London
STANDING by a pond at London Zoo, grabbing a moment to talk shop with an
American colleague on what was supposed to be a family outing, Karl Friston
tried to describe a new vision of the brain. Traditional thinking held that the
brain was some kind of computer, crunching its way through billions of inputs
each second, outputting consciousness. But said Friston, a theoretical
neurobiologist at London’s Institute of Neurology, it is more as if the arrival
of those inputs provokes a widespread disturbance in the brain.
Look, Friston told his Harvard friend, the brain is like this pond. You throw
in a pebble—the sensory input—and you get ripples. That’s the
neurons responding. Sure, the pattern says something about the way the pebble
hit the surface. But the pond is already covered in ripples caused by other
pebbles, so the pattern appears a little chaotic. And then once the ripples
spread out far enough to begin bouncing off the sides, he continued, the shape
of the pond begins to affect what is going on. The whole thing keeps evolving
and becoming more complex.
Advertisement
Yes, replied his friend, nodding furiously, and as we throw more and more
pebbles—or rather experiences—into the pond, we change the kind of
patterns it produces, and even the shape of the pond itself. This system has a
memory!
In the early 90s, in hundreds of private conversations like this, mind
scientists were groping their way towards a fresh view of the brain—one
based on the idea that mental states are dynamically evolved rather than
clinically computed. Back then, the arguments were little more than hand-waving
exercises. People were familiar with the new ideas about chaos, complexity and
nonlinear systems coming out of places like the Santa Fe Institute, but unsure
how they applied to the brain. Today, however, the dynamic revolution is
beginning to roll. At workshops and meetings around the world, researchers like
Friston are talking publicly about dynamic models of the brain, and the evidence
to support the new theories that is beginning to fall into place.
A replacement for the brain-as-computer model certainly seems overdue. The
textbook view has been that brain cells are simple logic gates, adding and
subtracting input spikes until some threshold level of charge is breached, at
which point they convulse to produce a spike of their own. The all-or-nothing
nature of a cell’s firing promised to lift neurons clear of the usual soupy
sloppiness of cellular processes, allowing the brain to carry out digitally
crisp, noise-free calculations.
The task for researchers was simply to discover how the output of each cell
encoded a message. In a chase likened to the hunt to crack the genetic code,
neuroscientists became obsessed with finding the “neural code”. They tried to
discover whether the message was contained in the strength of a spike, the
average number of spikes produced each second, or in the timing of the firing,
with information carried only on those spikes which were synchronised with
spikes from other cells
(see “Dot dot dot, dash dash dash”, Â鶹´«Ă˝, 18 May 1996, p 40).
But the neurons have proved slippery customers. “For 30 years we’ve been
going along quite nicely, with lots of expensive equipment, lots of expensive
people and lots of papers being produced, but finally the answers aren’t there.
We can’t even say what it is about the spike train of an individual neuron that
counts,” says Rodney Douglas of the Institute of Neuroinformatics in Zurich.
Much worse for the idea of a simple, crackable neural code are the smattering of
recent findings which show that the output of any individual neuron also depends
on what the brain happens to be thinking at the time. It’s as if rather than the
spikes combining to produce conscious awareness, consciousness is able to decide
how the cells should spike.
The search for the neural code began in earnest in the 1960s with David Hubel
and Torsten Wiesel’s Nobel prizewinning demonstration that certain cells in the
primary visual cortex—the first part of the higher brain to receive
sensory input from the eyes—fired only in response to the sight of a line
or edge, indeed, only to a line of the correct slope. The neurons represented
every possible orientation of a line at each point of the visual field, and were
lined up in the brain like dots on a TV screen, creating a physical “map” of the
input from the eye.
Other researchers soon showed that cells in different areas of the sensory
cortex made maps of the frequencies of sound, and even, in the case of touch, of
the contours of the body. In fact, the entire wrinkled surface of the cortex
seemed to be a mosaic of mapping, with the primary sensory areas being the first
rung of a hierarchy of processing. The primary maps were reworked, as the
message from one layer of cells, supposedly encoded in the neurons’ spike
trains, fed into the next. So, for example, about halfway up the visual
hierarchy, cells might fire in response to movement in a certain direction and
with a certain speed, or a certain shape of a certain colour. Sensory qualities
began to emerge. At the very top of the hierarchy, neurons would react only to
complete objects—say, the sight of a face or a hand. Each rung of the
hierarchy was built on the digital clarity of the spike pattern of the neurons
below, providing a way for the brain to compute a precise, conscious
representation of the real world. That, at least, was the theory.
The problem was that most of the evidence for the theory came from studies of
anaesthetised animals whose heads had been propped up in front of screens with
their eyelids pinned back. When, in the late 1980s, researchers developed
techniques that made it easier to record neural impulses from awake animals, the
story of brain cells as simple switches, hard-wired to respond to this line or
that movement, changed dramatically.
Take an experiment reported in Nature last year by neuroscientists
John Maunsell, at Baylor College of Medicine in Houston, Texas, and Stefan Treue
of the University of TĂĽbingen in Germany. They studied those neurons about
halfway up the visual hierarchy that deal with motion, in monkeys trained to
watch moving dots on a screen. When the monkeys did not have to follow any dot
in particular, the motion cells simply burst into life each time they spotted a
dot heading in their preferred direction. But as soon as the monkeys were asked
to concentrate on a single dot—they had been trained to do this without
moving their heads or their eyes—the cells became picky. When the target
dot came into view, the cells went wild, doubling their firing rate, while the
response from the same neurons to non-target dots moving in the correct
direction became weaker.
It all makes good psychological sense. The cells turn the volume up in
response to movement that is the focus of attention, and mute it in response to
other movement. But it also raises the question of how the brain’s mental state
is managing to transmogrify the cell’s spike pattern.
Spooky
Neuroscientists dread any hint that something spooky might be going on. They
try to slide past the problem of the brain’s mental state interfering with the
clarity of the long-sought neural code with euphemisms such as “selective
attention effects” or “state-dependent modulations”.
Yet Maunsell admits that his findings strike to the heart of the idea that
the brain works as an input-driven machine: “We are coming to the end of one
generation of effort,” he predicts. “The next generation is going to have to
look at the whole system [and] understand the effect that plans, decisions and
actions can have on what neurons do.”
Maunsell and Treue are not the only ones who have been backed into a corner
by their own data. A rash of similar findings is emerging from labs run by the
likes of Robert Desimone at the National Institute of Mental Health near
Washington DC and Richard Andersen at Caltech in Pasadena. One team of
researchers has even found that cells right at the bottom of the visual
hierarchy—those that take the “freshest” input from the eyes and might be
expected to be least influenced by the brain’s mental state—are also at
its mercy.
David Leopold and Nikos Logothetis, both also at Baylor, reported in
Nature last year the results of an experiment in which monkeys looked
through stereoscopic displays so that each eye saw a different
image—gratings angled in different directions. The brain makes sense of
such a conflict by allowing the view of one eye to dominate: the monkey is
consciously aware of seeing only a single image.
According to the old view of the brain, the cortex cells that get their input
direct from the eyes shouldn’t be involved in the mental jiggery-pokery that
suppresses the image from one eye—it should happen higher up the
hierarchy. Instead, Leopold and Logothetis found that the firing of about a
fifth of cells in the primary visual cortex depended on which image the monkeys
signalled they were seeing. Even at the lowest level, there was an attention
effect.
Booming with the enthusiasm of an outsider who is beginning to be proved
right, Scott Kelso, a dynamicist who studies the brain and behaviour at Florida
Atlantic University in Boca Raton, claims that results like these will only make
sense once the old notion of the brain processing encoded messages through
nothing more than a hierarchy of inputs and outputs is abandoned. Instead, he
says, neuroscience must make a fresh start and recognise that the brain is a
dynamical system—an organ that evolves its patterns of activity rather
than computes them.
The very word “dynamic” strikes fear into the hearts of many researchers,
relying as it does on the maths of chaos and complexity theory. Jargon such as
“metastability”, “critical boundaries” and “loosely coupled attractors” litters
the papers. Still, the champions of the dynamic view stress that a few simple
ideas are key.
Bursting forth
First, says Kelso, stop thinking of neurons as if they are exchanging
messages. Instead (to use another of the hydraulic metaphors favoured by
dynamicists), the spike patterns of a cell are like a whorl erupting in moving
water—a local expression of a much wider balance of forces. After all, it
is no secret that most of the 5000 input lines to the average brain cell are
actually parts of feedback loops returning via neighbouring neurons, or those
higher up the hierarchy. Barely a tenth of the connections come from sense
organs or mapping levels lower in the hierarchy. Every neuron is plumbed into a
sea of feedback. The signals coming up the chain may provide the seed of a
response, but in the end, the cell’s spike patterns evolve in concert with how
the rest of the brain is reacting to the stimulus. The spike pattern is less a
crisp code and more the chatterings of a system forever moving towards an
equilibrium.
This is good, as it means there is nothing spooky about how thoughts and
intentions, that is mental states, shape the activity of a neuron, and vice
versa. But it does mean that levels of consciousness matter, especially if you
are trying to make sense of a neuron’s spike train.
When a cell is firing in relative isolation—for example, when an animal
is unconscious—its response will be at its most hard-wired, a simple sum
of its sensory or lower inputs. Like a ringing phone, the neuron will announce
that it has a message, but no one lifts the receiver to get the conversation
going. But as the experiments with wide-awake monkeys show, as soon as a cell
becomes drawn into some greater wave of processing, its firing appears far less
hard-wired. Of course, it takes time for the wave to build up, which is why
attention effects usually show up about a tenth of a second behind the first
exposure to the focus of the attention.
The second crucial change needed in the thinking about neural processing, say
the dynamicists, is to realise that the brain is always in a state of tension,
its circuits drawn tight like the surface of Friston’s pond. Computer analogies
suggest that the brain is a blank screen until cells fire to light up a picture.
But almost every brain cell is constantly firing, a fact that has long troubled
neuroscientists. There is a steady tick-over of at least three or four spikes a
second even in an area of the brain that seems to be doing nothing. The
temptation is to dismiss this activity as meaningless, just a leakage of
current. But dynamicists say the spikes bouncing around the brain’s connections
must be maintaining it at a certain level of tone, giving each new input
something to disturb in the first place.
Going a step further, they argue, this background firing presumably creates
some meaning. But what? The brain stores memories as patterns of connections
between cells—new experiences prompt the strengthening of old connections,
or the growth of new ones. The tick-over firing echoing around the brain could
be a defocused representation of everything you have ever learnt or known. When
the brain processes new information, it is not a matter of lighting up dark
circuits but of driving a generalised, weakly defined state of representation
towards a specific one. The brain is always on, it just needs tuning in.
Hot spots
As the message of the dynamicists begins to sink in, neuroscientists are
having to think again about the way they do experiments and analyse their data.
The most obvious change, says Friston, is that researchers must allow enough
time to get an accurate fix on what a cell is up to. Indeed, to truly understand
a cell’s firing pattern, you need to know how far along its feedback trajectory
it has gone. At the moment, neuroscientists tend to concentrate on a cell’s
first reaction to a stimulus rather than waiting another tenth of a second or so
until the feedback has had long enough to focus what the cell is saying.
What’s more, in a dynamic scheme, cells apparently saying nothing (that show
no change in firing rate and therefore go unreported when the time comes to
write up a research paper) are still important. “Rather than talking about a
hunt for the neural code, we should be talking about a hunt for the
metric—the right kind of spatiotemporal measure to give the full picture
of how a cell’s response evolves,” Friston says.
Friston is as good as his word when it comes to his own interest in human
brain scanning. The standard approach to scanner studies is to look for brain
hot spots, the bits of the brain that have to work the hardest when a subject
does some mental or physical task. Like 20th-century phrenologists, researchers
look for the brain bump that “does” hand movements or mental imagery. But if the
brain really works by evolving patterns of connections, then it is how areas of
the brain, even those that appear quite faint on a scan, join together over time
that tells the true story.
As part of a two-day symposium on dynamical neuroscience at this October’s
meeting of the Society for Neuroscience in New Orleans, Friston attempted to
prove that the distinction between mapping brain hot spots and patterns of
connections is not purely academic. He reported an analysis of brain scan data
collected by magnetoencephalography, using a method of correlation that
highlights increases and decreases in activity in different parts of the brain
that occur over the same period. It turned out that high activity in an area at
the front of the brain called the prefrontal cortex, and low activity in an area
towards the back called the parietal cortex, are tightly coupled just when the
volunteer is deciding to make small hand movements. Usual methods of analysis
would have missed the link. What’s more, says Friston, it took a twentieth of a
second or more for this coupling to appear, a clear sign that the connection had
to evolve.
For now, however, dynamicists like Friston and Kelso are keeping a sense of
perspective. They know that convincing mainstream neurobiologists to stop
looking for machine-like order in a biological organ that thrives on the
creative energy of chaos and feedback is going to take more than a few
experiments and lots of enthusiasm. As Kelso said after the New Orleans
symposium: “If we are serious about the brain as a self-organising system, then
we need new tools, new concepts, a new language. Even the way we measure the
brain has to be different.” That process has started, and once it is complete,
the dynamicists say, neuroscience’s golden age of discovery will be ready to
begin.
- Activity changes in early visual cortex reflect monkeys’ percepts during binocular rivalry
by D. Leopold and N. Logothetis, Nature, vol 379, p 549 (1996) - Attentional modulation of visual motion processing in cortical areas MT and MST
by S. Treue and J. Maunsell, Nature, vol 382, p 539 (1996) - Neuronal transients
by K. Friston, Proceedings of the Royal Society London B, vol 261, p 401 (1995)
