B
uilt around biological-like informa-
tion-processing structures, artificial
neural networks (ANNs) have demon-
strated their power and usefulness in areas
where identification and adaptability are most
crucial. After sufficient training with a given
set of problem data (using an arbitrary learn-
ing rule), ANNs can independently form inter-
nal representations (models) of the data’s
underlying phenomenon. Current neural net-
work applications are in pattern recognition,
function approximation, optimization prob-
lems, and forecasting. For instance, due to
their associative capability and after fine-tun-
ing by sufficient training, ANNs can easily rec-
ognize geometrical patterns and plain
characters.
The usefulness of ANNs arises from their
adaptive and associative structure and their
parallel and distributed model of computa-
tion. Also, these networks have proven fault
tolerant, as the failure of few elements (in
large ANNs and in certain layers) does not
significantly affect the network’s functional
characteristics or performance. Furthermore,
they can be easily designed to incorporate
an even larger fault tolerance capability to
achieve higher reliability.
1
ANNs typically have a large number of
highly interconnected processing elements,
which constrains their hardware implemen-
tation and network architecture. To obtain
high density and processing-element con-
nectivity, most ANN architectures employ
some kind of resource sharing. A multilayer
structure allows processing elements to
share communication lines and control cir-
cuitry, so we chose to develop a multilay-
ered neuroprocessor based on pipelined
time-step interneural communication.
Other pulse stream architectures, such as
those proposed by Murray et al.,
2
use the
rate of impulses to represent the neural
states. Our architecture, however, represents
the neural states through pulse amplitude
modulation. Also, in our design, the analog
computation consists of integrating the bipo-
lar input pulses corresponding to the exci-
tatory and inhibitory activations. The
processing-element analog path consists of
CMOS transmission gates controlled by
buffered signals originating from the neuro-
processor control unit. The processing ele-
ments broadcast their output states, held on
a local capacitor, during their assigned time
slots. These features are desirable to meet
the design goals of versatility, high density,
high connectivity, and scalability.
Neuroprocessor architecture
Figure 1 shows the topology of our mul-
tilayered neural network. Using time-multi-
plexed pulse stream communication, the
processing elements of adjacent layers share
communication lines along which they
broadcast their output states during their
assigned time slots. The time slot assignment
is done by programming a cyclic shift regis-
ter that holds the processing-element
addresses in the programmed sequence.
Our design uses bit-serial downloading of
the network weights, which considerably
reduces the addressing hardware require-
ments. The analog processing-element out-
puts are available for external monitoring or
storage through the layer communication
lines. Global clocks and signals are buffered
and routed to each processing element.
58 IEEE M icro 0272-1732/97/$10.00 © 1997 IEEE
For efficient chip area
use and operating
speed, this multi-
layered pulse stream
neuroprocessor
design uses a mixed
analog-digital
implementation.
A T IME -M ULTIPLEXED
R ECONFIGURABLE
N EUROPROCESSOR
Fadi N. Sibai
Intel
Sunil D. Kulkarni
Voice-Tel Enterprises
.