Biomedical Engineering Reference
In-Depth Information
ing elements richly interconnected. At the present state of technological devel-
opment, the best way of emulating brains, and the behavior they generate is by
constructing "neuromorphs" (14), mimics of neurons fabricated in silicon with
VLSIs (very large scale integrated circuits), and interconnecting them with a
richness which approaches that of the central nervous system. Using current
technology, neuromorphic systems could be the brains of smart structures—
small devices that sense the real world and behave adaptively in it, or implanted
in humans to repair or extend their capabilities.
Part of the impetus for building artificial nervous systems is to gather in-
sights into how real nervous systems work (see Part III, section 5, this volume).
We argue that neuromorphic systems could foster understanding in ways that are
not easily achieved, if at all, by conventional digital computer simulations. A
common view is that in trying to understand intelligent systems a concern for
the details of implementation (i.e., the hardware) is unnecessary. We do not sub-
scribe to this view, but think that the nature of the neural machinery is closely
bound up with the solutions that have evolved to perform perceptual, cognitive,
and motor tasks. The kinds of computing operations these most naturally support
are therefore very different from those supported by a von Neumann architec-
ture. Neuromorphs, such as ours, patterned after biological neurons, depend
upon spike processing of information, giving them powerful signal processing
capabilities, indeed more powerful than the typical sigmoidal units used in arti-
ficial neural networks (10). Rather than review the field of neuromorphic engi-
neering generally, we refer the reader to some recent review articles and
collected papers (2,11,21), and instead focus on our neuromorphic system and
the approaches taken in our laboratories to some promising applications.
At least initially, neuromorphic systems are likely to be used where com-
pactness and low power consumption are at a premium, for example, as the
brains of autonomous vehicles and for neural prosthetics that interface between
the nervous system and artificial effectors or a patient's own musculature. There-
fore, in thinking about what kinds of networks to investigate, an underlying con-
sideration is that small systems must make good use of their resources. Dynamic
networks are of interest because modest numbers of interconnected neuro-
morphs should be able to store and process large amounts of information in the
transition of states they undergo. In the control of autonomous vehicles, a fast,
compact system is required that can efficiently utilize information about the en-
vironment. We are therefore exploring ways in which neuromorphic networks
can learn to behave adaptively through sensorimotor experience.
2.
THE NEURON AND THE NEUROMORPH
The design of our artificial dendritic tree (ADT) neuromorph is based on
what one might call the classical neuron as conceptualized in the mid 1950s (3).
The input structures of this neuron are the branched processes forming the den-
dritic tree, and to a lesser extent the cell body or soma from which the dendrites
