Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!linus!philabs!cmcl2!harvard!bu-cs!jam
From: jam@bu-cs.UUCP (Jonathan A. Marshall)
Newsgroups: net.ai
Subject: Re: neural networks
Message-ID: <612@bu-cs.UUCP>
Date: Wed, 14-May-86 16:44:09 EDT
Article-I.D.: bu-cs.612
Posted: Wed May 14 16:44:09 1986
Date-Received: Fri, 16-May-86 06:45:28 EDT
References: <837@mhuxt.UUCP> <175@sdics.UUCP> <13586@ucbvax.BERKELEY.EDU> <538@bu-cs.UUCP> <1583@mtuxo.UUCP>
Reply-To: jam@bu-cs.UUCP (Jonathan Marshall)
Organization: Boston Univ Comp. Sci.
Lines: 75

In article <1583@mtuxo.UUCP> orsay@mtuxo.UUCP (j.ratsaby) writes:
> In article <538@bu-cs.UUCP> jam@bu-cs.UUCP (Jonathan Marshall) writes:
>> 
>> Stephen Grossberg has been publishing on neural networks for 20 years.
>> He pays special attention to designing adaptive neural networks that
>> are self-organizing and mathematically stable.  Some good recent
>> references are:
>>  .
>>  .
>>  .
>> If anyone's interested, I can supply more references.

> I would like to ask you the following:
> From all the books that you read,was there any machine built or simulation
> that actually learned by adapting its inner structure ?

TRW is building a chip called the MARK-IV which implements some of
Grossberg's earlier adaptive neural networks.  The chip basically acts
as an adaptive pattern recognizer.

Also, Grossberg's group, the Center for Adaptive Systems, has
simulated some of his parallel learning algorithms in software.  In
particular, "masking fields" have been applied to speech-recognition,
the "boundary contour system" has been applied to visual pattern
segmentation, and other networks have been applied to symbolic
pattern-recognition.

> if so then what type of information was learned by the machine and in what
> quantities ? what action was taken to ask the machine to "remember" and 
> retrive information ? and finally , where are we standing today,that is, to
> your knowledge, which is the machine that behaves the closest to the
> biological brain ?
> I would very much apreciate reading some of your thoughts about the above,
>      thanks in advance.      joel Ratsaby      !mtuxo!orsay

The network simulations learned to discriminate patterns based on
arbitrary similarity measures.  They also performed associative
learning tasks that explain psychological data such as "inverted U,"
"overshadowing," "attentional priming," "speed-accuracy trade-off,"
and more.  The networks learned and remembered spatial patterns of
neural activity. The networks then later retrieved the patterns, using
them as "expectation templates" to match with newer patterns.  The
degree of match or mismatch determined whether (1) the newer patterns
were represented as instances of the "expected" pattern, or (2) a fast
parallel search was initiated for another matching template, or (3)
the new pattern was allocated its own separate representation as an
unfamiliar pattern.

One of Grossberg's main contributions to learning theory has been the
design of self-organizing associative learning networks.  His networks
function more robustly than most other designs because they are
self-scaling (big patterns get processed just as effectively as small
patterns), self-tuning (the networks dynamically adjust their own
capacities to simultaneously prevent saturation and suppress noise),
and self-organizing (learning occurs within the networks to produce
finer or coarser pattern discriminations, as required by experience).
Grossberg's mathematical analyses of "mass-action" systems enabled him
to design networks with these properties.

In addition, his networks are physiologically realistic and unify a
great deal of otherwise fragmented psychological data.  Read one or
two of his latest papers to see his claims.

The question of which _machine_ behaves closest to the biological
brain is not yet appropriate.  The candidates I know of are all
software simulations, with the possible exception of the TRW Mark-IV,
which is quite limited in capacity.  Other schemes, such as Hopfield
nets, are not mass-action (in the technical sense) simulations, and
hence fail to observe certain kinds of local-global tradeoffs that
characterize biological systems.

However, the situation is hopeful today.  More AI researchers have
been recognizing the importance of studying biological systems in
detail, to gain intuition and insight for designing adaptive neural
networks.