¿ìè¶ÌÊÓƵ

Here’s a recipe for doing neuroscience. Take one brain and slice thinly.
Study these slices to work out how the brain’s 100 billion neurons are inter-connected
(an electron microscope may help here). Now insert microelectrodes at strategic
points in the tissue. How do all these neurons communicate electrically?
(Remember that each neuron may receive electrical impulses from up to 100
000 others, so this may take some time). Finally, retire to a quiet room
and ponder how the neural circuits you have characterised produce such things
as learning and consciousness.

Real neuroscience may be a good deal more sophisticated than this but
it is propelled by a similar philosophy. How else can science tackle something
as complex as the human brain if not by a process of deconstruction? Carver
Mead of the California Institute of Technology in Pasadena has a novel answer:
reconstruct, don’t deconstruct. He sees neuroscience as entering a new,
‘synthetic’ phase in which researchers will build artificial brains from
scratch to learn how nervous systems work.

In 1991, Misha Mahowald, one of Mead’s team in Caltech’s Computation
and Neural Systems Laboratory, joined up with Rodney Douglas, a neurophysiologist
from the University of Oxford, to put this philosophy into practice. They
built a neuron. Neuroscientists sometimes use computer simulations of neurons
to test their theories; but it was not one of these. Nor was it a neural
network of the kind so beloved by computer scientists. Rather, it was a
silicon chip designed to mimic the electrical behaviour of a real neuron.

The flow of electrical current into and out of the chip is exactly like
that running through a typical nerve cell. ‘It’s like reverse engineering
the brain,’ says Douglas, referring to the practice where a manufacturer
of, say, washing machines buys its competitor’s latest model, has an engineer
take it to pieces to see how it works, and then builds a copy.

Mahowald and the rest of the Caltech team have been in the business
of reconstructing human organs in silicon for some time, having already
made analogues of the ear and several variants of the retina. Douglas spends
several months each year at Caltech on a visiting professorship and it was
there that he and Mahowald hatched the plan for a silicon neuron. ‘There
is an aesthetic to it .. making things that actually work,’ he says. Douglas
and his colleagues at Oxford had tried modelling the behaviour of neurons
with simulations on powerful computers. This works well with a small number
of simulated neurons, but as the neural network gets larger its complexity
makes the simulation very slow and computer time can be expensive. ‘The
silicon neuron is a much more potent tool than digital simulation, and
relatively cheap,’ Douglas says.

Last year, Mahowald moved to Oxford to join Douglas’s team. They have
now built and tested a chip carrying five neurons and are designing chips
with 64 neurons or more that will be combined on a circuit board to produce
a network of 1000 neurons. They believe these silicon neurons could be more
than just a research tool: they are ideally suited to processing the electrical
output of silicon retinas to form a complete vision system.

Mahowald has already developed an array of simpler silicon neurons that
can take information from two simple retinas and calculate how far away
an object is from the different perspective of each retina-the basis of
binary vision. A further refinement would be the addition of a silicon motor
system that could steer the ‘eyes’ to look at objects of interest. Because
the neuron chips are highly reliable and use little power, such a vision
system would be ideal for robots, especially those that explore hostile
environments such as the surface of other planets.

The research is based on the notion that real neurons behave like tiny
analogue computers. That is, they receive input signals from other neurons,
combine them and produce a single output signal. The dense network of nerve
fibres that interconnects neurons in the brain means that each cell may
receive as many as 100 000 electrical inputs at any one time. These inputs
cause neurons to increase or decrease the voltage across their cell membrane.
If the voltage exceeds a certain threshold, it sparks off rapid flows of
ions through the cell membrane. These ion flows are the electrical currents
that form the basis of a neuron’s output signal.

A crucial feature of a neuron’s ion currents is that they flow through
its membranes smoothly and at a range of rates dictated by conditions inside
the neuron. This ‘analogue’ behaviour is entirely different from the workings
of conventional computers. These deal only with digital data: circuits switch
between two states, on or off, to represent digital bits of 1 or 0. What
Mead and his researchers found when they began to contemplate modelling
neurons was that silicon circuits, if used in an analogue way, resemble
neurons.

‘The thing that was quite remarkable is that we have this technology
that was developed for a completely different reason, never anticipating
its use for modelling these neural systems. But it turns out to be remarkably
close to the kind of physics you need to do the modelling,’ says Mead.

The biological machinery with which real neurons send information to
each other is complex, to say the least. But two ideas are central: one,
the ends of the output fibres of neurons are connected to other neurons
through synapses, across which neurons can ‘fire’ neurotransmitter molecules;
two, these neurotransmitter molecules can indirectly alter the voltage across
a receiver neuron’s membrane. In certain circumstances the voltage change
will trigger an electrical impulse which leaves the body of the receiver
neuron to hurtle down its main output fibre (axon). At the end of this axon,
the impulse triggers the release of neurotransmitter, so the cycle often
repeats itself.

The task for Mahowald and her colleagues was to build a silicon neuron
capable of generating a similar kind of electrical impulse. This in turn
meant mimicking the tiny protein channels embedded in the membranes of neurons
which control the flow of electrical current into and out of the cell-and
hence the voltage across its membrane. It is the movement of ions through
these channels that creates and shapes the output impulses with which one
neuron communicates with others.

The mimickry was far from easy: the membranes of neurons are blessed
with a whole zoo of ion channels, and different types come into play during
the various phases of a nerve impulse. Most are specific to certain types
of ions-some let through only sodium ions, others only potassium ions, and
still others only calcium ions. Some channels open and close in response
to a change in membrane voltage; others respond to ions or messenger molecules.
Some open and close quickly; others act relatively slowly. And so on. The
various permutations of these characteristics make for a bewilderingly complex
array of molecules.

First in the researchers’ sights was the sodium channel that supplies
the initial surge of membrane current at the beginning of a nerve impulse.
The neuron must first receive enough inputs to push its membrane voltage
above a certain threshold. But when this happens, the channels open suddenly,
letting sodium ions flood across the neuron’s membrane. The pulling of this
voltage trigger results in a chain reaction: the influx of sodium ions changes
the voltage further, more sodium channels open, and more charged ions flood
in. All of which leads to a wave of electrical activity-an action potential-that
travels down the membrane of the neuron’s output fibre. Other ion channels
act to reset the membrane voltage at the end of this impulse and adjust
the rate at which neurons fire impulses.

PUMPING ION

Real neurons, then, owe their ability to compute inputs and generate
impulses to the teamwork of a diverse set of ion channels. In the silicon
neuron these channels are represented by transistors, while the cell membrane
whose ion currents and voltage they control is represented by a capacitor,
a device generally used in circuits to store charge. ‘The capacitor is a
model of the cell membrane and it stores the voltage that is the output
of the cell,’ says Mahowald.

But in what sense can a transistor mimic an ion channel? In a digital
circuit, transistors act simply as switches, turning current on or off in
response to an input of voltage. But in an analogue circuit, such as the
silicon neuron, they behave like taps: the higher the voltage at their input
terminal, the more current can flow between their other two terminals. Mahowald
and her colleagues connected transistors (ion channels, remember) to the
central capacitor (the neuron’s membrane) in such a way that the transistors
could trigger current to leak from one side of the capacitor to the other-just
as ions pass through a real neuron’s membrane. ‘We model each population
(of ion channel) as an independent circuit and these are all connected to
this one capacitor,’ says Mahowald.

So far so good. The complication was that the transistors had to mimic
ion channels down to the very last detail. They had to stay clamped shut
while their input voltages were low; yet snap open, producing a surge of
current, as soon as the input voltage reached a certain threshold. The easiest
way of picturing this behaviour is on a graph of current against voltage,
which approximates to a stretched ‘S’ shape (Figure 1). A second hurdle
was that the transistors representing ion channels had to close down after
this initial current surge. This was vital if the final output was to be
an impulse rather than a continuous signal.

The answer Mahowald and Douglas came up with is based on a circuit of
transistors called a differential pair . This circuit takes the voltage
across the capacitor (the ‘membrane’ voltage) and converts it into an output
with the required S-shape. The output is then fed to a ‘tap’ transistor
controlling current flow across the capacitor. A second differential pair,
positioned back-to-back with the first and fitted with a time delay circuit,
closes down the tap transistor after the intial current surge. This ensures
that the output signal is a sharp pulse of voltage.

CIRCUIT TRAINING

Douglas and Mahowald fabricated all the circuits on a silicon chip using
a technique, called complementary metal-oxide-silicon (CMOS), that is commonly
used to make chips with low power consumption. Also included were circuits
that check the voltage in various areas of the chip so that the researchers
could monitor its behaviour. According to Douglas, the methods used to do
this were similar to those used by neurophysiologists to study real neurons.

When the researchers tested the neuron, it behaved like a dream. A current
of 0.20 nanoamps was fed into the ‘interior’ side of the capacitor. This
current raised its voltage but not over the threshold, so no action potentials
were produced. A current of 0.26 nanoamps, however, pushed the membrane
potential over the edge: the potential across the capacitor jumped up to
3 volts then immediately sank back down to near its baseline value. The
silicon neuron produced voltage spikes at regular intervals until the current
was stopped. This is exactly how a real neuron computes: it converts a constant
input signal – the current- into a series of voltage spikes.

But the story doesn’t end there. For the silicon neuron that Mahowald
and Douglas finally built has two other important ‘biological’ talents.
First, it can adjust the rate at which it fires voltage spikes; and secondly,
it can discriminate between different levels of input, adapting its output
accordingly. When an input current is high, a real neuron will initially
fire very rapidly and then slow down after the first few output spikes.
This ability to adapt is vital to the workings of most nervous systems.
It is one of the reasons why a bright source of light shining in your eyes
is unbearable at first but soon becomes less painful.

The silicon neuron can copy these behaviours thanks to the inclusion
of additional transistor circuits designed to mimic three types of channel
used by real neurons to adjust their output signals. The neuron’s firing
rate is adjusted by a ‘fast’ potassium channel, so called because it opens
during the early stages of a nerve impulse. It slows down the firing rate
most when the input current only just raises the membrane potential over
the threshold. The other two channels are specifically required for adaptation.
One is a calcium channel, the other a special type of potassium channel
that opens only when the amount of calcium inside a cell exceeds a certain
threshold.

If the firing rate is high, calcium builds up inside the cell, with
each electrical impulse, and eventually triggers the potassium channel to
open. The resulting influx of potassium ions slows the firing rate. Mahowald
and Douglas needed a different type of circuit to mimic this behaviour,
called an accumulator. If the membrane capacitor fires repeatedly and quickly,
the circuit gradually accumulates charge. This charge is an indirect index
of the amount of calcium in the neuron; it also provides the input to the
differential pair. As the charge builds up on the accumulator, so the voltage
across the membrane capacitor falls, slowing the firing rate.

One of the great values of the silicon neuron to neurophysiologists
is its flexibility. By varying the voltage thresholds or time lags of certain
transistor circuits, researchers can change the type of neuron being modelled.
‘The idea of the model is that whatever channel you need, you can design
a circuit for it, and you can give its current a strength which is a controllable
parameter,’ says Mahowald. ‘So you can change what kind of neuron you have
by turning on or turning off these currents to varying degrees.’

But this flexibility carries a price. Once you move to chips with more
neurons, specifying all the characteristics of all the neurons becomes
difficult-there are just not enough pins connected to the chip. The researchers’
solution is to incorporate digital memory circuits into the chip which,
at regular intervals ‘instruct’ various circuits to change their properties.

Another major problem is connecting the neurons up in a network in such
a way that the outputs from some neurons become the inputs for others.
With the five-neuron chip, the researchers added circuits which mimic synapses-the
junctions that connect real neurons. When nerve impulses reach the synapses
at the end of an axon, they fire neurotransmitter molecules across the synapses
to the membranes of target neurons. In general, if impulses arrive at a
synapse at a high rate, the neurotransmitter rapidly builds up in the gap
and the signal received by the target neuron is bigger. The researchers
again used an accumulator to mimic this build-up.

One reason the researchers want to mimic synapses is that these tiny
connections are thought to play a vital part in learning-a fundamental
characteristic of all nervous systems. The basic idea is that learning involves
changes in the efficiencies, or strengths, of synapses connecting neurons
whose activity is in some way linked with the task or knowledge being learned.
In the synapse of a silicon neuron, this is simply done by adjusting the
maximum voltage transistor at the bottom of the differential pair. This
specifies the maximum conductivity of the synapse, so increasing it allows
the synapse to carry bigger signals.

In theory, the ability to strengthen synapses in this way should make
it possible to design an array of silicon neurons capable of learning. But
there is still a long way to go. In the researchers’ five-neuron chip, neurons
have synapses that are not yet connected to each other. So to feed the output
of one neuron to the input of another, the researchers use an external connection.
With larger networks of neurons such an approach would be physically impossible,
so Mahowald and Douglas have resorted to the methods of digital electronics.

Their technique relies on the fact that the neuron circuits are much
slower than digital circuits. Impulses can thus be ‘broadcast’ in digital
form to all other neurons on the chip, but only some neurons are programmed
to detect them and convert them back into analogue form. In the large network
the researchers are planning, all the neurons on a chip are connected to
a single digital communications link called a bus. The bus will also extend
from chip to chip to link all 1000 neurons together.

Each neuron has an ‘address’, a string of digital code, to identify
itself. When a neuron generates an impulse, the signal is converted into
a digital message carrying that neuron’s address, which is then fed into
the bus. All the other neurons are equipped with address decoders which
read whatever addresses are on the bus and if a particular neuron is programmed
to receive signals from the neuron with that address, it converts the address
back into an analogue input. Because the digital bus is so fast compared
with the neurons, individual impulses can be transported to any of the 1000
neurons in this way. With this system, researchers can program the neurons
into any shape of network they want.

* * *

Soul of a silicon neuron

The secret to making silicon chips behave like neurons is a circuit
known as the differential pair. Despite its name, the circuit comprises
three transistors.

The voltage applied to the bottom transistor is fixed, and defines the
maximum current that can flow through the circuit. The voltage applied to
the right-hand transistor is fixed at the threshold voltage. The membrane
voltage is applied to the left-hand transistor. The output is derived from
the amount of current that flows into the left-hand transistor. Although
this output is flowing into the circuit, its size is the important factor.

When the membrane potential applied to the left-hand transistor is very
small, the output current is also small because the right-hand transistor
is effectively ‘full on’ and its current flow takes up almost all the capacity
allowed by the bottom transistor. But as the membrane voltage gets nearer
to the threshold voltage things begin to change. The output current suddenly
increases and becomes equal to the current from the right-hand transistor
when the membrane and the threshold voltages are equal (because the ‘pressure’
of the flows are the same). As the membrane voltage continues to increase,
the output current also increases and squeezes out the other flow. The output
current levels out as it nears the maximum set by the bottom transistor
and stays at this value no matter how high the membrane voltage goes.

But how does this output current control the ion channel transistor?
The output current is supplied by a circuit called a current mirror. Whatever
current the current mirror supplies to the differential pair, it supplies
an equal current into another circuit. The size of this mirror current is
converted into a voltage which is in turn fed to the input terminal of the
ion channel transistor. So the ion-channel transistor reacts to the membrane
voltage in the required S-shape fashion.