It might seem that connectionism burst upon the unsuspecting world of cognitive
science in 1986 with the publication of Parallel Distributed Processing by
McClelland, Rumelhart and the PDP Research Group. Although this book had more
impact than any other publication in the history of connectionism, it was preceded
by more than 40 years of research into the computational abilities of networks of
simple computing elements working in parallel. This chapter will review a number of
key papers and books which are generally agreed to have influenced the direction
taken by connectionist research. The aim is to give a flavour of some earlier research,
not a systematic history. Detailed reviews of the period can be found in Levine
( 1983), Anderson and Rosenfeld ( 1988), Cowan and Sharp ( 1988) and the 1989
edition of Hinton and Anderson. Some of the papers referred to in this chapter are
difficult to find in their original form. Most have been reprinted, along with historical
and explanatory commentary, in a very useful collection of influential papers in
connectionism edited by Anderson and Rosenfeld ( 1988).

A key event in the history of connectionism was the publication in 1943 of a paper
called 'A logical calculus of the ideas immanent in nervous activity' by McCulloch
and Pitts. They demonstrated that a network of simple computing units, operating in
parallel, could perform logical operations. A crucial aspect of their modelling was
that the properties of the computing units were based on those of the neuron (as they
were believed to be at the time the paper was written). This approach has been
followed (up to a point) by contemporary connectionism. It was not followed by the
mainstream of cognitive function modellers, for whom the aim was to model what
the cognitive system achieved, not how it did it.

McCulloch and Pitts's 'neurons' had three properties: They were binary state

-314-