Newsgroups: comp.ai
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news3.near.net!transfer.stratus.com!news.mathworks.com!uhog.mit.edu!news!minsky
From: minsky@media.mit.edu (Marvin Minsky)
Subject: Re: AXIS Machine Learning Questions
Message-ID: <1995Jun9.193221.5124@media.mit.edu>
Sender: news@media.mit.edu (USENET News System)
Cc: minsky
Organization: MIT Media Laboratory
References: <3r9ju1$rn9@condor.ic.net>
Date: Fri, 9 Jun 1995 19:32:21 GMT
Lines: 60

In article <3r9ju1$rn9@condor.ic.net> pjackson@falcon.ic.net (Philip Jackson) writes:
>Dear Marengo Media Inc.,
>
>I've read with interest (parts of) the recent thread on your claims
>for a "machine learning breakthrough". I think most readers of this
>newsgroup would welcome such a breakthrough, but I hope you will 
>understand that most readers will probably also want to reserve 
>judgment regarding whether AXIS represents a breakthrough.  Anyway,
>I have the following comments and questions:

Agreed.  However, there's also a well-known theory of " adaptive logic
nets" developed by arms@cs.UAlberta.CA (Bill Armstrong) at University
of Alberta, Edmonton, Canada.  Armstrong's scheme is an incemental
reinforcement learning scheme applied to Boolean logic trees (in
contrast to the semilinear threshold units in most ANNs).  A couple of
years ago, he had full information available via ftp at
menaik.cs.ualberta.ca:arms.  

So, as an alternative to Philip Jackson's request, I'd like a
discussion of the similarities and differences fo AXIS with ALN,
because they sound so similar.  Here's an extract from the beginning
of one of Armstrong's early (1991) tutorials on ALN:

"The nodes (or adaptive logic elements) have two input leads. The input
signals x1, x2 are boolean ( 0 or 1).  A connection "weight" to be
used for each input ( to use the term for the usual kind of neural
net) is determined by a single bit of information.  The nonlinearity,
or "squashing function", used in ALNs is just a simple threshold
(comparison to a constant).  Specifically, if b1 and b2 are boolean,
then the element computes a boolean value which is 1 if and only if

     (b1 + 1) * x1 + (b2 + 1) * x2 >= 2.

"The four combinations of b1 and b2 ( 00, 11, 10, 01) generate the four
boolean functions of two variables: AND, OR, LEFT, and RIGHT, where
LEFT(x1, x2) = x1 and RIGHT(x1, x2) = x2.  For example 1 * x1 + 1 * x2
>= 2 if and only if both x1 AND x2 are 1.

"The LEFT and RIGHT functions, in effect, serve to provide some
flexibility in the interconnections.  They disappear after training is
complete.

"Two counters in the node count up and down in response to 0s and 1s
desired of the network when the input to the node is 01 or 10
respectively.  They try to determine the best function for the node
even in the presence of noise and conflicting training data.  The
counters also disappear after training is complete."

---

(Some of us tried a few pattern recognition examples and found that ALN did
indeed work well on some of them.  Unfortunately we didn't keep any
notes.)

