26-04-2012, 03:23 PM
Spiking neural networks
10.1.1.12.7093.pdf (Size: 450.51 KB / Downloads: 23)
Generations of artificial neurons
Artificial neural networks are already becoming a fairly old
technique within computer science; the first ideas and models
are over fifty years old. The first generation of artificial
neural networks consisted of McCulloch-Pitts threshold neurons
[15], a conceptually very simple model: a neuron sends
a binary ‘high’ signal if the sum of its weighted incoming
signals rises above a threshold value.
Biological background
Maass [19] correctly points out that it’s a bad idea to pour
water over your computer. It would most likely stop
functioning, defining a sharp contrast with the absolute need
for water all organisms in nature have. The neurons (see fig.
1a) in our brain find themselves surrounded by an artificial
ocean of salty extra-cellular fluid. Our wetware, as we call
our brain and parts of it, bears a lot of resemblance with the
wetware of creatures that still live in the ocean. Squid have
neurons up to 1.000 times larger than we do, making them
much easier to examine.
Spike coding
There are many different schemes for the use of spike timing
information in neural computation. Because of the nature of
this paper we’ll only cover two models here: the spike response
and the integrate-and-fire model. Both are instances
of the general threshold-fire model. The integrate-and-fire
model, which is very commonly used in networks of spiking
neurons, will be covered after the conceptually more simple
and general spike-response model. This model is simple to
understand and implement. However, as it approximates the
very detailed Hodgkin-Huxley model very well it captures
generic properties of neural activity [8,9].
Integrate-and-fire neurons
The most widely used and best-known model of thresholdfire
neurons, and spiking neurons in general, is the integrateand-
fire neuron [5,6]. This model is based on, and most easily
explained by, principles of electronics. Figure 3 shows schematic
drawings of both a real and an integrate-and-fire neuron.
A spike travels down the axon and is transformed by a
low-pass filter, which converts the short pulse into a current
pulse I(t-tj(f)) that charges the integrate-and-fire circuit.
Discussion
Neural structures are very well suited for complex information
processing. Animals show an incredible ease in coping
with dynamic environments, raising interest for the use of
artificial neural networks in tasks that deal with real-world
interactions. Over the years, three generations of artificial
neural networks have been proposed, each generation
biologically more realistic and computationally more powerful.
Spiking neural networks use the element of time in communicating
by sending out individual pulses. Spiking neurons
can therefore multiplex information into a single stream
of signals, like the frequency and amplitude of sound in the
auditory system [9].