Chapter 2. Prehension

Appendix C - Computational Neural Modelling 383
4. Assemblage Models. To capture the essence of behavioral
data, a conceptual coordinated control program model (Arbib,
198 1) can be written to describe the activation of schemas (units
of motor control and interactions with the environment). A
schema can be modelled as many networks (a network
assemblage) passing control information (activation lines) and
data (e.g., target location). This conceptual model can be
implemented in a language such as RS (Lyons & Arbib, 1989).
Computational network models that involve parallel distributed
processing (Rumelhart et al., 1986b) have been called connectionist
models (Feldman & Ballard, 1982) or artificial neural networks. The
key ingredient to this type of information processing is that it involves
the interactions of a large number of simple processing elements that
can either inhibit or excite each other. Thus they honor very general
neurobiological constraints, but using simplifying assumptions, are
motivated by cognitive phenomena and are governed primarly by
computational constraints (Churchland & Sejnowski, 1988). Each
element can be a simple model of a neuron, as in the third example
above, or it in itself can be a network of neurons, as in the network
assemblage example. Control theoretic models, while not usually
involving the interaction of a large number of elements, are useful for
bringing to bear the tools of modem control theory for information
processing, as was seen in Chapter 5. Models of individual neurons
are not discussed in this book, since we are more interested in how
networks of neurons can perform computations.
Parallel distributed processing is neurally-inspired, due to two
major factors (Rumelhart et al., 1986a). The first factor is time:
neurons are slow and yet can solve problems dealing with a large
number of simultaneous constraints in a relatively short time-period.
Feldman (1985) called this the ‘100 step program constraint’, since
one second of processing by neurons that work at the millisecond
range would take about 100 time-steps. The second reason is that the
knowledge being represented is in the connections, and not in ‘storage
cells’ as in conventional computers. If it is an adaptive model, a
distributed representation will allow the network to learn key
relationships between the inputs and outputs. Memory is constructed
as patterns of activity, and knowledge is not stored but created or
recreated each time an input is received. Importantly, as more
knowledge is added, weights are changed very slightly; otherwise,
existing knowledge will be wiped out. This leads to various
constraints on processing capabilities, as described in the next section.