BIOMIMETIC CORTICAL NANOCIRCUITS - Bio-Inspired and Nanoscale Integrated Computing - page 459

17.2.2. Justification for a Radical Change in Technology
When the hardware for interconnectivity and learning are included, along with the
possible end of the continued scaling of CMOS circuits, 1 a future CMOS cortex
would still be very large. Interconnectivity, even with some 3D interconnection
capabilities like flip-chip technology, would be a major problem, and plasticity
costly to implement. Connecting 10 4 presynaptic terminals to each neuron poses
technological challenges. Furthermore, the rigidity of conventional integrated
circuits makes plasticity virtually impossible to implement directly. We believe
that a change in technology is required to create circuits that exhibit human-like
intelligence, especially via plasticity.
Carbon nanotubes may support the scale and interconnection density of a
synthetic cortex. For this reason, we have begun preliminary studies into possible
carbon nanotube circuits that could form the basis for a synthetic cortex. These
technologies may, in the distant future, enable the construction of a reasonably
sized synthetic cortex, as we predict here. We believe that with the right
combination of manipulation techniques true 3D structures will be possible,
thus alleviating a large portion of the interconnection and scale issues.
17.2.3. Our Modeling Approach
We have begun circuit design with the excitatory synapse [15], modeling neuro-
transmitter action and ion channels. The synaptic circuits we have designed
translate higher-voltage short-duration signals to longer-duration signals of
smaller magnitude. Their behavior is chaotic [16], with present state sensitive to
initial conditions, and output behavior highly nonlinear with respect to the input
behavior. We have simulated a carbon nanotube transistor circuit model of this
neural synapse that captures, in a coarse manner, the actions of neurotransmitters,
ion channel and ion pump mechanisms, and temporal summation of PSPs. We
have chosen to focus on excitatory PSP's (EPSP's) first, and have chosen economy
of size over exact replication of waveforms, to facilitate scaling to cortical-sized
neural networks.
Our neural models focus on the ion channel, one fundamental unit of
processing, and model the complexities described here. Ion channels are molecular
structures that block or permit the flow of charge-carrying ions through the cell
membrane. Ion pumps subsequently return the neuron to a resting state by
restoring the concentration of ions relative to the extra-cellular fluid. Ion channels
may be voltage or chemically gated. Voltage-gated channels respond to membrane
potential differentials. Chemically gated ion channels open and close in response
to specific chemicals called transmitters and may be directly gated (ionotropic)
channels or indirectly gated (metabotropic) channels. The chemical process that
causes a metabotropic channel to open is called a secondary messenger cascade.
1 Recent Intel and IBM breakthroughs in dielectric and gate materials extend MOS scaling possibilities
in the immediate future [14].