308 CATALYZING INQUIRY
- The disparity between the information processing that can be done by digital computers and that
done by nervous systems is likely to be a consequence of the different way in which nerve tissue
represents and processes information, although this representation is not understood. - At the device level, nervous tissue operates on physical principles that are similar to those that
underlie semiconductor electronics.^11 Thus, differences between neural and silicon computation must
be the result of differences in computational architecture and representation. It is thus the higher-level
organization underlying neural computation that is of interest and relevance. Note also that for the
purposes of understanding neural signaling or computation, a neuron-by-neuron simulation of nervous
tissue per se cannot be expected to reveal very much about the principles of organization, though it may
be necessary for the development of useful artifacts (e.g., neural prostheses).
Some of the principles underlying neural computation are understood. For example, neurobiology
uses continuous adaptation rather than absolute precision in responding to analog inputs. The dynamic
range of the human visual system is roughly 10 decades in input light intensity—about 32 bits. But
biology doesn’t process visual signals with 32-bit precision; rather, it uses a 7- or 8-bit instantaneous
dynamic range and adapts the visual pathway’s operating point based on the background light inten-
sity. Although this approach is similar to the automatic gain control used in electronic amplifiers,
biology takes the paradigm much farther: adaptation pervades every level of the visual system, rather
than being concentrated just at the front end.^12
There are essentially two complementary approaches toward gaining a greater understanding of
neural information processing. One approach is to reproduce physiological phenomena to increase our
understanding of the nervous system.^13 A second approach is based on using a manageable subset of
neural properties to investigate emergent behavior in networks of neuron-like elements.^14 Those favor-
ing the first approach believe that these details are crucial to understanding the collective behavior of
the network and are developing probes that are increasingly able to include the relevant physiology.
Those favoring the second approach make the implicit assumption that reproducing many neurophysi-
ological details is secondary to understanding the collective behavior of nervous tissue, even while
acknowledging that only a detailed physiological investigation can reveal definitively whether the
details are in fact relevant.
What can be accomplished by building silicon circuits modeled after biology? First, once the neu-
ronal primitives are known, it will be possible to map them onto silicon. Once it is understood how
biological systems compute with these primitives, biologically based silicon computing will be possible.
Second, we can investigate how physical and technological limits, such as wire density and signal
delays and noise, constrain neuronal computation. Third, we can learn about alternative models of
computation. Biology demonstrates nondigital computing machines that are incredibly space- and
energy-efficient and that find adequate solutions to ill-posed problems naturally.
(^11) In both integrated circuits and nervous tissue, information is manipulated principally on the basis of charge conservation. In
the former, electrons are in thermal equilibrium with their surroundings and their energies are Boltzmann distributed. In the
latter, ions are in thermal equilibrium with their surroundings and their energies also are Boltzmann distributed. In semiconduc-
tor electronics, energy barriers are used to contain the electronic charge, by using the work function difference between silicon
and silicon dioxide or the energy barrier in a pn junction. In nervous tissue, energy barriers are also erected to contain the ionic
charge, by using lipid membranes in an aqueous solution. In both systems, when the height of the energy barrier is modulated,
the resulting current flow is an exponential function of the applied voltage, thus allowing devices that exhibit signal gain.
Transistors use populations of electrons to change their channel conductance, in much the same way that neurons use popula-
tions of ionic channels to change their membrane conductance.
(^12) Adaptation helps to explain why some biological neural systems never settle down—they can be built so that when faced
with unchanging inputs, the inputs are adapted away. This phenomenon helps to explain many visual aftereffects. A stabilized
image on the retina disappears after a minute or so, and the whole visual field appears gray.
(^13) M.A. Mahowald and R.J. Douglas, “A Silicon Neuron,” Nature 354(6354):515-518, 1991.
(^14) J. Hertz, A. Krogh, and R.G. Palmer, Introduction to the Theory of Neural Computation, Addison-Wesley, Reading, MA, 1991.