Page 126 - Contribution To Phenomenology

CONNECTIONISMAND PHENOMENOLOGY 119

neural network. At its core, the basic idea of such a system involves
interconnected units or nodes (something like neurons perhaps) that can
be activated either by external stimuli, i.e., input, or by other units. What
makes this network a connectionist one is that each of the units is
connected to at least a few, more usually a large number of other units
which it can then activate by sending out signals. This in turn leads the
latter units to send out waves of signals to other units, including the ones
from which they have just received signals (something like the activity of
synapses), until the system ultimately "settles in" to a kind of equilibrium
state. These signals can be either activating (turn on) or inhibitory (turn
off) signals; the connections are thought of as weighted so that the
strength with which a signal is passed on from one unit to the next
depends on a factor called the weight of the connection, a factor which,
along with the activation level of the units, is a variable.^ The processing
at each unit is determined solely by its initial state and the input from
the nodes which are connected to it, and the whole system consists of
nothing more than many such nodes and their connections.
So there is nothing in the system that controls the operation of the
system as a whole (hke the Central Processing Unit in a conventional
machine), nor anything in the system that "knows" the state of the
system as a whole. Instead of proceeding in a linear fashion, the systems
employ parallel processing, that is they do not proceed one step at a
time; rather, at any given time there will be a number of units passing
along or inhibiting activation to a number of others. Furthermore, though
the states that receive an interpretation—and thus, count as "represent-
ations" of something outside the system—may be those of individual units,
more often the activations of several units at once will be what is
interpreted. Thus the term "parallel distributed processing" (PDP) has
become another name for the kind of activity carried out by these
systems and as an adjective for approaches oriented on such systems.* In
general, however, the input presented to the system is interpreted as the

^ An only somewhat longer and much clearer introduction to the basic model can
be found in Tienson 1988, 6-13.

* Compare, for instance, the title of the two important volumes by Rummelhart and
McClelland that describe one research group's experience in designing and employing
such systems, volumes which were decisive in the emergence of connectionism as a
widespread movement: D. E. Rumelhart and J. L. McLelland, Parallel Distributed
Processing (Cambridge: MIT Press, 1986).