Biomedical Engineering Reference
In-Depth Information
The number of patterns that can be discriminated in this way is related to
the number of combinations of neurons that can be formed by the network, and
is far larger than the number of fixed points in a static network of the same size
(23). For computations with small networks—and with neuromorphs we are
limited to networks that are small compared to biological networks—a dynami-
cal network of this type is capable of much greater capacity as a memory or for
pattern recognition. We may expect that networks exploiting this kind of dy-
namical behavior will play a significant role as applications are realized.
5.
SENSORIMOTOR DEVELOPMENT IN A
NEUROMORPHIC NETWORK
In this section we present an experimental system of simulated neuro-
morphs that develops sensorimotor capabilities. It was studied to find ways of
using neuromorphs for controlling the behavior of an autonomous vehicle, but
the same principles could be applied in a neuromorphic prosthetic where arbi-
trary patterns of spikes need to be interpreted and used to control artificial effec-
tors or the patient's own muscles.
We have been investigating ways in which development could be per-
formed in neuromorphic circuits using principles that are neurobiologically
plausible. In animal nervous systems, the patterns of neural activity that emerge
during development, influenced both endogenously and exogenously, play a
vital role in establishing normal processing. While there is no compelling reason
to make the development of a neuromorphic "brain" or prosthetic neurobiologi-
cally plausible, using mechanisms evolved over eons would appear to be a good
strategy.
If one can point to a single principle underlying both the development of
neural connections and the subsequent adjustment of connections during learn-
ing, it would be the rule ascribed to the psychologist, D.O. Hebb. In essence, the
rule requires the strengthening of the connection between a sending, or pre-
synaptic neuron and a receiving, or postsynaptic neuron if the two fire together
close together in time. Otherwise, weakening of the connection may result, de-
pending on which of several formulations of the rule are applied. Recent neuro-
biological work has shown that the relative timing of pre- and postsynaptic
spiking is indeed responsible for potentiation and depression of synaptic efficacy
or weight (13,25).
To illustrate the neuromorphic approach, we present results on a simulation
of a network of spiking neurons that approximate the behavior of our silicon
neuromorphs. Previous theoretical work showed that the application of Hebb
rules can lead to the development of adaptive connections and realistic sensory
receptive fields (9). Although the intent here is to develop a controller for an
autonomous vehicle, the results demonstrate for a system of spiking neurons
how inputs from sources with very different characteristics can combine auto-
matically to generate order and useful functionality—in this case directional
