🚀 go-pugleaf

Source code for the latest version of the adaptive logic network (ALN)
simulation package, atree release 2.6 for Windows 3.x, was made
available via anonymous ftp from menaik.cs.ualberta.ca [129.128.4.241]
on July 3, 1992 (in the file pub/atre26.exe).

ALNs are multilayer perceptrons based on boolean logic rather than
arithmetic operations.  By means of encodings, they can also handle
continuous values as inputs and outputs.

The simulated (real) hardware resulting from ALN training on a typical
problem is a tree of about ten to twenty layers of NAND operations
(transistors).  Once the inputs are in place, such hardware can be
expected to perform a basic pattern recognition task in under 20
nanoseconds, faster than any other kind of digital system.  Besides
being useful where extreme speed is required, ALNs can be used in data
analysis applications where complex non-linear models are appropriate.

The atree 2.6 package includes a demo showing the advanced character 
recognition capabilities of adaptive logic networks.  US experts in the OCR 
field have said the demo is "quite impressive".  One reviewer even said it 
was "incredible".

Two weeks have passed since the announcement on comp.ai.neural-nets,
and over 274 direct copies have been taken via Internet and 175 via
CompuServe.  Altogether, about 4000 direct copies have been taken of
various versions of atree C and C++ source code, made freely available
over the past year.  Researchers around the world are using the atree
code to achieve successes in areas such as beef grading based on
ultrasound images, classification of particles from an accelerator,
and design of walking prostheses for patients with spinal cord damage.

ALNs are soon going to drive the typical, backprop-trained multilayer
perceptrons from the artificial neural network arena.  The first
reason is that ALNs are inherently faster in parallel hardware by
several orders of magnitude, hardware that may be as simple and
inexpensive as an off-the-shelf programmable logic chip. The second
reason is that if less than complete parallelism is used, non-parallel
logical operations can be done "lazily", omitting unnecessary work.
The third is that a safe design methodology, based on forcing
piecewise monotonicity, can be carried out for ALNs. The "arithmetic"
MLPs, whether trained by backprop or not, are much slower, have much
worse behavior as the size of the task grows, and can't support a safe
design methodology.  In short, they have already been made obsolete by
ALNs.

Although quite a few Canadian university researchers are ftp-ing the
atree 2.6 code, it is disappointing that few Canadian companies are in
the lists we have kept of sites taking the code on Internet.  On the
other hand, we are getting plenty of email from foreign companies that
are interested in ALNs.  Canada now has, in ALNs, the best artificial
neural network technology in the world, but the benefits will go to
companies that use it, not to those who ignore it.


--
***************************************************
Prof. William W. Armstrong, Computing Science Dept.
University of Alberta; Edmonton, Alberta, Canada T6G 2H1
arms@cs.ualberta.ca Tel(403)492 2374 FAX 492 1071