June 29, 2000
WHAT'S NEXT
An Electronic Circuit That Draws Its Inspiration From Life
By ANNE EISENBERG
o one knows exactly how the
brain processes information,
but scientists tinkering with
biologically inspired electronics
have begun trying to imitate in silicon what nature does so well in the
flesh.
An eclectic group of researchers
whose disciplines combine electrical
engineering, computer design, physics and neuroscience have built a
circuit of 16 artificial neurons with a
system of connections, or synapses,
designed to mimic the way the cerebral cortex computes: neither solely
analog nor solely digital, but a combination of both. The hybrid circuit,
built with transistors on a chip, is
based on theoretical models of the
brain as well as on empirical results
of animal studies.
A description of the brain-inspired
silicon circuit appeared in the June
22 issue of the journal Nature.
"It is a very simple circuit, but it
represents a good understanding and
a good starting point," said Dr. Chris
Diorio, a professor in the department
of computer science and engineering
at the University of Washington in
Seattle and one of the authors of a
Nature article commenting on the
circuit work.
"Now that we have this primitive
circuit, we can use this understanding to build larger-scale biologically
inspired circuits," Dr. Diorio said.
Such circuits may one day help
process auditory and visual information for robots as well as the output of
bionic implants to help the blind see
and the deaf hear.
Retinal chips may
someday be used as pre-processors
that feed into such circuits, just as
cochlear implants may one day be
linked to these circuits to aid hearing.
| |
|
A digital-analog hybrid
based both on
theoretical models of
the brain and on animal
studies.
| |
| |
The circuits, which draw little
power, may also be useful in the
future world of ubiquitous computing, when chips in things like mailboxes and toasters will have to interact with people in a way that uses
battery power efficiently.
The new circuit is unusual in that
both analog and digital features coexist simultaneously in it. "Our specific advance is a hybrid analog-digital circuit," explained Dr. Rahul
Sarpeshkar, a professor at the Massachusetts Institute of Technology
and a co-author of the paper. "Other
people have built hardware models
of neural networks. What's new is the
particular way we've used feedback
to compute in a hybrid fashion, neither purely analog or digital, but an
intimate mixture of both."
That design adaptation is significant because the kind of signaling
the brain uses is thought to be very
different from that used in digital
computers.
"Philosophers and psychologists
have long noted that human perception has both analog and digital characteristics," said Dr. Sebastian
Seung, a professor at M.I.T. and a co-author of the research report.
While a nerve cell can trigger a
nerve cell downstream to turn on or
off, a digital kind of information
transfer, analog interactions among
nerve cells involve graded effects.
"One neuron can make another active or inactive," Dr. Seung said,
"but the intensity of the activity varies in a continuous way."
Perception must make sharp, digital, yes-no distinctions like, Is that a
dog or a cat? But it also has to make
graded, analog distinctions, like identifying various shades of gray.
"We show that these operations
are not mutually exclusive, but can
coexist," Dr. Seung said, "suggesting
that the brain's microcircuitry may
be ideally matched to the dual nature
of perception."
The work on the circuit was accompanied by the development of a
mathematical theory describing how
analog amplification and digital selection can coexist in the same silicon circuit.
The theory and the circuit were
developed over many years by Drs.
Sarpeshkar, Seung and Richard H. R.
Hahnloser; from 1996 to 1999, they
collaborated at Lucent Technologies'
Bell Labs in Murray Hill, N.J., but
they are now at M.I.T. The two other
co-authors are Drs. Misha A.
Mahowald and Rodney J. Douglas,
both of the Institute of Neuroinformatics in Zurich; Dr. Mahowald is
now deceased. Much of the basic
research was sponsored by Lucent,
for whom both Drs. Seung and Sarpeshkar consult.
Dr. John Wyatt, a professor at
M.I.T. who has worked on retinal
implants that can communicate directly with the nervous systems of
blind patients, said that the new type
of circuit might be useful in the future because it might make it easier
to build devices that would pay attention to only the stimuli of interest,
ignoring any surrounding hubbub.
| |
|
A beginning step
toward devices that
may one day help the
blind see and the deaf
hear.
| |
| |
"What's interesting about this circuit," Dr. Wyatt said, "is that in its
combined analog-digital way, it attempts to select the inputs that are
the most important -- to pay attention -- and then to amplify these
inputs in an analog way while suppressing others."
Circuits that are completely digital find it hard to make those kinds of
choices, he said.
"This is an important idea to recreate in circuitry
because right now we don't have
much of a sense of where to focus,
when to shift," Dr. Wyatt said.
"Cochlear implant patients, for example, have enormous difficulties in
noisy environments, where irrelevant background sounds can easily
dominate the speech they wish to
listen to." Attention-selection circuitry might eventually be important in
helping people choose a particular
conversation.
The problem of how to design devices that can pay attention to the
right things is a stumbling block in
trying to build devices to handle
sight or hearing. "Right now the
brain seems an awful lot smarter
than anything we can build to make
sense of sensory data," Dr. Wyatt
said.
The efficiency of the biologically
inspired circuit may also prove important in for uses other than bionic
implants.
"One of the arguments for
making special-purpose hardware is
that it does the job in a much more
efficient way," Dr. Seung said. The
authors created special hardware
rather than writing a software program than simulated the circuit in
part for this reason.
"If you want
something that runs on extremely
low power, now for bionic implants
and in the future, when every device
has an embedded computer, the battery problem will be significant."
While the circuit shows great
promise, Dr. Rajesh P. N. Rao, an
associate at the Salk Institute for
Biological Studies and co-author of
the Nature article discussing the circuit research, pointed out that much
remained to be done.
"Right now, the chip is hard-wired," he said.
"In the brain,
though, the neurons are constantly
adjusting the connections between
themselves -- for example, when
children are learning a new language
or learning to recognize their mother." What the biologically inspired
designs of the future will need to do,
he said, is to build circuits with connections that can organize themselves the way the brain cells do,
turning the volume up or down on
nearby links. Those kinds of circuits
will come closer to imitating the
adaptability offered by biology.
What's Next is published on Thursdays in the Circuits section. Click here for a list of links to other columns in the series.
