|
|
Turing Neural Networks
It was in 1948 that Alan Turing wrote "Intelligent Machinery",
a fascinating investigation of different connectionist models
that would today be called neural networks. In this
experiment, the visitor can create and activate neuron-like
elements on the BioWall.
|
|
It was in 1948 that Alan Turing wrote a little-known report entitled "Intelligent
Machinery". At that time, he was employed at the National Physical
Laboratory (NPL) in London where he worked on the design of an electronic
computer - the Automatic Computing Engine (ACE). Turing never had
great interest in publicizing his ideas, so the paper went unpublished
until 1968, 14 years after his death.
Few people know that the "Intelligent Machinery" paper contains
a fascinating investigation of different connectionist models that
would today be called neural networks. It is amazing that his employer
at the National Physical Laboratory, Sir Charles Darwin, grandson
of the well-known English naturalist, dismissed the manuscript as
a "schoolboy essay". In describing randomly connected networks of
artificial neurons, Turing has written one of the first manifests
of the field of artificial intelligence (although he did not use
this term). Turing's neural networks have recently been investigated
in detail in a book.
Turing himself called his networks unorganized machines. He basically proposed
three types of machines: A-type, B-type, and P-type unorganized
machines. A-type and B-type machines are Boolean networks made up
of extremely simple, randomly interconnected NAND gates (i.e., neurons),
each having exactly two inputs (i.e., synapses) from other neurons.
The neurons are synchronized with a global clock signal. Unlike
A-type networks, Turing's B-type networks have modifiable interconnections
(basically a switch) and thus an external agent can "organize" these
machines (by enabling and disabling connections) to perform a required
job. The idea behind the introduction of B-type networks was to
open the possibility of reinforcing successful and useful links
and of cutting useless ones. His deeper motivation was to build
structures that can learn. On the other hand, the idea of organizing
an initially random network of neurons and connections is undoubtedly
one of the most significant aspects of Turing's paper.
Recently, Turing's neural networks have been implemented on the
BioWall's reconfigurable tissue. Each of the 3200 units of the machine
can be interactively configured by choosing one out of five possible
functions: (1) empty cell, (2) neuron, (3) connection, (4) synapse,
or (5) input cell. The user (the external supervisor) is invited
to discover and affect the behavior of the unorganized B-type machine
by opening and closing synapses (i.e., "organizing" the machine)
and by modifying the network's inputs. All modifications occur by
simply pressing on the respective touch-sensitive membranes. This
application is first and foremost a demonstration of Turing's neural
networks on reconfigurable hardware (to the best of our knowledge,
the first one). However, it also exemplifies the fusion of the ontogenetic
and epigenetic axes in a single artificial tissue.
For further information
Resources
|
|
|
|
|
Full view of
a network,
© Eurelios
47KB JPEG |
|
|
Partial view
of a network.
© Eurelios
54KB JPEG |
|
|
Neural network
construction.
© SAVE,
EPFL
2,845KB MPEG |
|
|
Turing neural
nets on the BioWall.
© SAVE,
EPFL
1,518KB MPEG |
|
|