578
14 Neural Networks and Brains
computer. 1 Thus, our electronic computers, made from interconnected ‘logic gates’,
with sufficient capacity, can perform arbitrarily close to any other constructed brain
based on logic elements. To the extent that our ‘wet brains’ act deterministically, we
expect our deterministic electronic brains can act as our biological brains do. This
is Turing’s theorem: Any deterministic computer is capable of producing the same
output as any other deterministic computer, irrespective of how it is constructed.
Our brains do not contain neurons acting as digital logic gates. But to a good
approximation, the input and output of a biological neuron can be digitized, in that
the neurotransmitters are made from individual molecules, and their effect on the
receiving neuron can be quantized. Our brains can be well approximated by a set of
idealized neurons.
However, brains that are constructed to act like human brains can be made
from electronic components. We can then use the “Turing Test” to see how good
an approximation we have. The Turing Test uses interrogation of the simulation
computer to see if its answers match that of our brain sufficiently that a human
cannot tell the difference.
Brains give life forms a sophisticated ability to coordinate actions, analyze and
respond to sensory inputs, control internal functions, balance objectives, anticipate
based on experience, and make plans. Besides analysis, there are two other major
advancements for life forms with brains: Learning and memory capacity. A brain
can be flexible, in that it is capable of learning. Learning is also a property of
individual protozoan, but brains add elaboration to the capacity to learning. Learning
strategies for survival, from experience and from communication with parents and
others, relieves the DNA from having to store such information, which often is
ephemeral, and therefore better stored outside the DNA. In fact, for mammals,
learning accounts for billions of times more information than that stored on
their DNA.
The fundamental element of a biological brain is a neuron, abstractly represented
in (Fig. 14.1). Biological networks of neurons learn by having activity affect their
response thresholds and their innervations. 2
Neuron inputs are carried on conduction lines called dendrites; the output on a
line called an axon, which divides into branches called axon terminals and ending in
terminal bulbs. The neuron can be either excitatory or inhibitory in action on another
neuron, muscle, or gland. The strengths of the inputs, {E i , I j } and the output
threshold T are ‘plastic’, in that conditions and history can affect these numbers. The
connection locus between terminal bulbs and dendrites of another neuron or muscle
are called ‘synapses’. The synapse and the state of the neural endings determine the
strengths {E i } and {I j }. Neurons can also sprout new dendrites and axon terminals.
Neurons require a recovery time before they are able to fire again, called the
‘refractory period’. At the end an absolute refractory period (about 2 ms), the neuron
1 Alan Turing, Computing Machinery and Intelligence, Mind LIX (236): 433–460, (October 1950),
doi:10.1093/mind/LIX.236.433
2 Donald O. Hebb, Organization of Behavior, [Wiley & Sons] (1949).
14 Neural Networks and Brains
computer. 1 Thus, our electronic computers, made from interconnected ‘logic gates’,
with sufficient capacity, can perform arbitrarily close to any other constructed brain
based on logic elements. To the extent that our ‘wet brains’ act deterministically, we
expect our deterministic electronic brains can act as our biological brains do. This
is Turing’s theorem: Any deterministic computer is capable of producing the same
output as any other deterministic computer, irrespective of how it is constructed.
Our brains do not contain neurons acting as digital logic gates. But to a good
approximation, the input and output of a biological neuron can be digitized, in that
the neurotransmitters are made from individual molecules, and their effect on the
receiving neuron can be quantized. Our brains can be well approximated by a set of
idealized neurons.
However, brains that are constructed to act like human brains can be made
from electronic components. We can then use the “Turing Test” to see how good
an approximation we have. The Turing Test uses interrogation of the simulation
computer to see if its answers match that of our brain sufficiently that a human
cannot tell the difference.
Brains give life forms a sophisticated ability to coordinate actions, analyze and
respond to sensory inputs, control internal functions, balance objectives, anticipate
based on experience, and make plans. Besides analysis, there are two other major
advancements for life forms with brains: Learning and memory capacity. A brain
can be flexible, in that it is capable of learning. Learning is also a property of
individual protozoan, but brains add elaboration to the capacity to learning. Learning
strategies for survival, from experience and from communication with parents and
others, relieves the DNA from having to store such information, which often is
ephemeral, and therefore better stored outside the DNA. In fact, for mammals,
learning accounts for billions of times more information than that stored on
their DNA.
The fundamental element of a biological brain is a neuron, abstractly represented
in (Fig. 14.1). Biological networks of neurons learn by having activity affect their
response thresholds and their innervations. 2
Neuron inputs are carried on conduction lines called dendrites; the output on a
line called an axon, which divides into branches called axon terminals and ending in
terminal bulbs. The neuron can be either excitatory or inhibitory in action on another
neuron, muscle, or gland. The strengths of the inputs, {E i , I j } and the output
threshold T are ‘plastic’, in that conditions and history can affect these numbers. The
connection locus between terminal bulbs and dendrites of another neuron or muscle
are called ‘synapses’. The synapse and the state of the neural endings determine the
strengths {E i } and {I j }. Neurons can also sprout new dendrites and axon terminals.
Neurons require a recovery time before they are able to fire again, called the
‘refractory period’. At the end an absolute refractory period (about 2 ms), the neuron
1 Alan Turing, Computing Machinery and Intelligence, Mind LIX (236): 433–460, (October 1950),
doi:10.1093/mind/LIX.236.433
2 Donald O. Hebb, Organization of Behavior, [Wiley & Sons] (1949).