IT’S certainly different, some might even say distasteful, but there’s no
doubt that Steve Potter has the strangest lab rat around. Its body is a virtual
one, running around inside a computer generated world. And its brain is like
nothing you’ve ever seen—an amorphous mass of brain cells from a real rat,
living in a shallow glass dish and wired up to the computer. Potter wants to see
if the “rat” can learn its way around, which means the neurons must sense and
remember the rat’s virtual world and control its movements, just as a real brain
would.
For decades, neuroscientists have struggled to understand just how tangles of
neurons let us carry out everyday wonders such as learning and moving. They’ve
studied the behaviour of single neurons, eavesdropped on the electrical signals
from various areas of the brain and even scanned the brains of people carrying
out complex tasks. But answers have been hard to come by.
It’s difficult to get neurons isolated in a dish to do what they would
normally do, and it’s tricky to see what’s going on at a cellular level when
you’re studying people inside a scanner. What’s needed is a way to bridge the
gap, a way to see neurons close up as they interact with the outside world.
Advertisement
Potter, a neuroscientist at Caltech in Pasadena, is one of several
researchers who think they’ve found a way. Some have built artificial machines
in which real nervous tissue connects directly to electronics. Others are taking
brainstems from fish to control robots or co-opting the brain of a live monkey
to move an artificial arm. As the living neural networks adapt to their
cybernetic parts, the researchers hope to see the learning process in
action.
But it doesn’t end there. Some scientists are toying with the idea of tapping
directly into the computing power of neurons, to build “neurocomputers” far more
powerful than their silicon cousins. They would have to train brain cells to
perform tasks that they wouldn’t normally do—an ambitious idea, not to
mention an ethical minefield—yet the work is well on its way.
William Ditto, co-founder of the Laboratory for Neuroengineering at the
Georgia Institute of Technology in Atlanta, argues that the first simple
neurocomputers may be built within five years. Eventually, they’ll carry out
tasks that conventional computers cannot seem to cope with, such as recognising
handwriting, speech and faces—things our brains can do in the blink of an
eye when only fragments of information are available. Biological systems could
also be better at routeing Internet traffic and navigating autonomous vehicles,
which are surprisingly complex tasks in computing terms.
Of course, it was hoped that artificial neural networks—silicon
versions of the real thing—would solve these problems. But these haven’t
lived up to their promise. Even a single neuron is a complicated
information-processing system, with thousands of inputs and outputs, combining
and modulating signals in vastly more complex ways than we understand. We are
nowhere near emulating one cell, let alone a network of neurons, says Joel
Davis, an expert on adaptive neural systems at the Office of Naval Research in
Arlington, Virginia.
For Peter Fromherz, a neuroscientist from the Max Planck Institute for
Biochemistry in Munich, neurocomputing is still a while away. He is working on
ways to mix biology and silicon so that he can probe individual neurons to study
learning and memory.
Last year, his team immobilised two snail neurons between silicon structures
that looked like picket fences etched onto a chip. Within two days the neurons
grew extensions to connect with each other. They could then exchange electrical
signals with one another, or with electrodes on the chip. The connection allows
him to see precisely how the cell responds to electrical signals, and with more
cells he hopes to study how a whole network changes physically to store a
memory. “We have the basic components for integrating digital electronics with
neural networks,” says Fromherz. “The next step is to have more neurons on the
chip. The goal is to make a very small learning network.”
But it’s too early to speculate where this may lead, says Fromherz. It’s true
that we could use living neural networks without understanding how they work,
but there are other problems. The notion of anchoring neurons to a chip wouldn’t
work for a neurocomputer of any significance, he says, simply because it would
need hundreds of thousands or even millions of neurons. The neurons would have
to be free to move around, making and breaking connections as they please.
Potter has gone at least some way towards tackling this problem. While he
doesn’t have the precise control over individual neurons that Fromherz has,
Potter is interested in how networks of neurons solve complex problems, and so
has thousands of neurons growing in wells etched in silicon, and 60 electrodes
that pick up the constant mass of electrical activity of the network. When
hundreds of neurons are connected together, certain behaviours start to emerge.
“I’m hoping that some of these emergent properties will be applicable to our
artificial neural networks, or perhaps to actual physical hardware,” he
says.
In Potter’s experiment, a program analyses the neurons’ electrical outputs,
looking for repeatable patterns of activity. The neurons seem to produce regular
patterns, each of which Potter arbitrarily assigns to control a specific
movement of the virtual rat. So pattern one might make it move to the right,
pattern two to the left, say. When the neurons are active, the rat moves about
according to the sequence of patterns they generate.
Potter also provides the neurons with feedback. Information about the rat’s
movements and “sensory” signals—from a kind of artificial touch and vision
system—is sent back to the neurons via the electrodes. The neurons respond
by changing their firing patterns and connections—making the rat move
differently (see Diagram).
These changes should reveal just how adaptable
the brain is. Given the right feedback, Potter hopes that his artificial
animal—he calls it an animat—will become familiar with its virtual
surroundings and that this learning will be reflected in changing signals and
connections within the neural network. Ultimately, if the animat bumps into a
virtual wall, it should learn to avoid the wall next time round.
Whether or not this will actually happen, it’s too early to tell. But
research in Chicago suggests he might be on the right track. Neuroscientist
Sandro Mussa-Ivaldi of Northwestern University Medical School believes he’s
already seen learning in action. Instead of using dissociated brain cells,
Mussa-Ivaldi’s team is using whole brainstems extracted from
lampreys—eel-like fish older than dinosaurs—to control a modern-day
robot. Like Potter, he’s hoping to develop an experimental system that can be
used to understand the properties of biological neural networks.
The robot has light sensors, and the electrical signal they generate is fed
into one of the neural circuits in the brainstem. The circuit generates signals
that are fed back to the robot’s wheels. This odd combination of organic brain
and inorganic body creates a robot that may move towards a light or away from
it.
Mussa-Ivaldi took a robot that had an affinity for light, and held it
motionless for five minutes with an intense light shining on its left side. When
he freed the robot into normal light, the researchers got a big surprise. It
ignored lights on the left and moved only towards lights on the right. It seemed
to have learned to override its natural affinity.
While the robots are giving Mussa-Ivaldi the opportunity to design
experiments to study that learning process, he’s also interested in how we might
imitate it, even how we might tap into the neural tissue to develop new methods
of computation. The ability to adapt, as his lamprey robot did, is what will set
apart living neural computers from conventional ones, he believes. The
challenge will be to learn enough about such behaviour to control it so that
robots learn to do useful tasks.
But there are moral issues to be resolved before we move too far down this
road. The death of animals is accepted by society because there’s no other way
of finding out the details of how neurons keep us thinking and moving, and
because it could lead eventually to treatments for brain disorders such as
Alzheimer’s and Parkinson’s diseases, perhaps even paralysis. But creating a new
type of computer or robot is clearly not reason enough to sacrifice animals. It
might be more acceptable if we could use neurons grown entirely in culture, from
stem cells perhaps. Indeed Potter has kept alive neuronal cultures for 18 months
now.
Using brain tissue within machines raises some unique questions, however,
says Paul Root Wolpe of the Center for Bioethics at the University of
Pennsylvania, and NASA’s chief bioethicist. “You are creating an organism that
by its very definition could not exist in nature,” says Wolpe. But he insists
that as long as the tissue is collected in a way that’s ethically acceptable,
without causing pain or distress, it doesn’t mean we should never build such
organisms. “What I’m saying is then we need to start to have a moral
conversation about the implications—about how far we should take them, and
what we shouldn’t do.”
That conversation is not going to be easy. For instance, how do we know if
the “creature” we have built is capable of feeling pain? What are its rights?
“What really matters here is what the creature is capable of sensing and doing,”
says philosopher and cognitive scientist Andy Clark of the University of Sussex.
The complexity of the robot’s behaviour should help us judge it. “That’s roughly
how we judge the animals we find in the world,” says Clark. Something with the
complex behaviour of a dog or a chimp would raise greater ethical concerns
something that behaved more like a worm.
Miguel Nicolelis and his colleagues at Duke University in Durham, North
Carolina, are raising ethical issues of a different type. They are probing the
brain of a living animal. It sounds gruesome, but the aim is clear. Nicolelis
wants to understand how large populations of cells interact and process
information so that he can develop interfaces that control prosthetics for
people who are paralysed. At the moment, his team is learning to control a robot
arm using signals from an owl monkey’s brain. As the monkey reaches for food,
surgically implanted microelectrodes in its cortex read the neural signals. A
computer analyses these signals, recognises patterns of activity in the brain,
and predicts the motion of the monkey’s arm. These predictions guide the robotic
arm so that when the monkey moves its arm the robotic arm moves too—in
eerie coordination.
Does this mean the monkey’s brain is being used as a computer? In a very real
sense, yes, says Ditto. And by doing these kinds of experiments, we get to
understand the brain better, which could eventually help us to control it for
other purposes.
But getting the brain to do our bidding would require feedback. For instance,
if the robotic arm was reaching for an object, sending feedback to the monkey’s
brain could help the brain learn how to do it right. Nicolelis’s team has just
finished developing a system for delivering such feedback. One is
visual—the monkey can watch the position of the robot arm displayed as a
dot on a screen. The other is tactile—the monkey’s skin can be stimulated
to provide feedback. It might even be possible to directly stimulate sensory
regions of the brain.
The monkey’s brain will have to figure out what the feedback means—just
as people with cochlear implants must learn to interpret the new spectrum of
sounds they hear. What’s crucial is that the brain will be learning about an
external environment, while the environment, represented by the computer, is
learning about the monkey’s brain signals. “It’s a two-way street. In essence
you are generating an interaction between a brain and an artificial device,”
says Nicolelis.
Mandayam Srinivasan from the Massachusetts Institute of Technology, who
worked with Nicolelis on this project, conjures up an even more fantastic
future. One day human brains might be wired up to transmit signals and `receive
feedback. Astronauts could be trained using a virtual reality system where
objects bounce as they would on Mars. Before leaving Earth the astronaut’s brain
could learn how to handle gravity on the Red Planet.
For Wolpe, the idea of tapping into brains—human or animal—raises
the biggest issues. Should we be allowed to control other brains? “If we implant
things in people’s brains that allow impulses to move from the brain out to the
world, it isn’t that much of a conceptual leap to think of impulses moving from
the world back into the brain in such a way as to affect its functioning or its
cognitive ability,” he says. “We begin to think about remotely controlling
brains, including human brains. Then who has the right or the ability to do
ٳ?”
The obvious answers—that we should never use this kind of technology to
control people’s thought processes or to turn animals into slaves—are the
easy ones, says Wolpe. But what about the grey areas, where things are not so
easily defined? “The real ethical struggle will happen over the more subtle
issues, and we don’t know what those subtle issues will be.”