Page 202 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

192 CHAPTER 9 Theory of the Brain and Mind: Visions and History

1. EARLY HISTORY
The early history of the neural network field is discussed thoroughly in Chapter 2 of
[1] and will be summarized here. It began essentially in the 1940s when advances
in neuroscience converged with the early development of digital computers. Many
scientists and engineers in that period were taken with an analogy between
computers and real brains, based on the fact that neurons are all-or-none, either firing
or not firing, just as binary switches in a digital computer are either on or off. This
analogy stimulated the development of the science of cybernetics [2] and the first
significant article about the logic of networks of neurons [3].
McCulloch and Pitts [3] proved mathematically that a network of all-or-none
neurons could be constructed so that one neuron in the network fires selectively in
response to any given spatiotemporal array of firings of other neurons in the network.
While the all-or-none nature of the network units and the focus on single neurons were
oversimplifications, some of McCulloch and Pitts’ constructions anticipated themes
that are common in more recent neural networks. For example, some of their networks
presage the distinction developed in popular later models between input, output, and
hidden units [4]. The input units react to particular data features from the environment
(e.g., “cold object on skin,” “black dot in upper left corner,” “loud noise to the right”).
The output units generate particular organismic responses (e.g., “I feel cold,” “the
pattern is a letter A,” “walk to the right”). The hidden units are neither input nor output
units themselves but, via network connections, influence output units to respond to
prescribed patterns of input unit firings or activities. The input-output-hidden trilogy
can at times be seen as analogous to the distinction between sensory neurons, motor
neurons, and all other neurons (interneurons) in the brain.
Also, one of McCulloch and Pitts’ examples involved constructing a network to
feel heat when a cold object is removed. This is an example of a contrast effect,
which has been modeled more recently using either opponent processing via
transmitter depletion (Grossberg, 1972a, [5,6]), or temporal differences based on
reward prediction errors [7e11].
Networks of the McCulloch-Pitts type do not include learning: their responses
remain uniform over time, although some of them store a memory of previous
stimulation via reverberation. Yet under the influence of psychology, other researchers
drew the distinction between mechanisms for short-term memory (STM) and for long-
term memory (LTM; [12,13]). Hebb [12] expressed the emerging consensus in the
field that reverberatory loops (say, between cortex and thalamus in the brain) could
be a mechanism for STM, but that LTM required structural or biochemical changes
in synapses or neurons.
Hebb sought a neural basis for classical conditioning. He predicted that paired
presynaptic and postsynaptic stimuli would lead to increases in the strength of a
synapse, anticipating later findings in neuroscience about synaptic plasticity
[14,15]. His idea was widely accepted in both neuroscience and neural modeling,
but some neuroscientists and modelers expressed concern that Hebb’s law could
lead to instability and needed to be counteracted by a law for reducing synaptic