During the last years intensive research has been performed
in the area of neural network systems and their applications.
Also some efforts have been invested exploring hardware
implementations of these neural network systems, mainly
because a good specialized hardware would exploit much more
efficiently and faster the potential of neural network based
systems. Most of the efforts in hardware are oriented towards
the construction of special purpose digital computers for
applications with neural networks (see for example [1]-[4]), and
fewer people are researching on analog implementations for
neural networks [5]-[12]. However, very few people are
looking for hardware implementations based on pulsing models
of the neurons [13]-[17]. The reason might be that at this
moment it is not sufficiently clear what the advantages of neural
systems based on pulsing neurons might be, even though the
living brains use such neurons.
On one hand, one obvious reason that justifies the use of
pulses is that, since such systems operate on the basis of
averaging principles, they are inherently more tolerant to
imperfections and nonidealhies in the components.
On the other hand, one of the biggest problems in hardware
neural network implementations is the efficient storage of the
vast amount of synaptic weight values of neural network
systems. Although we don’t address such a problem in this
paper, may be we can justify the use of pulsing neural network
systems as candidates to produce very efficient weight storage
networks.