COPElAND & PROUDfOOT | 311
incidental noise: other medical applications include the detection of lung nodules and heart
arrhythmias, and the prediction of patients’ reactions to drugs. Connectionist networks are
good at recognizing objects, and are used to recognize faces and in optical character recog-
nition. Business applications include loan risk assessment, real estate valuation, bankruptcy
prediction, and share price prediction, while telecommunications applications include control
of telephone switching networks, and echo cancellation in modems and on satellite links.
Turing’s anticipation of connectionism
Modern connectionists look back to Frank Rosenblatt, who published his first of many papers
on the topic in 1957, as the founder of their approach.^3 Yet Turing had already investigated a type
of connectionist network as early as 1948, in a little-read paper titled ‘Intelligent machinery’.^4
Written while Turing was working for the National Physical Laboratory in London, the man-
uscript did not meet with his employer’s approval. Sir Charles Darwin, the rather headmasterly
director of the laboratory and grandson of the great naturalist, dismissed it as a ‘schoolboy’s
essay’.^5 In reality, this far-sighted paper was the first manifesto of the field of AI. Although it
remained unpublished until 1968,^6 14 years after his death, Turing not only set out the funda-
mentals of connectionism but also brilliantly introduced many of the concepts that were later to
become central in AI, in some cases after reinvention by others (see Chapter 25).
Educating neurons
In ‘Intelligent machinery’, Turing invented a kind of neural network that he called a ‘B-type
unorganised machine’, consisting of artificial neurons and devices that modify the connections
between the neurons (Fig. 29.2). B-type machines may contain any number of neurons con-
nected in any pattern, but are always subject to the restriction that each neuron-to-neuron
connection passes through a modifier device.
All connection modifiers have two training fibres (Fig. 29.2). Applying a pulse to one of
them sets the modifier to ‘pass mode’, in which an input (0 or 1) passes through unchanged
and becomes the output. A pulse on the other fibre places the modifier in ‘interrupt mode’, in
figure 29.2 Two interconnected B-type neurons (the circles). Each neuron has two inputs and executes the simple
logical operation of ‘not and’ (NAND): if both inputs are 1, the output is 0; otherwise, the output is 1. Each connec-
tion passes through a modifier (the black square) that is set either to allow data to pass unchanged or to destroy the
transmitted information. Switching the modifiers from one mode to the other enables the network to be trained.
Jack Copeland. All rights reserved.