BIOMIMETIC CORTICAL NANOCIRCUITS - Bio-Inspired and Nanoscale Integrated Computing - page 461

neural model on specific applications [30]. Sato et al. [31] use stochastic logic to
obtain analog behavior from digital circuits. Chen and Shi [32] use pulse-width
modulation. Linares-Barranco et al. [33] describe a CMOS implementation of
oscillating neurons. Fu et al.

[34] present thin-film analog artificial neural

networks.
The master's thesis by Chao describes a basic CMOS neuron, designed by
Parker, with learning capabilities [35]. An 8-transistor CMOS synapse [36] is close
in scale and nature to our current synapse circuit. Analog synapses have been
reported by Pinto et al. [37] and Lee et al. [38] and a phase-lock loop synapse
has been reported by Volkovskii [39]. Elias [40] has performed modeling of
dendritic trees that are similar to our models. The primary difference involves our
use of transistors and his use of resistors and capacitors for dendritic computa-
tions. In addition, his simple synapses involve single transistors. Noteworthy
neurons capable of learning have been proposed [35, 41-47]. Koosh and Good-
man [42] put a digital computer in the loop for training, control and weight
updates, and the neural network is entirely analog, a style realized by several
research groups.
While many neurons in the literature have some biomimetic features (e.g., [22,
26, 40, 50]), the complete range of neural variations has not been implemented
in a single model or even in the variety of neuron models distributed throughout
the research community. The most important research on neural interconnec-
tivity is Boahen's neurogrid project, with roots in the pioneering research by
Mahowald [23].
Commercial minimally biomimetic neural networks incorporating learning
are available, and in use by the high-energy physics community. However, in
contrast to our model, there is little correspondence in the majority of these
models between individual circuit elements and specific physiological mechanisms
in the biological neuron. The correspondence between specific biological mechan-
isms and circuit elements allows us to vary synapse behavior easily with control
inputs. This and our choice of carbon nanotube technology differentiates us from
related work.

17.3.2. Nanotechnology
While nanowires and nanotubes form the basis for a significant volume of
nanotechnology research, the focus here is on the nanotechnologies incorporated
into the neural circuits described in this chapter. Nanowires [11, 12, 51-55] are
one-dimensional structures with diameters typically around 10 l nm and lengths
up to tens of microns. Zhou's group is one of the leaders working on the synthesis
and device applications of various semiconductive nanowires such as In 2 O 3
(shown in Fig. 17.1a), SnO 2 , ZnO, Si and Ge. These materials can be very good
electrical conductors, and their conductance can be varied over many orders of
magnitude by applying a gate voltage bias [55]. Zhou has demonstrated high-
performance transistors based on many kinds of nanowires that can work as the
active channel in artificial neurons.