Lippard.Multics 1986-10-30 15:52:46 mst Thu
Subject:  Neuron Chips
Date:  Thursday, 30 October 1986 15:49 mst
From:  Paul Dickson <Dickson>
To:  {mbx >udd>m>jjl>misc>misc}

The following isn't humor.  But I thought it was a worth-while miscellaneous
computer related news item.

          Paul Dickson

_________________________________

[From November 1986 Issue of BYTE]

 "Neuron" Chips Emulate Brain Cells, Hold Promise of Much Faster Processors


          Perhaps the most complex data communications system of all is
the neural network found in even the simplest of animals.  These extremely
complex interconnecting structures allow most animals to perform
pattern-recognition tasks that even today can be approximated only by the
largest supercomputer.

          One division of AT&T Bell Laboratories (Holmdel, NJ) is trying
to develop better pattern-recognition capabilities by emulating very
simple neural networks on integrated circuit chips.  Some of this work has
evolved from studies at another division of Bell Labs concentrating on the
neural networks of slugs.

          So far, the researchers have designed three electronic neural
network (ENN) chips.  The first, with 22 neurons and 22 input channels,
was successfully tested last March.  A second chip, with 54 neurons and
input channels, was successfully tested in September.  This chip contains
almost 3000 synapses, connecting each input with each electronic neuron.
Each synapse is a programmable resistor, which can be adjusted during
EEN's learning period.

          Design work has just been completed on a 256-neuron chip, which
will be built using a combination of standard CMOS and electron beam
lithography.  A fourth chip, with 512 neurons, is in the design stages.

          The 54-neuron chip has been tested using simple search tasks.
The chip was first "taught" a list of names.  It was then shown a new name
and asked which name on the list was most like it.

          Because of the inherently parallel structure of the network, it
can perform such tasks much faster than a conventional processor.
According to Larry Jackel, head of Bell Labs' Device Structure Research
Department, ENN chips can perform these tasks 100 to 1000 times faster
than a conventional computer, and perhaps 10 to 30 times faster than
special-purpose hardware.  The response time of each electronic neuron is
only 400 nanoseconds, much faster than a biological neuron.

          The relatively small size of the resistors used on these chips
makes possible a very high chip density, higher than that associated with
conventional transistor-based circuits.  Complexity of future chips may be
somewhat limited, however, by the number of input connectors that can be
added.  Multiplexing the connectors may ease this problem but then might
cause its own bottleneck.  Jackel says his group is also working on
combining the EENs in hierarchically structured gangs.