BIOMIMETIC CORTICAL NANOCIRCUITS - Bio-Inspired and Nanoscale Integrated Computing - page 456

As technologies become smaller and less expensive, there is the possibility of
constructing neural structures on the scale of a human brain in the foreseeable
future. Researchers are beginning to design, simulate, and construct cortical
circuits with the goal of determining the feasibility of implementing a synthetic
cortex (e.g., [1-3]). Jeff Hawkins, inventor of the PalmPilot, presents an eloquent
informal discussion of the brain [4] that motivates our biomimetic assumptions.
Therefore, we have started to examine the feasibility of building extremely large-
scale neural systems using nanotechnology.
There is strong motivation for developing synthetic cortical circuits. We
believe revolutionary changes both in technology and computing paradigms are
essential in order to extend the scope of hard computational problems that can be
solved. Synthetic brain structures could be used to solve some problems that have
eluded conventional approaches. A synthetic cortex that could solve some difficult
image understanding and speech recognition problems would be invaluable.
Robots that possessed these capabilities could care for the elderly, react in
situations too dangerous for humans, and provide intelligent support for routine
human activities like vehicle collision avoidance [5]. Biomimetic cortical circuits
could also be used as prosthetic devices [6], although there are other significant
engineering challenges to such devices.
Although the timeline for large-scale biomimetic processing is some decades in
the future, research on comprehensive models of large networks of cortical
neurons is in the early stages (e.g., [7, 8]). A future synthetic cortex could be
constructed using custom mixed-signal circuits that mimic the activites of
individual neurons or simulated using parallel processing. There are fundamental
challenges with both approaches, most critical of which are the issues of scale,
connectivity and plasticity. The scale and interconnectivity of a synthetic cortex is
daunting, with about 100 billion neurons, each possessing an average of 10,000
(and up to 100,000) distinct synapses [9]. The axon of each neuron fans out to
around 10,000 presynaptic terminals. While some connections to each neuron
originate in proximal (near) neurons, some originate in distal (far) neurons,
posing interconnection problems for the candidate modeling technologies.
Further, new synapses can form in a neuron in as little as an hour, expanding
networks of cells and creating new networks. Such plasticity appears to be
required for learning. Modeling this plasticity, coupled with nonstructural
changes to the neuron as a result of learning, further complicates solutions to
the synthetic cortex.
Constructing a biomimetic neuron poses additional challenges. The biological
synapse has a complex physiology that we believe should be modeled. One of the
complexities of neural tissue is the existence of transmitters, chemical messengers
that can decrease or increase the excitability of the postsynaptic receptors to
stimuli by the presynaptic cells, possibly by altering cell membrane conductance to
charge-carrying ions via chemically gated ion channels. A further complication
of transmitter function is via the long-term retrograde process that directly
or indirectly modulates transmitter release in the presynaptic junction, a form
of extremely local feedback. Transmitters acting via secondary messengers can