At 08:21 PM 4/9/04 +0200, Eugen Leitl wrote:
>It should look a lot like a Golgi stain of your neocortex, though, the
Sorry the below is long, but its subscription only, and the comparisons
to man-made networks are worth reading.
Science, Vol 301, Issue 5641, 1870-1874 , 26 September 2003
Communication in Neuronal Networks
Simon B. Laughlin1 and Terrence J. Sejnowski2,3*
Brains perform with remarkable efficiency, are capable of prodigious
computation, and are marvels of communication. We are
beginning to understand some of the geometric, biophysical, and energy
constraints that have governed the evolution of cortical
networks. To operate efficiently within these constraints, nature has
optimized the structure and function of cortical networks with
design principles similar to those used in electronic networks. The
brain also exploits the adaptability of biological systems to
reconfigure in response to changing needs.
1 Department of Zoology, University of Cambridge, Downing Street,
Cambridge CB2 3EJ, UK.
2 Howard Hughes Medical Institute, Salk Institute for Biological
Studies, La Jolla, CA 92037, USA.
3 Division of Biological Sciences, University of California, San Diego,
La Jolla, CA 92093, USA.
Science, Vol 301, Issue 5641, 1870-1874 , 26 September 2003
[DOI: 10.1126/science.1089662]
Previous Article
Table of Contents
Next Article
Communication in Neuronal Networks
Simon B. Laughlin1 and Terrence J. Sejnowski2,3*
Brains perform with remarkable efficiency, are capable of prodigious
computation, and are marvels of communication. We are
beginning to understand some of the geometric, biophysical, and energy
constraints that have governed the evolution of cortical
networks. To operate efficiently within these constraints, nature has
optimized the structure and function of cortical networks with
design principles similar to those used in electronic networks. The
brain also exploits the adaptability of biological systems to
reconfigure in response to changing needs.
1 Department of Zoology, University of Cambridge, Downing Street,
Cambridge CB2 3EJ, UK.
2 Howard Hughes Medical Institute, Salk Institute for Biological
Studies, La Jolla, CA 92037, USA.
3 Division of Biological Sciences, University of California, San Diego,
La Jolla, CA 92093, USA.
* To whom correspondence should be addressed. E-mail: terry(a)salk.edu
Neuronal networks have been extensively studied as computational
systems, but they also serve as communications networks in
transferring large amounts of information between brain areas. Recent
work suggests that their structure and function are
governed by basic principles of resource allocation and constraint
minimization, and that some of these principles are shared with
human-made electronic devices and communications networks. The
discovery that neuronal networks follow simple design rules
resembling those found in other networks is striking because nervous
systems have many unique properties.
To generate complicated patterns of behavior, nervous systems have
evolved prodigious abilities to process information.
Evolution has made use of the rich molecular repertoire, versatility,
and adaptability of cells. Neurons can receive and deliver signals at up
to 105 synapses and can
combine and process synaptic inputs, both linearly and nonlinearly, to
implement a rich repertoire of operations that process information (1).
Neurons can also
establish and change their connections and vary their signaling
properties according to a variety of rules. Because many of these
changes are driven by spatial and
temporal patterns of neural signals, neuronal networks can adapt to
circumstances, self-assemble, autocalibrate, and store information by
changing their properties
according to experience.
The simple design rules improve efficiency by reducing (and in some
cases minimizing) the resources required to implement a given task. It
should come as no surprise
that brains have evolved to operate efficiently. Economy and efficiency
are guiding principles in physiology that explain, for example, the way
in which the lungs, the
circulation, and the mitochondria are matched and coregulated to supply
energy to muscles (2). To identify and explain efficient design, it is
necessary to derive and
apply the structural and physicochemical relationships that connect
resource use to performance. We consider first a number of studies of
the geometrical constraints
on packing and wiring that show that the brain is organized to reduce
wiring costs. We then examine a constraint that impinges on all aspects
of neural function but has
only recently become apparentenergy consumption. Next we look at
energy-efficient neural codes that reduce signal traffic by exploiting
the relationships that
govern the representational capacity of neurons. We end with a brief
discussion on how synaptic plasticity may reconfigure the cortical
network on a wide range of
time scales.
Geometrical and Biophysical Constraints on Wiring
Reducing the size of an organ, such as the brain, while maintaining
adequate function is usually beneficial. A smaller brain requires fewer
materials and less energy for
construction and maintenance, lighter skeletal elements and muscles for
support, and less energy for carriage. The size of a nervous system can
be reduced by
reducing the number of neurons required for adequate function, by
reducing the average size of neurons, or by laying out neurons so as to
reduce the lengths of their
connections. The design principles governing economical layout have
received the most attention.
Just like the wires connecting components in electronic chips, the
connections between neurons occupy a substantial fraction of the total
volume, and the wires (axons
and dendrites) are expensive to operate because they dissipate energy
during signaling. Nature has an important advantage over electronic
circuits because
components are connected by wires in three-dimensional (3D) space,
whereas even the most advanced VLSI (very large scale integration)
microprocessor chips use
a small number of layers of planar wiring. [A recently produced chip
with 174 million transistors has seven layers (3).] Does 3D wiring
explain why the volume
fraction of wiring in the brain (40 to 60%; see below) is lower than in
chips (up to 90%)? In chips, the components are arranged to reduce the
total length of wiring.
This same design principle has been established in the nematode worm
Caenorhabditis elegans, which has 302 neurons arranged in 11 clusters
called ganglia. An
exhaustive search of alternative ganglion placements shows that the
layout of ganglia minimizes wire length (4).
Cortical projections in the early sensory processing areas are
topographically organized. This is a hallmark of the six-layer
neocortex, in contrast to the more diffuse
projections in older three-layer structures such as the olfactory
cortex and the hippocampus. In the primary visual cortex, for example,
neighboring regions of the
visual field are represented by neighboring neurons in the cortex.
Connectivity is much higher between neurons separated by less than 1 mm
than between neurons
farther apart (see below), reflecting the need for rapid, local
processing within a cortical columnan arrangement that minimizes wire
length. Because cortical neurons
have elaborately branched dendritic trees (which serve as input
regions) and axonal trees (which project the output to other neurons),
it is also possible to predict the
optimal geometric patterns of connectivity (57), including the optimal
ratios of axonal to dendritic arbor volumes (8). These conclusions were
anticipated nearly 100
years ago by the great neuroanatomist Ramon y Cajal: "After the many
shapes assumed by neurons, we are now in a position to ask whether this
diversity... has been
left to chance and is insignificant, or whether it is tightly regulated
and provides an advantage to the organism.... We realized that all of
the various conformations of
the neuron and its various components are simply morphological
adaptations governed by laws of conservation for time, space, and
material" [(9), p. 116].
The conservation of time is nicely illustrated in the gray matter of
the cerebral cortex. Gray matter contains the synapses, dendrites, cell
bodies, and local axons of
neurons, and these structures form the neural circuits that process
information. About 60% of the gray matter is composed of axons and
dendrites, reflecting a high
degree of local connectivity analogous to a local area network. An
ingenious analysis of resource allocation suggests that this wiring
fraction of 60% minimizes local
delays (10). This fraction strikes the optimum balance between two
opposing tendencies: transmission speed and component density. Unlike
the wires in chips,
reducing the diameter of neural wires reduces the speed at which
signals travel, prolonging delays. But it also reduces axon volume, and
this allows neurons to be
packed closer together, thus shortening delays.
Global Organization of the Communication Network
Long-range connections between cortical areas constitute the white
matter and occupy 44% of the cortical volume in humans. The thickness of
gray matter, just a few
millimeters, is nearly constant in species that range in brain volume
over five orders of magnitude. The volume of the white matter scales
approximately as the 4/3
power of the volume of the gray matter, which can be explained by the
need to maintain a fixed bandwidth of long-distance communication
capacity per unit area of
the cortex (11) (Fig. 1). The layout of cortical areas minimizes the
total lengths of the axons needed to join them (12). The prominent folds
of the human cortex allow
the large cortical area to be packed in the skull but also allow
cortical areas around the convolutions to minimize wire length; the
location of the folds may even arise
from elastic forces in the white matter during development (13).
Fig. 1. Cortical white and gray matter
volumes of 59 mammalian species are related by a power law that spans
five to six
orders of magnitude. The line is the least
squares fit, with a slope around 1.23 1 0.01 (mean 1 SD) and correlation
of 0.998.
The number of white matter fibers is
proportional to the gray matter volume; their average length is the
cubic root of that
volume. If the fiber cross section is
constant, then the white matter volume should scale approximately as the
4/3 power of the
gray matter volume. An additional factor
arises from the cortical thickness, which scales as the 0.10 power of
the gray matter
volume. [Adapted from (11)] [View Larger
Version of this Image (44K GIF file)]
The global connectivity in the cortex is very sparse, and this too
reduces the volume occupied by long-range connections: The probability
of any two cortical neurons
having a direct connection is around one in a hundred for neurons in a
vertical column 1 mm in diameter, but only one in a million for distant
neurons. The distribution
of wire lengths on chips follows an inverse power law, so that shorter
wires also dominate (14). If we created a matrix with 1010 rows and
columns to represent the
connections between every pair of cortical neurons, it would have a
relatively dense set of entries around the diagonal but would have only
sparse entries outside the
diagonal, connecting blocks of neurons corresponding to cortical areas.
The sparse long-range connectivity of the cortex may offer some of the
advantages of small-world connectivity (15). Thus, only a small fraction
of the computation
that occurs locally can be reported to other areas, through a small
fraction of the cells that connect distant cortical areas; but this may
be enough to achieve activity
that is coordinated in distant parts the brain, as reflected in the
synchronous firing of action potentials in these areas, supported by
massive feedback projections
between cortical areas and reciprocal interactions with the thalamus
(16, 17).
Despite the sparseness of the cortical connection matrix, the potential
bandwidth of all of the neurons in the human cortex is around a terabit
per second (assuming a
maximum rate of 100 bits/s over each axon in the white matter),
comparable to the total world backbone capacity of the Internet in 2002
(18). However, this
capacity is never achieved in practice because only a fraction of
cortical neurons have a high rate of firing at any given time (see
below). Recent work suggests that
another physical constraintthe provision of energylimits the brain's
ability to harness its potential bandwidth.
Energy Usage Constrains Neural Communication
As the processor speeds of computers increase, the energy dissipation
increases, so that cooling technology becomes critically important.
Energy consumption also
constrains neural processing. Nervous systems consume metabolic energy
continuously at relatively high rates per gram, comparable to those of
heart muscle (19).
Consequently, powering a brain is a major drain on an animal's energy
budget, typically 2 to 10% of resting energy consumption. In humans this
proportion is 20%
for adults and 60% for infants (20), which suggests that the brain's
energy demands limit its size (21).
Energy supply limits signal traffic in the brain (Fig. 2). Deep
anesthesia blocks neural signaling and halves the brain's energy
consumption, which suggests that about
50% of the brain's energy is used to drive signals along axons and
across synapses. The remainder supports the maintenance of resting
potentials and the vegetative
function of neurons and glia. Cortical gray matter uses a higher
proportion of total energy consumption for signaling, more than 75%
(Fig. 2), because it is so richly
interconnected with axons and synapses (21). From the amounts of energy
used when neurons signal, one can calculate the volume of signal traffic
that can be
supported by the brain's metabolic rate. For cerebral cortex, the
permissible traffic is 5 action potentials per neuron per second in rat
(Fig. 2) (22) and <1 per
second in human (23). Given that the brain responds quickly, the
permissible level of traffic is remarkably low, and this metabolic limit
must influence the way in which
information is processed. Recent work suggests that brains have
countered this severe metabolic constraint by adopting energy-efficient
designs. These designs
involve the miniaturization of components, the elimination of
superfluous signals, and the representation of information with
energy-efficient codes.
Fig. 2. Power consumption limits neural
signaling rate in the gray matter of rat cerebral cortex. Baseline
consumption is set by
the energy required to maintain the
resting potentials of neurons and associated supportive tissue (r.p.)
and to satisfy their
vegetative needs (nonsignaling). Signaling
consumption rises linearly with the average signaling rate (the rate at
which neurons
transmit action potentials). The measured
rates of power consumption in rat gray matter vary across cortical areas
and limit
average signaling rates to 3 to 5.5 Hz.
Values are from (19), converted from rates of hydrolysis of adenosine
triphosphate
(ATP) to W/kg using a free energy of
hydrolysis for a molecule of ATP under cellular conditions of 1019 J.
[View Larger Version of this Image (20K
GIF file)]
Miniaturization, Energy, and Noise
The observation that 1 mm3 of mouse cortex contains 105 neurons, 108
synapses, and 4 km of axon (24) suggests that, as in chip design, the
brain reduces energy
consumption by reducing the size and active area of components. Even
though axon diameter is only 0.3 5m (on average), sending action
potentials along these
"wires" consumes more than one-third of the energy supplied to cortical
gray matter (22). Thus, as with computer chips, an efficient layout
(discussed above) and a
high component density are essential for energy efficiency but, as is
also true for chips, miniaturization raises problems about noise.
When a neuron's membrane area is reduced, the number of molecular pores
(ion channels) carrying electrical current falls, leading to a decline
in the signal-to-noise
ratio (SNR) (2527). The noise produced by ion channels, and by other
molecular signaling mechanisms such as synaptic vesicles, is potentially
damaging to
performance. However, the effects of noise are often difficult to
determine because they depend on interactions between signaling
molecules in signaling systems.
These interactions can be highly nonlinear (e.g., the voltage-dependent
interactions between sodium and potassium ion channels that produce
action potentials) and
can involve complicated spatial effects (e.g., the diffusion of
chemical messengers between neurons and the transmission of electrical
signals within neurons). A new
generation of stochastic simulators is being developed to handle these
complexities and determine the role played by molecular noise and
diffusion in neural signaling
(26, 28, 29). With respect to miniaturization, stochastic simulations
(25) show that channel noise places a realistic ceiling on the wiring
density of the brain by setting a
lower limit of about 0.1 5m on axon diameter.
The buildup of noise from stage to stage may be a fundamental
limitation on the logical depth to which brains can compute (30). The
analysis of the relationships
among signal, noise, and bandwidth and their dependence on energy
consumption will play a central role in understanding the design of
neural circuits. The cortex has
many of the hallmarks of an energy-efficient hybrid device (28). In
hybrid electronic devices, compact analog modules operate on signals to
process information, and
the results are converted to digital data for transmission through the
network and then reconverted to analog data for further processing.
These hybrids offer the ability
of analog devices to perform basic arithmetic functions such as
division directly and economically, combined with the ability of digital
devices to resist noise. In the
energy-efficient silicon cochlea, for example, the optimal mix of
analog and digital data (that is, the size and number of operations
performed in analog modules) is
determined by a resource analysis that quantifies trade-offs among
energy consumption, bandwidth for information transmission, and
precision in analog and digital
components. The obvious similarities between hybrid devices and neurons
strongly suggest that hybrid processing makes a substantial contribution
to the energy
efficiency of the brain (31). However, the extent to which the brain is
configured as an energy-efficient hybrid device must be established by a
detailed resource
analysis that is based on biophysical relationships among energy
consumption, precision, and bandwidth in neurons.
Some research strongly suggests that noise makes it uneconomical to
transfer information down single neurons at high rates (29, 31). Given
that a neuron is a
noise-limited device of restricted bandwidth, the information rate is
improved with the SNR, which increases as the square root of the number
of ion channels, making
improvements expensive (25). Thus, doubling the SNR means quadrupling
the number of channels, the current flow, and hence the energy cost.
Given this
relationship between noise and energy cost, an energy-efficient nervous
system will divide information among a larger number of relatively noisy
neurons of lower
information capacity, as observed in the splitting of retinal signals
into ON and OFF pathways (32). Perhaps the unreliability of individual
neurons is telling us that the
brain has evolved to be energy efficient (31).
Saving on Traffic
Energy efficiency is improved when one reduces the number of signals in
the network without losing information. In the nervous system, this
amounts to an economy
of impulses (33) that has the additional advantage of increasing
salience by laying out information concisely. Economy is achieved by
eliminating redundancy. This
important design principle is well established in sensory processing
(34). Redundancy reduction is a goal of algorithms that compress files
to reduce network traffic.
In the brain, efficiency is improved by distributing signals
appropriately in time and space. Individual neurons adopt distributions
of firing rate (35, 36) that maximize
the ratio between information coded and energy expended. Networks of
neurons achieve efficiency by distributing signals sparsely in space and
time. Although it was
already recognized that sparse coding improves energy efficiency (37),
it was Levy and Baxter's detailed analysis of this problem (38) that
initiated theoretical studies
of energy-efficient coding in nervous systems. They compared the
representational capacity of signals distributed across a population of
neurons with the costs
involved. Sparse coding schemes, in which a small proportion of cells
signal at any one time, use little energy for signaling but have a high
representational capacity,
because there are many different ways in which a small number of
signals can be distributed among a large number of neurons. However, a
large population of
neurons could be expensive to maintain, and if these neurons rarely
signal, they are redundant. The optimum proportion of active cells
depends on the ratio between
the cost of maintaining a neuron at rest and the extra cost of sending
a signal. When signals are relatively expensive, it is best to
distribute a few of them among a large
number of cells. When cells are expensive, it is more efficient to use
few of them and to get all of them signaling. Estimates of the ratio
between the energy demands of
signaling and maintenance suggest that, for maximum efficiency, between
1% and 16% of neurons should be active at any one time (22, 23, 38).
However, it is
difficult to compare these predictions with experimental data; a major
problem confronting systems neuroscience is the development of
techniques for deciphering
sparse codes.
There is an intriguing possibility that the energy efficiency of the
brain is improved by regulating signal traffic at the level of the
individual synaptic connections between
neurons. A typical cortical neuron receives on the order of 10,000
synapses, but the probability that a synapse fails to release
neurotransmitter in response to an
incoming signal is remarkably high, between 0.5 and 0.9. Synaptic
failures halve the energy consumption of gray matter (22), but because
there are so many
synapses, the failures do not necessarily lose information (39, 40).
The minimum number of synapses required for adequate function is not
known. Does the
energy-efficient cortical neuron, like the wise Internet user, select
signals from sites that are most informative? This question draws energy
efficiency into one of the
most active and important areas of neuroscience: synaptic plasticity.
Reconfiguring the Network
Long-distance communication in the brain occurs through all-or-none
action potentials, which are transmitted down axons and converted to
analog chemical and
electrical signals at synapses. The initiation of action potentials in
the cortex can occur with millisecond precision (41) but, as we have
just discussed, the
communication at cortical synapses is probabilistic. On a short time
scale of milliseconds to seconds, presynaptic mechanisms briefly
increase or decrease the
probability of transmission at cortical synapses over a wide range,
depending on the previous patterns of activity (42). On longer time
scales, persistent correlated
firing between the presynaptic and postsynaptic neurons can produce
long-term depression or potentiation of the synaptic efficacy, depending
on the relative timing of
the spikes in the two neurons (43).
A new view of the cortical network is emerging from these discoveries.
Rather than being a vast, fixed network whose connection strengths
change slowly, the
effective cortical connectivity is highly dynamic, changing on fast as
well as slow time scales. This allows the cortex to be rapidly
reconfigured to meet changing
computational and communications needs (44). Unfortunately, we do not
yet have techniques for eavesdropping on a large enough number of
neurons to determine
how global reconfiguration is achieved. Local field potentials (LFPs),
extracellular electric fields that reflect the summed activity from
local synaptic currents and other
ion channels on neurons and glial cells, may provide hints of how the
flow of information in cortical circuits is regulated (16). Oscillations
in the 20- to 80-Hz range
occur in the LFPs, and the coherence between spikes and these
oscillations has been found to be influenced by attention and working
memory (45, 46).
Conclusions
The more we learn about the structure and function of brains, the more
we come to appreciate the great precision of their construction and the
high efficiency of their
operations. Neurons, circuits, and neural codes are designed to
conserve space, materials, time, and energy. These designs are exhibited
in the geometry of the
branches of dendritic trees, in the precise determination of wiring
fractions, in the laying out of maps in the brain, in the processing of
signals, and in neural codes. It is
less obvious, but highly likely, that the unreliability of single
neurons is also a mark of efficiency, because noise in molecular
signaling mechanisms places a high price
on precision. To an extent yet to be determined, the noise and
variability observed among neurons is compensated by plasticitythe
ability of neurons to modify their
signaling properties. Neural plasticity also has the potential to
direct the brain's scarce resources to where they will be of greatest
benefit.
