兺 ␳ij共Sជ l兲aj共t兲 + ␰i共t兲,
dt
j
共26兲
where ␰i共t兲 is an external Gaussian noise. In this model it
is assumed that the stimulus Sជ l influences the matrix ␳
ij
and increments ␴i only in the subnetwork Nl. Each increment ␴i controls the time constant of an initial exponential growth from the resting state ai共t兲 = 0. As shown
by Aframovich, Zhigulin, et al. 共2004兲 to assure that the
SHS is in the phase space of the system 共26兲 the following inequalities must be satisfied:
␴ik−1
␴ ik
␴ik+1
␴ ik
⬍ ␳ik−1ik ⬍
␴ik−1
␴ ik
− 1 ⬍ ␳ik+1ik ⬍
␳iik ⬎ ␳ik−1ik +
+ 1,
␴ik+1
␴ ik
␴i − ␴ik−1
␴ ik
,
.
共27兲
共28兲
共29兲
␴im is the increment of the mth saddle whose unstable
manifold is one dimensional; ␳ik±1ik is the strength of the
inhibitory connection between neighboring saddles in
the heteroclinic chain. The computer modeling result of
a network with parameters that satisfy 共27兲–共29兲 is
shown in Fig. 45.
In the next section we discuss some experiments that
support the SHS paradigm.
4. Relation to experiments
The olfactory system may serve as one example of a
neural system that generates transient, but trial-to-trial
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1251
FIG. 45. 共Color online兲 Time series of the activity of a WLC
network during ten trials 共only 20 neurons are shown兲: simulations of each trial were started from a different random initial
condition. In this plot each neuron is represented by a different color and its level of activity by the saturation of the color.
From Afraimovich, Zhigulin, et al., 2004.
reproducible, sequences of neuronal activity which can
be explained with the WLC principle. The complex intrinsic dynamics in the antennal lobe 共AL兲 of insects
transform static sensory stimuli into spatiotemporal patterns of neural activity 共Laurent et al., 2001兲. Several
experimental results about the reproducibility of the
transient spatiotemporal AL dynamics have been published 共Stopfer et al., 2003; Galan et al., 2004; Mazor and
Laurent, 2005兲 共see Fig. 46兲. In experiments described by
Galan et al. 共2004兲 bees were presented with different
odors, and neural activity in the AL was recorded using
calcium imaging. The authors analyzed the transient trajectories in the projection neuron activity space and
found that trajectories representing different trials of
stimulation with the same odor were very similar. It was
shown that after a time interval of about 800 ms different odors are represented in phase space by different
static attractors, i.e., the transient spatiotemporal patterns converge to different spatial patterns of activity.
However, the authors emphasize that due to the reproducibility of the transient dynamics some odors were
recognized in the early transient stage as soon as 300 ms
after the onset of the odor presentation. It is highly
likely that the transient trajectories observed in these
experiments represent realizations of a SHS.
The generation of reproducible sequences plays also a
key role in the high vocal center 共HVC兲 of the songbird
system 共Hahnloser et al., 2002兲. Like a CPG, this neural
system is able to generate sparse spatiotemporal patterns without any rhythmic stimuli in vitro 共Solis and
Perkel, 2005兲. In its projections to the premotor nucleus
RA, HVC in an awake singing bird sends sparse bursts
of high-frequency signals once for each syllable of the
song. These bursts have an interspike interval about
2 ms and last about 8 ms within a syllable time scale of
100– 200 ms. The bursts are shown for several HVC
→ RA projection neurons in Fig. 47. The HVC also contains many inhibitory interneurons 共Mooney and
Prather, 2005兲. The interneurons burst densely throughout the vocalizations, in contrast to the bursting of the
RA-projecting HVC neurons at single precise timings. A
1252
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 47. 共Color online兲 HVC songbird patterns. Spike raster
plot of ten HVC共RA兲 neurons recorded in one bird during
singing. Each row of tick marks shows spikes generated during
one rendition of the song or call; roughly ten renditions are
shown for each neuron. Modified from Hahnloser et al., 2002.
FIG. 46. 共Color兲 Transient AL dynamics. Top panel: Trajectories of the antennal lobe activity during poststimulus relaxation
in one bee. Modified from Galan et al., 2004. Bottom panel:
Visualization of trajectories representing the response of a PN
population in a locust AL over time. Time-slice points were
calculated from 110 PN responses to four concentrations 共0.01,
0.05, 0.1, 1兲 of three odors, projected onto three dimensions
using locally linear embedding, an algorithm that computes
low-dimensional, neighborhood-preserving embeddings of
high-dimensional inputs 共Roweis and Saul, 2000兲. Modified
from Stopfer et al., 2003.
plausible hypothesis is that HVC’s synaptic connections
are nonsymmetric and WLC can be a mechanism of the
neural spatiotemporal pattern generation of the song.
This would provide a basis for the reproducible patterned output from the HVC when it receives a song
command stimulus.
D. Sequence learning
Sequence learning and memory as sequence generation require temporal asymmetry in the system. Such
asymmetry can result from specific properties of the network connections, in particular, asymmetry of the connections, or can result from temporal asymmetry in the
dynamical features of individual neurons and synapses,
or both. The specific dynamical mechanisms of sequence
learning depend on the time scale of the sequence that
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
this neural system needs to learn. Learning of fast sequences, 20– 30 ms and faster, needs precise synchronization of the spikes or phases of neural waves. One possible mechanism for this can be the learning of synfire
waves. For slow sequences, like autonomous repetitive
behavior, it would be preferable to learn relevant behavioral events that typically occur on the time scale of hundreds of milliseconds or slower and the switching 共transitions兲 between them. Networks whose dynamics are
based on WLC are able to do such a job. We consider
here slow sequence learning and spatial sequential
memory 共SSM兲.
The idea is that sequential memory is encoded in a
multidimensional dynamical system with a SHS. Each of
the saddle points represents an event in a sequence to be
remembered. Once the state of the system approaches
one fixed point representing a certain event, it is drawn
along an unstable separatrix toward the next fixed point,
and the mechanism repeats itself. The necessary connections are formed in the learning phases by different sensory inputs originated by sequential events.
Seliger et al. 共2003兲 have discussed a model of the SSM
in the hippocampus. It is well accepted that the hippocampus plays the central role in acquisition and processing information related to representing motion in physical space. The most spectacular manifestation of this
role is the existence of so-called place cells which repeatedly fire when an animal is in a certain spatial location
共O’Keefe and Dostrovsky, 1971兲. Experimental research
also favors an alternative concept of spatial memory
based on a linked collection of stored episodes 共Wilson
and McNaughton, 1993兲. Each episode comprises a sequence of events, which, besides spatial locations, may
include other features of the environment 共orientation,
odor, sound, etc.兲. It is plausible to describe the corresponding learning with a population model that represents neural activity by rate coding. Seliger et al. 共2003兲
have proposed a two-layer dynamical model of SSM that
can answer the following key questions: 共i兲 How is a
certain event, e.g., an image of the environment, recorded in the structure of the synaptic connections be-
Rabinovich et al.: Dynamical principles in neuroscience
tween multiple sensory neurons 共SNs兲 and a single principal neuron 共PN兲 during learning? 共ii兲 What kind of
cooperative dynamics forces individual PNs to fire sequentially, in a way that would correspond to a specific
sequence of snapshots of the environment? 共iii兲 How
complex should this network be in order to store a certain number of different episodes without mixing different events or storing spurious episodes?
The two-layer structure of the SSM model is reminiscent of the projection network implementation of the
normal form projection algorithm 共NFPA兲; see Baird
and Eeckman 共1993兲. In the NFPA model, the dynamics
of the network is cast in terms of normal form equations
which are written for amplitudes of certain normal
forms corresponding to different patterns stored in the
system. The normal form dynamics can be chosen to
follow certain dynamical rules. Baird and Eeckman
共1993兲 have shown that a Hopfield-type network with
improved capacity can be built using this approach. Furthermore, it has been suggested 共Baird and Eeckman,
1993兲 that specific choices of the coupling matrix for the
normal form dynamics can lead to multistability among
more complex attracting sets than simple fixed points,
such as limit cycles or even chaotic attractors. For example, quasiperiodic oscillations can be described by a
normal form that corresponds to a multiple Hopf bifurcation 共Guckenheimer and Holmes, 1986兲. As shown below, a model of SSM after learning is completed can be
viewed as a variant of the NFPA with a specific choice of
normal form dynamics corresponding to winnerless
competition among different patterns.
To illustrate these ideas consider a two-level network
of Ns SNs 关xi共t兲兴 and Np principal neurons 关ai共t兲兴. One
can reasonably assume that sensory neurons do not have
their own complex dynamics and are slaved either to
external stimuli in the learning or storing regime or to
the PNs in the retrieval regime. In the learning regime,
xi共t兲 is a binary input pattern consisting of 0’s and 1’s.
Np
During the retrieval phase, xi共t兲 = 兺j=1
Pijaj共t兲, where Pij is
the Ns ⫻ Np projection matrix of connections among SNs
and PNs.
The PNs are driven by SNs during the learning phase,
but they also have their own dynamics controlled by inhibitory interconnections. When learning is complete,
the direct driving from SNs is disconnected. The equations for the PN rates ai共t兲 read
N
N
s
p
dai共t兲
= ai共t兲 − ai共t兲 兺 Vijaj共t兲 + ␣ai 兺 PTij xj共t兲 + ␰共t兲,
dt
j=1
j=1
共30兲
where ␣ ⫽ 0 in the learning phase and ␣ = 0 in the retrieval phase, and PTij is the projection matrix. The coupling between SNs and PNs is bidirectional. The last
term on the right-hand side of Eq. 共30兲 represents small
positive external perturbations which can input signals
from other parts of the brain that control learning and
retrieval dynamics.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1253
After a certain pattern is presented to the model, the
sensory stimuli reset the state of the PN layer according
NS T
to the projection rule ai共t兲 = 兺j=1
Pij xj共t兲, but ai共t兲 change
according to Eq. 共30兲.
The dynamics of SNs and PNs during the learning and
retrieval phases have two learning processes: 共i兲 forming
the projection matrix Pij which is responsible for connecting a group of sensory neurons of the first layer corresponding to a certain stored pattern to a single PN
which represents this pattern at the PN level; and 共ii兲
learning of the competition matrix Vij which is responsible for the temporal 共logical兲 ordering of the sequential
memory.
The slow learning dynamics of the projection matrix is
controlled by the following equation:
Ṗij = ⑀ai共␤xj − Pij兲
共31兲
with ⑀ Ⰶ 1. We assume that initially all Pij connections
are nearly identical Pij = 1 + ␩ij, where ␩ij are small random perturbations, 兺j␩ij = 0, 具␩2ij典 = ␩20 Ⰶ 1. Additionally,
we assume that initially the matrix Vij is purely competitive: Vii = 1 and Vij = V0 ⬎ 1 for i ⫽ j.
Suppose we want to memorize a certain pattern A in
our projection matrix. We apply a set of inputs Ai corresponding to the pattern A of the SNs. As before, we
assume that external stimuli render the SNs in one of
two states: excited, Ai = 1, and quiescent, Ai = 0. The initial state of the PN layer is fully excited: ai共0兲 = 兺jPijAj.
According to the competitive nature of interactions between PNs after a short transient, only one of them, the
neuron A which corresponds to the maximum ai共0兲, remains excited and the others become quiescent. Which
neuron becomes responsible for the pattern A is actually
random, as it depends on the initial projection matrix
Pij. It follows from Eq. 共31兲 that for small ⑀ synapses of
suppressed PNs do not change, whereas synapses of the
single excited neuron evolve such that connections between excited SNs and PNs neurons amplify toward ␤
⬎ 1, and connections between excited PNs and quiescent
SNs decay to zero. As a result, the first input pattern will
be recorded in one of the matrix Pij rows, while other
rows will remain almost unchanged. Now suppose that
we want to record a second pattern different from the
first one. We can repeat the procedure described above,
namely, apply external stimuli associated with pattern B
to the SNs, project them to the initial state of the PN
layer, ai共0兲 = 兺jPijBj, and let the system evolve. Since synaptic connections from SNs suppressed by the first pattern to neuron A have been eliminated, a new set of
stimuli corresponding to pattern B will excite neuron A
more weakly than most of the others, and competition
will lead to selection of one PN B different from neuron
A. In this way we can record as many patterns as there
are PNs.
The sequential order of the patterns recorded in the
projection network is determined by the competition
matrix Vij, Eq. 共30兲. Initially it is set to Vij = V0 ⬎ 1 for i
⫽ j and Vii = 1 which corresponds to winner-take-all competition. The goal of sequential spatial learning is to
1254
Rabinovich et al.: Dynamical principles in neuroscience
E. Sequences in complex systems with random connections
FIG. 48. Amplitudes of principal neurons during the memory
retrieval phase in a two-layer dynamical model of sequential
spatial memory. 共a兲 Periodic retrieval, two different test patterns presented; 共b兲 aperiodic retrieval with modulated inhibition 共see text兲. Modified from Seliger et al., 2003.
record the transition of pattern A to pattern B in the
form of suppressing the competition matrix element
VBA. We suppose that the slow dynamics of the nondiagonal elements of the competition matrix are controlled by the delay-differential equation
V̇ij = ⑀ai共t兲aj共t − ␶兲共V1 − Vij兲,
共32兲
where ␶ is constant. Equation 共32兲 shows that only the
matrix elements corresponding to ai共t兲 ⫽ 0 and aj共t − ␶兲
⫽ 0 are changing toward the asymptotic value V1 ⬍ 1 corresponding to the desired transition. Since most of the
time, except for short transients, only one PN is excited,
only one of the connections Vij is changing at any time.
As a result, an arbitrary, nonrepeating, sequence of patterns can be recorded.
When a test pattern T is presented to the sensory
layer, xi共0兲 = T共i兲 ai共0兲 = 兺iPijTTj, and T resembles one of
the recorded patterns, this will initiate a periodic sequence of patterns corresponding to the previously recorded sequence in the network. Figure 48 shows the
behavior of principal neurons after different initial patterns resembling different digits have been presented. In
both cases, the system quickly settles onto a cyclic generation of patterns associated with a given test pattern.
At any given time, except for a short transient time between patterns, only a single PN is on, corresponding to
a particular pattern.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
The level of cellular and network complexity in the
nervous system leads one to ask: How do evolution and
genetics build a complex brain? Comparative studies of
the neocortex indicate that early mammalian neocortices were composed of only a few cortical fields and in
primates the neocortex expanded dramatically; the number of cortical fields increased and the connectivity between them became very complex. The architecture of
the microcircuitry of the mammalian neocortex remains
largely unknown in terms of cell-to-cell connections;
however, the connections of groups of neurons with
other groups are becoming better understood thanks to
new anatomical techniques and the use of slice techniques. Many parts of the neocortex developed under
strict genetic control as precise networks with connections that appear similar from animal to animal. Kozloski et al. 共2001兲 discussed visual networks in this context. However, the local connectivity can be probabilistic
or random as a consequence of experience-dependent
plasticity and self-organization 共Chklovskii et al., 2004兲.
In particular, the imaging of individual pyramidal neurons in the mouse barrel cortex over a period of weeks
共Maravall et al., 2004兲 showed that sensory experience
drives the formation and elimination of synapses and
that these changes might underlie adaptive remodeling
of neural circuits.
Thus the brain appears as a compromise between existing genetic constraints and the need to adapt, i.e., networks are formed by both genetics and activitydependent or self-organizing mechanisms. This makes it
very difficult to determine the principles of network architecture and to build reasonable dynamical models
that are able to predict the reactions of a complex neural
system to changes in the environment; we have to take
into account that even self-organized networks are under genetic control but in a different sense. For example,
genetics can control the average balance between excitatory and inhibitory synaptic connections, sparseness of
the connections, etc. The point of view that the infant
cortex is not a completely organized machine is based on
the supposition that there is insufficient storage capacity
in the DNA to control every neuron and every synapse.
This idea was formulated first by Alan Turing in 1948
共Ince, 1992兲.
A simple calculation reveals that the total size of the
human genome can specify the connectivity of about 105
neurons. The human brain actually contains around 1011
neurons. Let us say that we have N neurons. Each neuron requires Np log2N bits to completely specify its connections, where p is the average number of connections.
Therefore we need at least N2p log2N bits to specify the
entire on-off connectivity matrix of N neurons. If the
connectivity degree p is not very sparse then we just
need N2 bits. So, if we solve min共N2 , N2p log2N兲 = 3.3
⫻ 109 base pairs in the human genome using a connectivity degree of 1%, we obtain a maximum of 105 neurons that can be completely specified. Since we do not
know how much of the genome is used for brain connec-
1255
Rabinovich et al.: Dynamical principles in neuroscience
tivity, it is not possible to narrow down the estimation.
Nevertheless, it does not make sense to expect the
whole genome to specify all connections in the brain.
This simple estimate makes clear that learning and synaptic plasticity have a very important role in determining the connectivity of brain circuits.
The dynamics of complex network models are difficult
to dissect. The mapping of the corresponding local and
global bifurcations in a low-dimensional system has been
extensively studied. To perform such analysis in highdimensional systems is very demanding if not impossible. Average measures, such as mean firing rates, average membrane potential, correlations, etc., can help us
to understand the dynamics of the network as a function
of a few variables. One of the first models to use a meanfield approach was the Wilson-Cowan model 共Wilson
and Cowan, 1973兲. Individual neurons in the model resemble integrate-and-fire neurons with a membrane integration time ␮ and a refractory period r. Wilson and
Cowan’s main hypothesis is that the unreliable individual responses, when grouped together, can lead to
more reliable operations. The Wilson-Cowan formalism
can be reduced to the following equations:
⳵E共x,t兲
= − E共x,t兲 + 关1 − rE共x,t兲兴
␮
⳵t
⫻ Le
−
␮
冕
冋冕
E共y,t兲wee共y,x兲dy
册
I共y,t兲wei共y,x兲dy + Se共x,t兲 ,
−
冕
冋冕
dxi共t兲
=⌰
␮
dt
␮
共33兲
E共y,t兲wie共y,x兲dy
册
I共y,t兲wii共y,x兲dy + Si共x,t兲 ,
共34兲
where E共x , t兲 and I共x , t兲 are the proportions of firing
neurons in the excitatory and inhibitory population, the
coordinate x is a continuous variable that represents the
position in the cortical surface, wee, wei, wie, and wii are
the connectivity weights, and Se and Si are external inputs to the excitatory and inhibitory populations, respectively. The gain functions Le and Li basically reflect the
expected proportions of excitatory and inhibitory neurons receiving at least threshold excitation per unit of
time. One subtle trick used in the derivation of this
model is that the membrane integration time is introduced through synaptic connections. The model expressed in this form attempts to eliminate the uncertainty of single neurons by grouping them according to
those with reliable common responses. We are still left
with the problem of what to expect in a network of clusters connected randomly to each other. Here we will
discuss it in more detail.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
冉兺
NE
j=1
冊
NI
wEE
ij xj共t兲
E
− 兺 wEI
− xi共t兲,
ij yj共t兲 + Si
j=1
共35兲
⳵I共x,t兲
= − I共x,t兲 + 关1 − rI共x,t兲兴
⳵t
⫻ Li
In a random network of excitatory and inhibitory neurons, it is not uncommon to find oscillatory activity 共Jin,
2002; Huerta and Rabinovich, 2004兲. However, it is more
interesting to study the transient behavior of neural recurrent networks. These are fast behaviors and important for sensory processing and for the control of motor
commands. In studying this one needs to address two
main issues: 共i兲 whether it is possible to consistently find
networks with random connections, described by equations similar to Eqs. 共33兲 and 共34兲, behaving regularly,
and 共ii兲 whether transient behavior in these networks is
reproducible.
Huerta and Rabinovich 共2004兲 showed, using the
Wilson-Cowan formalism, periodic sequential activity
共limit cycles兲 is more likely to be found in regions of the
control parameter space where inhibitory and excitatory
synapses are slightly out of balance. However, reproducible transient dynamics is more likely found in the region of parameter space far from balanced excitation
and inhibition. In particular, the authors investigated the
model
dyi共t兲
=⌰
dt
冉
NE
NI
j=1
j=1
冊
兺 wIEij xj共t兲 − 兺 wIIij yj共t兲 + SIi − yi共t兲,
共36兲
where x共t兲 and yi共t兲 represent the fractions of active neurons in cluster i of the excitatory and inhibitory populations, respectively. The numbers of excitatory and inhibitory clusters are NE and NI. The labels E and I are used
to denote quantities associated with the excitatory or
inhibitory populations, respectively. The external inputs
SE,I are instantaneous kicks applied to a fraction of the
total population at time zero. The gain function is ⌰共z兲
= 兵tanh关共z − b兲 / ␴兴 + 1其 / 2, with a threshold b = 0.1 below
the excitatory and inhibitory synaptic strength of a
single connection. Clusters are taken to have very sharp
thresholds of excitability by choosing ␴ = 0.01. There is a
wide range of values that generates similar results. The
time scale is set as done by Wilson and Cowan 共1973兲,
␮ = 10 ms. The connectivity matrices wXY
have entries
ij
drawn from a Bernoulli process 共Huerta and Rabinovich, 2004兲. The main control parameters in this problem
are the probabilities of connections from population to
population.
Now we can answer the following question: What kind
of activity can a network with many neurons and random connections produce? Intuition suggests that the
answer has to be a complex multidimensional dynamics.
However, this is not the case 共Fig. 49兲: most observable
stimulus-dependent dynamics are more simple and reproducible; periodic, transient, or chaotic 共also low dimensional兲.
1256
Rabinovich et al.: Dynamical principles in neuroscience
FIG. 50. 共Color兲 Spatiotemporal patterns of coordinated
rhythms induced by stimuli in a model of the inferior olive.
Several structures with different frequencies can coexist simultaneously in a commensurate representation of the spiking frequencies when several stimuli are present. Incommensurate
stimuli are introduced in the form of current injections in different clusters of the network. 共Panels on the right show the
positions of the input clusters.兲 These current injections induce
different spiking frequencies in the neurons. Colors in these
panels represent different current injections, and thus different
spiking frequencies in the input clusters. Top row shows the
activity of a network with two different input clusters. Bottom
row shows the activity of a network with 25 different input
clusters. Sequences develop in time from left to right. Regions
with the same color have synchronous behavior. Color bar
maps the membrane potential. Red corresponds to spiking
neurons 共−45 mV is above the firing threshold in the model兲.
Dark blue means hyperpolarized activity. Bottom panel shows
the activity of a single neuron with subthreshold oscillations
and spiking activity. Modified from Varona, Aguirre, et al.,
2002.
FIG. 49. 共Color online兲 Three-dimensional projections of
simulations of random networks of 200 neurons. For illustrative purposes we show three types of dynamics that can be
generated by a random network: 共top兲 chaos, 共middle兲 limit
cycle 共both in the areas of parameter space that are close to
balanced兲, and 共bottom兲 transient dynamics 共far from balanced兲.
This is a very important point for understanding cortex dynamics that involves the cooperative activity of
many complex networks 共units or microcircuits兲. From
the functional point of view, the stimulus-dependent dynamics of the cortex can be considered as a coordinated
behavior of many units with low-dimensional transient
dynamics. This is the basis of a new approach to cortex
modeling named the “liquid-state machine” 共Maass et
al., 2002兲.
F. Coordination of sequential activity
Coordination of different sequential behaviors is crucially important for survival. From the modeling point of
view it is a very complex problem. The IO 共a network
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
already discussed in Sec. III.B.2兲 has been suggested as a
system that coordinates motor voluntary movements involving several simultaneous rhythms 共Llinás and Welsh,
1993兲. Here an example of how subthreshold oscillations
coordinate different incommensurate rhythms in a commensurate fashion is shown. In the IO, neurons are electrically coupled to their close neighbors. Their activity is
characterized by subthreshold oscillations and spiking
activity 共see Fig. 50兲. The cooperative dynamics of the
IO under the action of several incommensurate inputs
has been modeled by Varona, Aguirre, et al. 共2002兲. The
results of these large-network simulations show that the
electrical coupling of IO neurons produces quasisynchronized subthreshold oscillations. Because spiking activity can happen only on top of these oscillations, incommensurate inputs can produce regions with different
commensurate spiking frequencies. Several spiking frequencies are able to coexist in these networks. The coexistence of different rhythms is related to the different
clusterization of the spatiotemporal patterns.
Another important question related to coordination
of several sequential behaviors concerns the dynamical
principles that can be a basis for fast neuronal planning
and reaction to a changing environment. One might
Rabinovich et al.: Dynamical principles in neuroscience
think that the WLC principle can be generalized in order to organize the sequential switching according to 共i兲
the learned skill and 共ii兲 the dynamical sensory inputs.
The corresponding mathematical model might be similar
to Eqs. 共26兲–共29兲 together with a learning rule similar to
Eq. 共25兲. Stimuli Sl change sequentially and the timing of
each step 共the time that the system spends moving from
the vicinity of one saddle to the vicinity of the next one;
see Fig. 51兲 should be coordinated with the time of
change in the environment. In recurrent networks, as a
result of learning, the stimulus can go sequentially to the
specific goal of an optimal heteroclinic sequence among
many such sequences that exist in the phase space of the
model. What is important is that at the same time, i.e., in
parallel with the choosing of the rest of the motor plan,
the already existing part of the motor activity plan is
executed.
The two ideas just discussed can be applied to the
cerebellar circuit, which is an example of a complex recurrent network 共see Fig. 52兲. To give an impression of
the complexity of the cerebellar cortex we note that it is
organized into three layers: the molecular layer, the
Purkinje cell layer, and the granule cell layer. Only two
significant inputs reach the cerebellar cortex: mossy fibers and climbing fibers. Mossy fibers are in the majority
共4:1兲 and carry a wealth of sensory and contextual information of multiple modalities. They make specialized
excitatory synapses in structures called “glomeruli” with
the dendrites of numerous granule cells. Granule cell
axons form parallel fibers that run transversely in the
molecular layer, making excitatory synapses with
Purkinje cells. Each Purkinje cell receives ⬇150 000 synapses. These synapses are thought to be major storage
sites for the information acquired during motor learning.
The Purkinje cell axon provides the only output from
the cerebellar cortex. This is via the deep cerebellar nuclei. Each Purkinje cell receives just one climbing fiber
input from the inferior olive, but this input is very powerful because it involves several hundreds of synaptic
contacts. The climbing fiber is thought to have a role in
teaching in the cerebellum. The Golgi cell is excited by
mossy fibers and granule cells and exercises an inhibitory feedback control upon granule cell activity. Stellate
and basket cells are excited by parallel fibers in order to
provide feedforward inhibition to Purkinje cells.
The huge number of inhibitory neurons and the architecture of the cerebellar networks 共de Zeeuw et al.,
1998兲 support the generalized WLC mechanism for coordination. A widely discussed hypothesis is that the
specific circuitry of the IO, cerebellar cortex, and deep
cerebellar nuclei called the slow loop 共see Fig. 52兲 can
serve as a dynamical working memory or as a neuronal
clock with ⬇100-ms cycle time which would make it easy
to connect it to behavioral time scales 共Kistler and de
Zeeuw, 2002; Melamed et al., 2004兲.
Temporal coordination and, in particular, synchronization of neural activity is a robust phenomenon, frequently observed across populations of neurons with diverse membrane properties and intrinsic frequencies. In
the light of such diversity the question of how precise
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1257
FIG. 51. Illustration of the learned sequential switching in a
recurrent network with WLC dynamics: Thin lines, possible
learned sequences; thick line, sequential switching chosen online by the dynamical stimulus.
synchronization can be achieved in heterogeneous networks is critical. Several mechanisms have been suggested and many of them require an unreasonably high
degree of network homogeneity or very strong connectivity to achieve coherent neural activity. As discussed
above 共Sec. II.A.4兲, in a network of two synaptically
coupled neurons STDP at the synapse leads to the dynamical self-adaptation of the synaptic conductance to a
value that is optimal for the entrainment of the postsynaptic neuron. It is interesting to note that just a few
STDP synapses are able to make the entrainment of a
FIG. 52. A schematic representation of the mammalian cerebellar circuit. Arrows indicate the direction of transmission
across each synapse. Sources of mossy fibers: Ba, basket cell;
BR, brush cell; cf, climbing fiber; CN, cerebellar nuclei; Go,
Golgi cell; IO, inferior olive; mf, mossy fiber; pf, parallel fiber;
PN, pontine nuclei; sb and smb, spiny and smooth branches of
P cell dendrites, respectively; PC, Purkinje cell; bat, basket cell
terminal; pcc, P cell collateral; no, nucleo-olivary pathway; nc,
collateral of nuclear relay cell. Modified from Voogd and
Glickstein, 1998.
1258
Rabinovich et al.: Dynamical principles in neuroscience
heterogeneous network of electrically coupled neurons
more effective 共Zhigulin and Rabinovich, 2004兲. It has
been shown that such a network oscillates with a much
higher degree of coherence than when it is subject to
stimulation that is mediated by STDP synapses as compared with stimulation through static synapses. The observed phenomenon depends on the number of stimulated neurons, the strength of electrical coupling, and
the degree of heterogeneity. In reality, long-term plasticity depends not only on spike timing 共STDP兲 but also on
the firing rate and the cooperativity among different
neuronal inputs 共Sjöström et al., 2001兲. This makes modeling self-organization and learning more challenging.
Real behavior in nonstationary or complex environments, as already discussed, requires switching between
different sequential activities. Jancke et al. 共2000兲 have
identified distributed regions in different parts of the
cortex that are involved in the switching among sequential movements. It is important for dynamical modeling
that this differential pattern of activation is not seen for
simple repetitive movements. Thus such movements are
too simple to evoke additional activation. This means
that a dynamical model that aims to describe the sequential behavior in general has to correctly describe
the switching from a low-dimensional subspace to a
high-dimensional state space, and vice versa. There are
no general methods for describing multidimensional dissipative nonlinear systems with such transient but reproducible dynamics. We think that the WLC principle
might be the first step in this direction.
V. CONCLUSION
Physicists, mathematicians, and physiologists all agree
that an important attribute of any dynamical model of
CNS activity is that not only should it be able to fit the
available anatomical and physiological data, but it
should also be capable of explaining function and predicting behavior. However, the ways in which physicists
and mathematicians, on one hand, and physiologists, on
the other hand, use modeling are based on their own
experience and views and thus are different. In this review we tried to bring these different viewpoints closer
together and, using many examples from the sensory,
motor, and central nervous systems, discussed just a few
principles like reproducibility, adaptability, robustness,
and sensitivity.
Let us return to the questions formulated at the beginning of the review:
• What can nonlinear dynamics learn from neuroscience?
• What can
dynamics?
neuroscience
learn
from
nonlinear
After reading this review, we hope the reader can join
us in integrating the key messages in our presentation.
Perhaps we may offer our compact formulation.
Addressing the first question of what nonlinear dynamics can learn from neuroscience:
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
• The most important activities of neuronal systems
are transient and cannot be understood by analyzing
attractor dynamics alone. These need to be augmented by reliable descriptions of stimulusdependent transient motions in state space as this
comprises the heart of most neurobiological activity.
Nonetheless, because the dynamics of realistic neuronal models are strongly dissipative, their stimulusdependent transient behavior is strongly attracted to
some low-dimensional manifolds embedded in the
high-dimensional state space of the neural network.
It is a strong stimulus to nonlinear dynamics to develop a theory of reasonably low-dimensional transient activity and, in particular, to consider the local
and global bifurcations of such objects as homoclinic
and heteroclinic trajectories.
• For many dynamical problems of neuroscience, in
contrast to traditional dynamical approaches, the initial conditions do matter crucially. Persistent neuronal activity 共i.e., dynamical memory兲, stimulusdependent
transient
competition,
stimulusdependent transient synchronization, and stimulusdependent synaptic plasticity are all aspects of this.
Clearly, addressing these important phenomena will
require an expansion in our approaches to dynamical
systems.
Addressing the second question of what neuroscience
can learn from nonlinear dynamics:
• Dynamical models confirm the key role of inhibition
in neuronal systems. The function of inhibition is not
just to organize a balance with excitation in order to
stabilize a network but much more: 共a兲 inhibitory
networks can generate rhythms, such as reproducible
and adaptive motor rhythms in CPGs, or gamma
rhythms in the brain; 共b兲 they are responsible for the
transformation of an identity sensory code to a spatiotemporal code important for better recognition in
an acoustically cluttered environment; and 共c兲 thanks
to inhibition, neural systems can be at the same time
very sensitive to their input and robust against noise.
• Dynamical chaos is not just a fundamental phenomenon but also important for the survival of living organisms. Neuronal systems may use chaos for the organization of nontrivial behavior such as the
irregular hunting-swimming of Clione and for the organization of higher brain functions.
• The improvement in yield, stability, and longevity of
multielectrode recordings, new imaging techniques,
combined with new data processing methods, have
allowed neurophysiologists to describe brain activities as the dynamics of spatiotemporal patterns in
some virtual space. We think this is a basis for building a bridge between transient large-scale brain activity and animal behavior.
And finally as we pursue the investigation of dynamical principles in neuroscience, we hope that eventually
not to see these two questions apart from one another
Rabinovich et al.: Dynamical principles in neuroscience
but as an integrated approach to deep and complex scientific problems.
ACKNOWLEDGMENTS
We thank Ramon Huerta and Thomas Nowotny for
their help, and Rafael Levi and Valentin Zhigulin for
useful comments. This work was supported by NSF
Grant No. NSF/EIA-0130708, and Grant No. PHY
0414174; NIH Grant No. 1 R01 NS50945 and Grant No.
NS40110; MEC BFI2003-07276, and Fundación BBVA.
GLOSSARY
AL
AMPA receptors
bumps
CA1
carbachol
Clione
CNS
CPG
depolarization
dynamic clamp
GABA
heteroclinic loop
heteroclinic trajectory
antennal lobe, the first site of
sensory integration from the olfactory receptors of insects.
transmembrane receptor for
the neurotransmitter glutamate
that mediates fast synaptic
transmission.
spatially localized regions of
high neural activity.
subsystem of the hippocampus
with a very active role in general memory.
chemical that induces oscillations in in vitro preparations.
marine mollusk whose nervous
system is frequently used in
neurophysiology studies.
central nervous system.
central pattern generator, a
small neural circuit that can
produce stereotyped rhythmic
outputs without rhythmic sensory or central input.
any change in the neuron membrane potential that makes it
more positive than when the
cell is in its resting state.
a computer setup to insert virtual conductances into a neural
membrane typically used to
add synaptic input to a cell by
calculating the response current to a specific presynaptic input.
neurotransmitter of typically
inhibitory synapses; they can be
mediated by fast GABA共A兲 or
slow GABA共B兲 receptors.
a closed chain of heteroclinic
trajectories.
trajectory that lies simultaneously on the stable manifold
of one saddle point 共or limit
cycle兲 and the unstable manifold of another saddle 共or limit
cycle兲 connecting them.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
HH
1259
Hodgkin-Huxley
neuron
model.
HVC
high vocal center in the brain of
songbirds.
hyperpolarization
any change in the neuron membrane potential that makes it
more negative than when the
cell is in its resting state.
interneurons
neurons whose axons remain
within a particular brain region
as contrasted with projection
neurons, which have axons projecting to other brain regions,
or with motoneurons, which innervate muscles.
IO
inferior olive, a neural system
that is an input to the cerebellar cortex presumably involved
in motor coordination.
KCs
Kenyon cells, interneurons of
the mushroom body of insects.
Kolmogorov-Sinai
a measure of the degree of preentropy
dictability of further states visited by a chaotic trajectory
started within a small region in
a state space.
LP
lateral pyloric neuron of the
crustacean
stomatogastric
CPG.
LTD
long-term depression, activitydependent decrease of synaptic
efficacy transmission.
LTM
long-term memory.
LTP
long-term
potentiation,
activity-dependent reinforcement of synaptic efficacy transmission.
Lyapunov exponents ␭j the rate of exponential divergence from perturbed initial
conditions in the jth direction
of the state space. For trajectories belonging to a strange attractor the spectrum ␭j is independent of initial conditions
and characterizes the stable
chaotic behavior.
MCs
microcircuits; circuits composed of a small number of
neurons that perform specific
operational tasks.
mushroom body
lobed subsystem of the insect
brain involved in classification,
learning, and memory of odors.
mutual information
a measure of the independence
of two signals X and Y, i.e., the
information of X that is shared
by Y. In the discrete case, if the
joint probability density function of X and Y is p共x , y兲
= P共X = x , Y = y兲, the probability
1260
neuromodulators
neurotransmitters
pacemaker
PD
phase synchronization
共locking兲
place cell
plasticity
PN
Purkinje cell
RA
receptor
receptor neuron
SHS
SN
SSM
statocyst
STDP
STM
structural stability
synapse
Rabinovich et al.: Dynamical principles in neuroscience
density function of X alone is
f共x兲 = P共X = x兲, and the probability density function of Y
alone is g共y兲 = P共Y = y兲, then
the mutual information of X
and Y is given by I共X , Y兲
= 兺x,yp共x , y兲log2关p共x , y兲 / f共x兲g共y兲兴.
a substance other than a neurotransmitter, released by neurons that can affect the intrinsic
and synaptic dynamics of other
neurons.
chemicals that are used to relay
at the synapses the signals between neurons.
neuron or circuit that has endogenous rhythmic activity.
pyloric dilator neuron of the
crustacean CPG.
the onset of a certain relationship between the phases of
coupled self-sustained oscillators.
a type of neuron found in the
hippocampus that fires strongly
when an animal is in a specific
location in an environment.
changes that occur in the organization of synaptic connections or intracellular dynamics.
projection or principal neurons.
main cell type of the cerebellar
cortex.
premotor nucleus of the songbird brain.
a protein on the cell membrane
that binds to a neurotransmitter, neuromodulator, or other
substance, and initiates the cellular response to the ligand.
sensory neuron.
stable heteroclinic sequence.
sensory neuron.
sequential spatial memory.
balance organ in some invertebrates that consists of a spherelike structure containing a mineralized mass 共statolith兲 and
several sensory neurons also
called statocyst receptors.
spike-timing-dependent plasticity.
short-term memory.
condition in which small
changes in the parameters do
not change the topology of the
phase portrait in the state
space.
specialized junction through
which neurons signal to one an-
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
synfire chain
WLC
other. There are at least three
different types of synapses: excitatory and inhibitory chemical synapses and electrical synapses or gap junctions.
propagation of synchronous
spiking activity in a sequence of
layers of neurons belonging to
a feedforward network.
winnerless-competition
principle for the nonautonomous
transient dynamics of neural
systems receiving external
stimuli and exhibiting sequential switching among temporal
“winners.”
REFERENCES
Abarbanel, H., R. Huerta, and M. I. Rabinovich, 2002, Proc.
Natl. Acad. Sci. U.S.A. 99, 10132.
Abarbanel, H. D. I., 1997, Analysis of Observed Chaotic Data
共Springer, New York兲.
Abarbanel, H. D. I., R. Huerta, M. I. Rabinovich, N. F.
Rulkov, P. F. Rowat, and A. I. Selverston, 1996, Neural Comput. 8, 1567.
Abarbanel, H. D. I., M. Rabinovich, and M. Sushchik, 1993,
Introduction to Nonlinear Dynamics for Physicists 共World Scientific, Singapore兲.
Abeles, M., 1991, Corticonics: Neural Circuits of the Cerebral
Cortex 共Cambridge University Press, Cambridge, England兲.
Abeles, M., H. Bergman, E. Margalit, and E. Vaadia, 1993, J.
Neurophysiol. 79, 1629.
Adrian, E. D., and Y. Zotterman, 1926, J. Physiol. 共London兲
61, 151.
Afraimovich, V., V. Zhigulin, and M. Rabinovich, 2004, Chaos
14, 1123.
Afraimovich, V. S., M. I. Rabinovich, and P. Varona, 2004, Int.
J. Bifurcation Chaos Appl. Sci. Eng. 14, 1195.
Aihara, K., 2002, Proc. IEEE 90, 919.
Amari, S., 1977, Biol. Cybern. 27, 77.
Amit, D., and N. Brunel, 1997a, Cereb. Cortex 7, 237.
Amit, D. J., and N. Brunel, 1997b, Cereb. Cortex 7, 237.
Andronov, A., 1933, in The First All-Union Conference on
Auto-oscillations 共GTTI, Moscow Leningrad兲, pp. 32–71.
Andronov, A., A. Vitt, and S. Khaikin, 1949, Theory of Oscillations 共Princeton University Press, Princeton, NJ兲.
Andronov, A. A., and L. Pontryagin, 1937, Dokl. Akad. Nauk
SSSR 14, 247.
Arbib, M. A., P. Érdi, and J. Szentágothai, 1997, Neural Organization: Structure, Function, and Dynamics 共Bradford Book/
MIT, Cambridge, MA兲.
Arieli, A., A. Sterkin, A. Grinvald, and A. Aertsen, 1996, Science 273, 1868.
Arnold, V., V. Afrajmovich, Y. Ilyashenko, L. Shilnikov, and N.
Kazarinoff, 1999, Bifurcation Theory and Catastrophe Theory
共Springer, New York兲.
Ashby, W. R., 1960, Design for a Brain, 2nd ed. 共Wiley, New
York兲.
Ashwin, P., M. Field, A. Rucklidge, and R. Sturman, 2003,
Chaos 13, 973.
Aswin, P., and M. Timme, 2005, Nature 共London兲 436, 36.
Rabinovich et al.: Dynamical principles in neuroscience
Aviel, Y., C. Mehring, M. Abeles, and D. Horn, 2003, Neural
Comput. 15, 1321.
Baird, B., and F. Eeckman, 1993, in Associative Neural Memories: Theory and Implementation, edited by M. H. Hassoun
共Oxford University Press, New York兲, p. 135.
Bao, W., and J.-Y. Wu, 2003, J. Neurophysiol. 90, 333.
Barlow, H. B., 1972, Perception 1, 371.
Bartos, M., Y. Manor, F. Nadim, E. Marder, and M. P. Nusbaum, 1999, J. Neurosci. 19, 6650.
Bartos, M., and M. P. Nusbaum, 1997, J. Neurosci. 17, 2247.
Bi, G., 2002, Biol. Cybern. 87, 319.
Bi, G., and M. Poo, 1998, J. Neurosci. 18, 10464.
Bi, G., and M. Poo, 2001, Annu. Rev. Neurosci. 24, 139.
Bliss, T. V., and T. Lomo, 1973, J. Physiol. 共London兲 232, 331.
Block, H. D., 1962, Rev. Mod. Phys. 34, 123.
Bondarenko, V., G. S. Cymbalyuk, G. Patel, S. Deweerth, and
R. Calabrese, 2003, Neurocomputing 52-54, 691.
Borda, J. C., 1781, Memoire sur les elections au scrutin 共Histoire de l’Academie Royale des Sciences, Paris兲.
Bressloff, P., 2002, Phys. Rev. Lett. 89, 088101.
Bressloff, P., and J. D. Cowan, 2002, Neural Comput. 14, 493.
Bressloff, P., J. Cowan, M. Golubitsky, P. Thomas, and M.
Weiner, 2001, Philos. Trans. R. Soc. London, Ser. B 356, 299.
Brody, C., and J. Hopfield, 2003, Neuron 37, 843.
Brunel, N., 2003, Cereb. Cortex 13, 1151.
Brunel, N., and X. Wang, 2001, J. Comput. Neurosci. 11, 63.
Busse, F., and K. Heikes, 1980, Science 208, 173.
Calabrese, R., F. Nadim, and O. Olsen, 1995, J. Neurobiol. 27,
390.
Canavier, C., D. Baxter, J. Clark, and J. Byrne, 1993, J. Neurophysiol. 69, 2252.
Canavier, C., J. Clark, and J. Byrne, 1990, Biophys. J. 57, 1245.
Casado, J., 2003, Phys. Rev. Lett. 91, 208102.
Castelo-Branco, M., R. Goebel, S. Neuenschwander, and W.
Singer, 2000, Nature 共London兲 405, 685.
Cateau, H., and T. Fukai, 2001, Neural Networks 14, 675.
Cazelles, B., M. Courbage, and M. I. Rabinovich, 2001, Europhys. Lett. 56, 504.
Chay, T., 1985, Physica D 16, 223.
Chechik, G., 2003, Neural Comput. 15, 1481.
Chklovskii, D. B., B. W. Mel, and K. Svoboda, 2004, Nature
共London兲 431, 78.
Chow, C., and N. Kopell, 2000, Neural Comput. 12, 1643.
Cohen, A., P. Holmes, and R. Rand, 1982, J. Math. Biol. 13,
345.
Cohen, M. A., and S. Grossberg, 1983, IEEE Trans. Syst. Man
Cybern. 13, 815.
Connors, B., 2002, Nature 共London兲 420, 133.
Connors, B. W., and Y. Amitai, 1997, Neuron 18, 347.
Contreras, D., A. Destexhe, T. Sejnowski, and M. Steriade,
1996, Science 274, 771.
Coombes, S., and A. Osbaldestin, 2000, Phys. Rev. E 62, 4057.
Cossart, R., D. Aronov, and R. Yuste, 2003, Nature 共London兲
423, 283.
Cowan, N., 2001, Behav. Brain Sci. 24, 87.
Crawford, J., 1991, Rev. Mod. Phys. 63, 991.
Crevier, D., and M. Meister, 1998, J. Neurophysiol. 79, 1869.
Cross, M., and P. Hohenberg, 1993, Rev. Mod. Phys. 65, 851.
Curts, C. E., and M. D’Esposito, 2003, Trends Cogn. Sci. 7,
415.
Cymbalyuk, G., O. Gaudry, M. Masino, and R. Calabrese,
2002, J. Neurosci. 22, 10580.
deCharms, R., and M. Merzenich, 1996, Nature 共London兲 361,
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1261
610.
deCharms, R. C., 1998, Proc. Natl. Acad. Sci. U.S.A. 95, 15166.
de Nó, R. L., 1938, J. Neurophysiol. 1, 207.
de Ryter van Steveninck, R., G. Lewen, S. Strong, R. Koberle,
and W. Bialek, 1997, Science 275, 1805.
Desmedt, J., and C. Tomberg, 1994, Neurosci. Lett. 168, 126.
de Zeeuw, C. I., J. I. Simpson, C. C. Hoogenaraad, N. Galjart,
S. K. E. Koekkoek, and T. J. H. Ruigrok, 1998, Trends Neurosci. 21, 391.
Diesmann, M., M.-O. Gewaltig, and A. Aertsen, 1999, Nature
共London兲 402, 529.
Diwadkar, V. A., P. A. Carpenter, and M. Just, 2000, Neuroimage 11, 85.
Doboli, S., A. A. Minai, and P. Best, 2000, Neural Comput. 12,
1009.
Dragoi, G., K. Harris, and G. Buzsaki, 2003, Neuron 39, 843.
Du, X., B. K. Ghosh, and P. Ulinski, 2005, IEEE Trans.
Biomed. Eng. 52, 566.
Duda, R., E. Hart, and D. Stork, 2001, Pattern Classification
共Wiley, New York兲.
Durstewitz, D., J. Seamans, and T. Sejnowski, 2000, Nat. Neurosci. 3, 1184.
Egorov, A., B. Hamann, E. Fransen, M. Hasselmo, and A.
Alonso, 2002, Nature 共London兲 420, 173.
Eguia, M. C., M. I. Rabinovich, and H. D. I. Abarbanel, 2000,
Phys. Rev. E 62, 7111.
Elhilali, M., J. Fritz, D. Klein, J. Z. Simon, and S. Shamma,
2004, J. Neurosci. 24, 1159.
Elowitz, M., and S. Leibler, 2000, Nature 共London兲 403, 335.
Elson, R. C., A. I. Selverston, H. D. I. Abarbanel, and M. I.
Rabinovich, 2002, J. Neurophysiol. 88, 1166.
Elson, R. C., A. I. Selverston, R. Huerta, N. F. Rulkov, M. I.
Rabinovich, and H. D. I. Abarbanel, 1998, Phys. Rev. Lett.
81, 5692.
Engel, A. K., P. Fries, and W. Singer, 2001, Nat. Rev. Neurosci.
2, 704.
Ermentrout, B., 1992, Neural Networks 5, 415.
Ermentrout, G., and J. Cowan, 1979, Biol. Cybern. 34, 137.
Ermentrout, G. B., 1998, Rep. Prog. Phys. 61, 353.
Ermentrout, G. B., and N. Kopell, 1984, SIAM J. Math. Anal.
15, 215.
Fano, R. M., 1961, Ed., Transmission of Information: A Statistical Theory of Communications 共MIT, New York兲.
Fatt, P., and B. Katz, 1952, J. Physiol. 共London兲 117, 109.
Fell, J., G. Fernandez, P. Klaver, C. Elger, and P. Fries, 2003,
Brain Res. Rev. 42, 265.
Feudel, U., A. Neiman, X. Pei, W. Wojtenek, H. Braun, M.
Huber, and F. Moss, 2000, Chaos 10, 231.
FitzHugh, R., 1961, Biophys. J. 1, 445.
Fitzpatrick, D. C., R. Batra, T. R. Sanford, and S. Kuwada,
1997, Nature 共London兲 388, 871.
Foss, J., A. Longtin, B. Mensour, and J. Milton, 1996, Phys.
Rev. Lett. 76, 708.
Freeman, W., 1972, Progress in Theoretical Biology 共Academic,
New York兲, Vol. 2.
Freeman, W., 2000, Neurodynamics: An Exploration in Mesoscopic Brain Dynamics 共Springer, New York兲.
Fuhrmann, G., I. Segev, H. Markram, and M. Tsodyks, 2002, J.
Neurophysiol. 87, 140.
Fukai, T., and S. Tanaka, 1997, Neural Comput. 9, 77.
Galan, R. F., S. Sachse, C. G. Galizia, and A. V. M. Herz, 2004,
Neural Comput. 16, 999.
1262
Rabinovich et al.: Dynamical principles in neuroscience
Gallager, R. G., 1968, Ed., Information Theory and Reliable
Communication 共Wiley, New York兲.
Garcia-Sanchez, M., and R. Huerta, 2003, J. Comput. Neurosci. 15, 5.
Gavrilov, N., and A. Shilnikov, 2000, Am. Math. Soc. Transl.
200, 99.
Georgopoulos, A. P., A. B. Schwartz, and R. E. Kettner, 1986,
Science 233, 1416.
Gerstner, W., and W. Kistler, 2002, Spiking Neuron Models
共Cambridge University Press, Cambridge, England兲.
Getting, P., 1989, Annu. Rev. Neurosci. 12, 185.
Gewaltig, M.-O., M. Diesmann, and A. Aertsen, 2001, Neural
Networks 14, 657.
Ghose, G., and J. Maunsell, 1999, Neuron 24, 79.
Glass, L., 1995, The Handbook of Brain Theory and Neural
Networks 共MIT, Cambridge兲, pp. 186–189.
Goldman-Rakic, P., 1995, Neuron 14, 477.
Goroff, D., 1992, Ed., New Methods of Celestial Mechanics
共AIP, New York兲.
Gray, C., P. Konig, A. Engel, and W. Singer, 1989, Nature
共London兲 338, 334.
Grillner, S., 2003, Nat. Rev. Neurosci. 4, 573.
Grossberg, S., 1973, Stud. Appl. Math. 52, 213.
Gu, H., M. Yang, L. Li, Z. Liu, and W. Ren, 2003, Phys. Lett. A
319, 89.
Guckenheimer, J., S. Gueron, and R. Harris-Warrick, 1993,
Philos. Trans. R. Soc. London, Ser. B 341, 345.
Guckenheimer, J., and P. Holmes, 1986, Nonlinear Oscillations,
Dynamical Systems, and Bifurcations of Vector Fields
共Springer, New York兲.
Guckenheimer, J., and P. Holmes, 1988, Math. Proc. Cambridge Philos. Soc. 103, 189.
Hass, J., T. Nowotny, and H. D. I. Abarbanel, 2006, J. Neurophysiol., doi:10.1152/jn.00551.2006.
Hahnloser, R. H., A. A. Kozhevnikov, and M. S. Fee, 2002,
Nature 共London兲 419, 65.
Hamilton, K., and J. Kauer, 1985, Brain Res. 338, 181.
Hatsopoulos, N., L. Paninski, and J. Donoghue, 2003, Exp.
Brain Res. 149, 478.
Hebb, R., 1949, The Organization of Behavior 共Wiley, New
York兲.
Hindmarsh, J. L., and R. M. Rose, 1984, Proc. R. Soc. London,
Ser. B 221, 87.
Hodgkin, A. L., and A. F. Huxley, 1952, J. Physiol. 共London兲
117, 500.
Hopfield, J., and C. Brody, 2001, Proc. Natl. Acad. Sci. U.S.A.
98, 1282.
Hopfield, J., and C. Brody, 2004, Proc. Natl. Acad. Sci. U.S.A.
101, 337.
Hopfield, J. J., 1982, Proc. Natl. Acad. Sci. U.S.A. 79, 2554.
Huang, X., W. C. Troy, Q. Yang, H. Ma, C. R. Laing, S. J.
Schiff, and J.-Y. Wu, 2004, J. Neurosci. 24, 9897.
Huerta, R., and M. I. Rabinovich, 2004, Phys. Rev. Lett. 93,
238104.
Huerta, R., M. Rabinovich, H. H. D. I. Abarbanel, and M.
Bazhenov, 1997, Phys. Rev. E 55, R2108.
Huerta, R., P. Varona, M. Rabinovich, and H. Abarbanel, 2001,
Biol. Cybern. 84, L1.
Ichinohe, N., F. Fujiyama, T. Kaneko, and K. S. Rockland,
2003, J. Neurosci. 23, 1372.
Ikegaya, Y., G. Aaron, R. Cossart, D. Aronov, I. Lampl, D.
Ferster, and R. Yuste, 2004, Science 304, 559.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
Ince, D., 1992, Ed., Mechanical Intelligence: Collected Works of
A. M. Turing 共North-Holland, Amsterdam兲.
Ito, M., 1982, Annu. Rev. Neurosci. 5, 275.
Izhikevich, E., 2004, IEEE Trans. Neural Netw. 15, 1063.
Izhikevich, E., 2006, Dynamical Systems in Neuroscience: The
Geometry of Excitability and Bursting 共MIT, Cambridge,
MA兲.
Jackson, A., V. Gee, S. Baker, and R. Lemon, 2003, Neuron 38,
115.
Jancke, L., M. Himmelbach, N. Shah, and K. Zilles, 2000, Neuroimage 12, 528.
Jefferys, J., R. Traub, and M. Whittington, 1996, Trends Neurosci. 19, 202.
Jin, D. Z., 2002, Phys. Rev. Lett. 89, 208102.
Jones, S. R., B. Mulloney, T. J. Kaper, and N. Kopell, 2003, J.
Neurosci. 60, 3457.
Kandel, E. R., J. H. Schwartz, and T. M. Jessell, 2000, Principles of Neural Science 共McGraw-Hill, New York兲.
Kaplan, D., and L. Glass, 1995, Understanding Nonlinear Dynamics 共Springer, New York兲.
Katz, B., 1969, The Release of Neurotransmitter Substances
共Thomas, Springfield, IL兲.
Kay, L., 2003, Chaos 13, 1057.
Kepecs, A., and J. Lisman, 2003, Network Comput. Neural
Syst. 14, 103.
Kim, U., T. Bal, and D. A. McCormick, 1995, J. Neurophysiol.
74, 1301.
Kimpo, R. R., F. E. Theunissen, and A. J. Doupe, 2003, J.
Neurosci. 23, 5730.
Kistler, W. M., and C. I. de Zeeuw, 2002, Neural Comput. 14,
2597.
Koch, C., 1999, Biophysics of Computation 共Oxford University
Press, New York兲.
Konishi, M., 1990, Cold Spring Harbor Symp. Quant. Biol. 55,
575.
Korn, H., and P. Faure, 2003, C R. Seances Soc. Biol. Fil 326,
787.
Kozloski, J., F. Hamzei-Sichani, and R. Yuste, 2001, Science
293, 868.
Krupa, P., 1997, J. Nonlinear Sci. 7, 129.
Kuramoto, Y., 1984, Chemical Oscillations, Waves, and Turbulence 共Springer, New York兲.
Kuznetsov, Y., 1998, Elements of Applied Bifurcation Theory,
2nd ed. 共Springer, New York兲.
Lai, Y. C., M. A. F. Harrison, M. G. Frei, and I. Osorio, 2003,
Phys. Rev. Lett. 91, 068102.
Laing, C. R., W. C. Troy, B. Gutkin, and G. B. Ermentrout,
2002, SIAM J. Appl. Math. 63, 62.
Landau, L. D., and E. M. Lifshitz, 1987, Fluid Mechanics 共Pergamon, New York兲.
Lapicque, L., 1907, J. Physiol. Pathol. Gen. 9, 620.
Lashley, K., 1960, in The Neuropsychology of Lashley, edited
by F. A. Beach, D. O. Hebb, C. T. Morgan, and H. W. Nissen
共McGraw-Hill, New York兲, pp. 506–521.
Latorre, R., F. B. Rodriquez, and P. Varona, 2006, Biol. Cybern. 95, 169.
Lau, P., and G. Bi, 2005, Proc. Natl. Acad. Sci. U.S.A. 102,
10333.
Laurent, G., 1996, Trends Neurosci. 19, 489.
Laurent, G., and H. Davidowitz, 1994, Science 265, 1872.
Laurent, G., M. Stopfer, R. W. Friedrich, M. I. Rabinovich,
and H. D. I. Abarbanel, 2001, Annu. Rev. Neurosci. 24, 263.
Leitch, B., and G. Laurent, 1996, J. Comp. Neurol. 372, 487.
Rabinovich et al.: Dynamical principles in neuroscience
LeMasson, G., S. R.-L. Masson, D. Debay, and T. Bal, 2002,
Nature 共London兲 417, 854.
Lestienne, R., 2001, Prog. Neurobiol. 共Oxford兲 65, 545.
Levi, R., P. Varona, Y. I. Arshavsky, M. I. Rabinovich, and A.
I. Selverston, 2004, J. Neurophysiol. 91, 336.
Levi, R., P. Varona, Y. I. Arshavsky, M. I. Rabinovich, and A.
I. Selverston, 2005, J. Neurosci. 25, 9807.
Lewis, T., and J. Rinzel, 2003, J. Comput. Neurosci. 14, 283.
Leznik, E., and R. Llinas, 2002, Ann. N.Y. Acad. Sci. 978, 529.
Lin, L., R. Osan, S. Shoham, W. Jin, W. Zuo, and J. Tsien,
2005, Proc. Natl. Acad. Sci. U.S.A. 102, 6125.
Lindner, B., J. Garca-Ojalvo, A. Neiman, and L. SchimanskyGeier, 2004, Phys. Rep. 392, 2004.
Llinás, R., and J. P. Welsh., 1993, Curr. Opin. Neurobiol. 3, 958.
Loebel, A., and M. Tsodyks, 2002, J. Comput. Neurosci. 13,
111.
Lotka, A. J., 1925, Elements of Physical Biology 共Williams &
Wilkins Co., Baltimore兲.
Maass, W., T. Natschläger, and H. Markram, 2002, Neural
Comput. 14, 2531.
Machens, C., T. Gollisch, O. Kolesnikova, and A. Herz, 2005,
Neuron 47, 447.
Machens, C., R. Romo, and C. Brody, 2005, Science 307, 1121.
Maeda, Y., and H. Makino, 2000, BioSystems 58, 93.
Maeda, Y., K. Pakdaman, T. Nomura, S. Doi, and S. Sato, 1998,
Biol. Cybern. 78, 265.
Mainen, Z., and T. Sejnowski, 1995, Science 268, 1503.
Malenka, R. C., and R. A. Nicoll, 1999, Science 285, 1870.
Malkov, V., M. Rabinovich, and M. M. Sushchik, 1996, Proceedings of Nizhny Novgorod University, edited by V. D.
Shalfeev 共Nizhny Novgorov Univerity, Nizhny Novgorod兲, p.
72.
Mandelblat, Y., Y. Etzion, Y. Grossman, and D. Golomb, 2001,
J. Comput. Neurosci. 11, 43.
Maravall, M., I. Y. Y. Koh, W. Lindquist, and K. Svoboda,
2004, Cereb. Cortex 10, 1093.
Marder, E., and R. L. Calabrese, 1996, Physiol. Rev. 76, 687.
Marder, E., L. A. G. Turrigiano, Z. Liu, and J. Golowasch,
1996, Proc. Natl. Acad. Sci. U.S.A. 93, 13481.
Martin, S. J., P. D. Grimwood, and R. G. M. Morris, 2000,
Annu. Rev. Neurosci. 23, 649.
Mazor, O., and G. Laurent, 2005, Neuron 48, 661.
McClurkin, J. W., L. M. Optican, B. J. Richmond, and T. J.
Gawne, 1991, Science 253, 675.
McCormick, D., Y. Shu, A. Hasenstaub, M. Sanches-Vives, M.
Badoual, and T. Bal, 2003, Cereb. Cortex 13, 1219.
McCulloch, W. S., and W. Pitts, 1943, Bull. Math. Biophys. 5,
115.
Mehta, M. R., A. K. Lee, and M. A. Wilson, 2002, Nature
共London兲 417, 741.
Melamed, O., W. Gerstner, W. Maas, M. Tsodyks, and H.
Markram, 2004, Trends Neurosci. 27, 11.
Mooney, R., and J. Prather, 2005, J. Neurosci. 25, 1952.
Morris, C., and H. Lecar, 1981, Biophys. J. 35, 193.
Nádasdy, Z., 2000, J. Physiol. 共Paris兲 94, 505.
Nagumo, J., S. Arimoto, and S. Yoshizawa, 1962, Proc. IRE 50,
2061.
Nakahara, H., and K. Doya, 1998, Neural Comput. 10, 113.
Nase, G., W. Singer, H. Monyer, and A. K. Engel, 2003, J.
Neurophysiol. 90, 1115.
Nenadic, Z., B. K. Ghosh, and P. Ulinski, 2002, IEEE Trans.
Biomed. Eng. 49, 753.
Nicholls, J. G., A. R. Martin, and B. G. Wallace, 1992, From
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1263
Neuron to Brain: A Cellular and Molecular Approach to the
Function of the Nervous System 共Sinauer Associates, Sunderland, MA兲.
Nowotny, T., 2003, URL http://inls.ucsd.edu/$\sim$nowotny/
dynclamp.html
Nowotny, T., and R. Huerta, 2003, Biol. Cybern. 89, 237.
Nowotny, T., R. Huerta, H. Abarbanel, and M. Rabinovich,
2005, Biol. Cybern. 93, 436.
Nowotny, T., M. I. Rabinovich, R. Huerta, and H. D. I. Abarbanel, 2003, J. Comput. Neurosci. 15, 271.
Nowotny, T., V. P. Zhigulin, A. I. Selverston, H. D. I. Abarbanel, and M. I. Rabinovich, 2003, J. Neurosci. 23, 9776.
Nystrom, L. E., T. S. Braver, F. W. Sabb, M. R. Delgado, D. C.
Noll, and J. D. Cohen, 2000, Neuroimage 11, 424.
O’Keefe, J., and J. Dostrovsky, 1971, Brain Res. 34, 171.
O’Reilly, R., R. Busby, and R. Soto, 2003, The Unity of
Consciousness—Binding, Integration and Dissociation 共Oxford University Press, Oxford兲, Chap. 3, p. 2.5.
Oscarsson, O., 1980, in The Inferior Olivary Nucleus, edited by
J. Courville, C. de Montigny, and Y. Lamarre 共Raven, New
York兲, pp. 279–289.
Ott, E., 1993, Chaos in Dynamical Systems 共Cambridge University Press, Cambridge, England兲.
Panchin, Y., Y. Arshavsky, T. Deliagina, L. Popova, and G.
Orlovsky, 1995, J. Neurophysiol. 73, 1924.
Panzeri, S., S. Schultz, A. Treves, and E. Rolls, 1999, Proc. R.
Soc. London, Ser. B 266, 1001.
Perez-Orive, J., O. Mazor, G. C. Turner, S. Cassenaer, R. I.
Wilson, and G. Laurent, 2002, Science 297, 359.
Persi, E., D. Horn, V. Volman, R. Segev, and E. Ben-Jacob,
2004, Neural Comput. 16, 2577.
Pikovsky, A., M. Rosenblum, and J. Kurths, 2001, Synchronization: A Universal Concept in Nonlinear Sciences 共Cambridge University Press, Cambridge, England兲.
Pinto, R., P. Varona, A. Volkovskii, A. Szücs, H. Abarbanel,
and M. Rabinovich, 2000, Phys. Rev. E 62, 2644.
Poincaré, A., 1905, Le Valeur de la Science 共Flammarion,
Paris兲.
Poincaré, H., 1892, Méthodes Nouvells de la Mécanique Céleste
共Gauthier-Villars, Paris兲.
Postlethwaite, C. M., and J. H. P. Dawes, 2005, Nonlinearity 18,
1477.
Pouget, A., P. Dayan, and R. Zemel, 2000, Nat. Rev. Neurosci.
1, 125.
Prinz, A., L. F. Abbott, and E. Marder, 2004a, Trends Neurosci. 27, 218.
Prinz, A. A., D. Bucher, and E. Marder, 2004b, Nat. Neurosci.
7, 1345.
Rabinovich, M., A. Ezersky, and P. Weidman, 2000, The Dynamics of Patterns 共World Scientific, Singapore兲.
Rabinovich, M., R. Huerta, and P. Varona, 2006, Phys. Rev.
Lett. 96, 014101.
Rabinovich, M., A. Volkovskii, P. Lecanda, R. Huerta, H. D. I.
Abarbanel, and G. Laurent, 2001, Phys. Rev. Lett. 87, 068102.
Rabinovich, M. I., and H. D. I. Abarbanel, 1998, Neuroscience
87, 5.
Rabinovich, M. I., and R. Huerta, 2006, Phys. Rev. Lett. 97,
188103.
Ramirez, J., A. Tryba, and F. Pena, 2004, Curr. Opin. Neurobiol. 6, 665.
Reinagel, P., and R. C. Reid, 2002, J. Neurosci. 22, 6837.
Reyes, A. D., 2003, Nat. Neurosci. 6, 593.
Riesenhuber, M., and T. Poggio, 1999, Neuron 24, 87.
1264
Rabinovich et al.: Dynamical principles in neuroscience
Rinzel, J., D. Terman, X. Wang, and B. Ermentrout, 1998, Science 279, 1351.
Robinson, H., and N. Kawai, 1993, J. Neurosci. Methods 49,
157.
Rodriguez, E., N. George, J. Lachaux, J. Martinerie, B.
Renault, and F. Varela, 1999, Nature 共London兲 397, 430.
Rosenblatt, F., 1962, Principles of Neurodynamics: Perceptions
and the Theory of Brain Mechanisms 共Spartan Books, New
York兲.
Roskies, A., 1999, Neuron 24, 7.
Rowe, D., 2002, Behav. Brain Sci. 24, 5.
Roweis, S., and L. Saul, 2000, Science 290, 2323.
Rubin, J. E., and D. Terman, 2004, J. Comput. Neurosci. 16,
211.
Rulkov, N. F., 2002, Phys. Rev. E 65, 041922.
Saari, G., 1995, Basic Geometry of Voting 共Springer-Verlag,
Berlin兲.
Sanchez-Vives, M., and D. McCormick, 2000, Nat. Neurosci. 3,
1027.
Schutter E. D., 2002, Curr. Biol. 12, R363.
Schweighofer, N., K. Doya, H. Fukai, J. V. Chiron, T. Furukawa, and M. Kawato, 2004, Proc. Natl. Acad. Sci. U.S.A.
101, 4655.
Scott, A., 2004, Ed., Encylopedia of Nonlinear Science 共Routledge, New York兲.
Segundo, J. P., and D. H. Perkel, 1969, UCLA Forum Med. Sci.
11, 349.
Segundo, J. P., G. Sugihara, P. Dixon, M. Stiber, and L. F.
Bersier, 1998, Neuroscience 87, 741.
Seliger, P., L. S. Tsimring, and M. I. Rabinovich, 2003, Phys.
Rev. E 67, 011905.
Selverston, A., 2005, Cell Mol. Neurobiol. 25, 223.
Selverston, A., M. Rabinovich, H. Abarbanel, R. Elson, A.
Szücs, R. Pinto, R. Huerta, and P. Varona, 2000, J. Physiol.
共Paris兲 94, 357.
Senseman, D. M., and K. A. Robbins, 1999, J. Neurosci. 19,
RC3 1.
Seung, H., D. Lee, B. Reis, and D. Tank, 2000, J. Comput.
Neurosci. 9, 171.
Seung, H. S., 1998, in Advances in Neural Information Processing Systems, edited by M. I. Jordan, M. J. Kearns, and S. A.
Solla 共MIT, Cambridge, MA兲, Vol. 10.
Seung, H. S., and H. Sompolinsky, 1993, Proc. Natl. Acad. Sci.
U.S.A. 90, 10749.
Shadlen, M. N., and W. T. Newsome, 1998, J. Neurosci. 18,
3870.
Sharp, A., M. O’Neil, L. Abbott, and E. Marder, 1993, Trends
Neurosci. 16, 389.
Shepherd, G., 1998, The Synaptic Organization of the Brain
共Oxford University Press, New York兲.
Sherman, A., and J. Rinzel, 1992, Proc. Natl. Acad. Sci. U.S.A.
89, 2471.
Shilnikov, A., and G. Cymbalyuk, 2005, Phys. Rev. Lett. 94,
048101.
Shu, Y., A. Hasenstaub, and D. McCormick, 2003, Nature
共London兲 423, 288.
Silberberg, G., S. Grillner, F. LeBeau, R. Maex, and H.
Markram, 2005, Trends Neurosci. 28, 541.
Simmers A. J., and M. Moulins, 1988, J. Neurophysiol. 59, 740.
Singer, W., 1999, Neuron 24, 49.
Singer, W., 2001, Ann. N.Y. Acad. Sci. 929, 123.
Singer, W., and C. M. Gray, 1995, Annu. Rev. Neurosci. 18,
555.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
Sjöström, P. J., G. G. Turrigiano, and S. B. Nelson, 2001, Neuron 32, 1149.
Softky, W. R., 1995, Curr. Opin. Neurobiol. 5, 239.
Solis, M., and D. Perkel, 2005, J. Neurosci. 25, 2811.
Sompolinsky, H., and I. Kanter, 1986, Phys. Rev. Lett. 57, 2861.
Sompolinsky, H., H. Yoon, K. Kang, and M. Shamir, 2001,
Phys. Rev. E 64, 051904.
Soto-Trevino, C., P. Rabbah, E. Marder, and F. Nadim, 2005, J.
Neurophysiol. 94, 590.
Soto-Trevino, C., K. A. Thoroughman, E. Marder, and L. E.
Abbott, 2001, Nat. Neurosci. 4, 297.
Stein, S. G., S. Grillner, A. I. Selverston, and G. S. Douglas,
1997, Eds., Neurons, Networks, and Motor Behavior 共MIT,
Cambridge, MA兲.
Stent, G. S., and W. O. Friesen, 1977, Biol. Cybern. 28, 27.
Stone, E., and D. Armbruster, 1999, Chaos 9, 499.
Stone, E., and P. Holmes, 1990, SIAM J. Appl. Math. 50, 726.
Stopfer, M., V. Jayaraman, and G. Laurent, 2003, Neuron 39,
991.
Strogatz, S. H., 2001, Nonlinear Dynamics and Chaos: With
Applications to Physics, Biology, Chemistry and Engineering
共Perseus Books Group, Cambridge, MA兲.
Szekely, G., 1965, Acta Physiol. Acad. Sci. Hung. 27, 285.
Szücs, A., R. D. Pinto, M. I. Rabinovich, H. D. I. Abarbanel,
and A. I. Selverston, 2003, J. Neurophysiol. 89, 1363.
Szücs, A., P. Varona, A. Volkovskii, H. D. I. Abarbanel, M.
Rabinovich, and A. Selverston, 2000, NeuroReport 11, 563.
Szücs, A., A. I. Selverston, M. I. Rabinovich, and H. D. I.
Abarbanel, 2004, Soc. Neurosci Abs. 420, 4.
Tams, G., A. Lorincz, A. Simon, and J. Szabadics, 2003, Science 299, 1902.
Terman, D., A. Bose, and N. Kopell, 1996, Proc. Natl. Acad.
Sci. U.S.A. 93, 15417.
Thompson, A. M., and J. Deuchars, 1994, Trends Neurosci. 17,
119.
Tiitinen, H., J. Sinkkonen, K. Reinikainen, K. Alho, J. Lavikainen, and R. Naatanen, 1993, Nature 共London兲 364, 59.
Tohya, S., A. Mochizuki, and Y. Iwasa, 2003, J. Theor. Biol.
221, 289.
Traub, R. D., and R. Miles, 1991, Ann. N.Y. Acad. Sci. 627,
277.
Tsodyks, M., T. Kenet, A. Grinvald, and A. Arieli, 1999, Science 286, 1943.
Tsodyks, M., and H. Markram, 1997, Proc. Natl. Acad. Sci.
U.S.A. 94, 719.
Tsuda, I., 1991, World Futures 32, 167.
Turrigiano, G. G., E. Marder, and L. Abbott, 1996, J. Neurophysiol. 75, 963.
Vaadia, E., I. Haalman, M. Abeles, H. Bergman, Y. Prut, H.
Slovin, and A. Aertsen, 1995, Nature 共London兲 373, 515.
van Essen, D. C., 1979, Annu. Rev. Neurosci. 2, 227.
van Vreeswijk, C., L. F. Abbott, and G. B. Ermentrout, 1994, J.
Comput. Neurosci. 1, 313.
van Vreeswijk, V., and H. Sompolinsky, 1996, Science 274,
1724.
Varona, P., C. Aguirre, J. J. Torres, M. I. Rabinovich, and H.
D. I. Abarbanel, 2002, Neurocomputing 44-46, 685.
Varona, P., M. I. Rabinovich, A. I. Selverston, and Y. I. Arshavsky, 2002, Chaos 12, 672.
Varona, P., J. J. Torres, H. D. I. Abarbanel, M. I. Rabinovich,
and R. Elson, 2001, Biol. Cybern. 84, 91.
Varona, P., J. J. Torres, R. Huerta, H. D. I. Abarbanel, and M.
I. Rabinovich, 2001, Neural Networks 14, 865.
Rabinovich et al.: Dynamical principles in neuroscience
Venaille, A., P. Varona, and M. I. Rabinovich, 2005, Phys. Rev.
E 71, 061909.
Vogel, E. K., and M. G. Machizawa, 2004, Nature 共London兲
428, 748.
Vogels, T., K. Rajan, and L. Abbott, 2005, Annu. Rev. Neurosci. 28, 357.
Volterra, V., 1931, in Animal Ecology, edited by R. N. Chapman 共McGraw-Hill, New York兲, pp. 409–448.
von der Malsburg, C., 1999, Neuron 24, 95.
von der Malsburg, C., and W. Schneider, 1986, Biol. Cybern.
54, 29.
Voogd, J., and M. Glickstein, 1998, Trends Neurosci. 21, 370.
Wang, X., 2001, Trends Neurosci. 24, 455.
Wang, X.-J., and J. Rinzel, 1995, The Handbook of Brain
Theory and Neural Networks 共MIT, Cambridge, MA兲, p. 686.
Rev. Mod. Phys., Vol. 78, No. 4, October–December 2006
1265
Waugh, F., C. Marcus, and R. Westervelt, 1990, Phys. Rev.
Lett. 64, 1986.
Wilson, H. R., 1999, Spikes, Decisions, and Actions 共Oxford
University Press, New York兲.
Wilson, H. R., and J. D. Cowan, 1973, Kybernetik 13, 55.
Wilson, M. A., and B. L. McNaughton, 1993, Science 261, 1055.
Wilson, M. A., and B. McNaughton, 1994, Science 265, 676.
Wolfe, J., and K. Cave, 1999, Neuron 24, 11.
Yuste, R., J. MacLean, J. Smith, and A. Lansner, 2005, Nat.
Rev. Neurosci. 6, 477.
Zeeman, E., and M. Zeeman, 2002, Nonlinearity 15, 2019.
Zhigulin, V. P., and M. I. Rabinovich, 2004, Neurocomputing
58-60, 373.
Zhigulin, V. P., M. I. Rabinovich, R. Huerta, and H. D. I. Abarbanel, 2003, Phys. Rev. E 67, 021901.
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