Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!usc!sdd.hp.com!uakari.primate.wisc.edu!aplcen!jhunix!ins_atge
From: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards)
Newsgroups: comp.ai
Subject: Re: What Has Traditional AI Accomplished?
Message-ID: <6691@jhunix.HCF.JHU.EDU>
Date: 24 Oct 90 01:25:45 GMT
References: <1990Oct16.135631.6444@cbnewsj.att.com> <69929@lll-winken.LLNL.GOV> <3740@media-lab.MEDIA.MIT.EDU>
Organization: The Johns Hopkins University - HCF
Lines: 57

In article <3740@media-lab.MEDIA.MIT.EDU> minsky@media-lab.media.mit.edu (Marvin Minsky) writes:
>In that connection, there is wisdom in Thomas G Edwards' remarks in
><6664@jhunix.HCF.JHU.EDU>:
>>  I see a future where inductive learning by small homogeneous NNs
>>  is used in combination with more traditional AI type goal building.

>Divide-and-conquer is surely needed for circle-inside-square.  Note
>that we still don't nkow how the brain does it.

One leader in the strategy to create connectionist structures capable
of divide-and-conquer style compositional learning is Jeurgen
Schmidhuber of T.U.M.  He imagines compositional learneds utilizing
three kinds of networks.

One is a program executer, which receives as input a start
situation, a (sub)goal situation, and external senses.  This module
produces output which allows a robot to achieve the transformation
from the start situation to the (sub)goal situation.

The second structure is an evaluator, which receives a start situation
and (sub)goal situation as input.  It produces output which indicates
whether there is a program executer network which can perform the
transformation from start to (sub)goal states.

The third structure is a subgoal generator.  It receives as input a
start situation and a goal situation.  It produces a subgoal
as an output.

All networks are continually running, recurrent neural networks.

The subgoal generator is trained by applying a start and goal input to
it, and also applying the start state and output from the generator
to one evaluator network, and the output from the generator and the
goal state to another evaluator network.  The subgoal generator is
trained until the output of the two evaluator networks indicate there
is a program executer which can perform the transition from the
start state to the subgoal from the generator, and from that
subgoal to the goal (or the net is trained until a local minimum is
found, indicating that a new program executer(s) has to be
developed to solve the problem).

Obviously there needs to be much more work in this area.
It is important that connectionist researchers get as familiar
with training continually running recurrent neural networks with
both supervised and unsupervised methods as they are with 3 layer
feedforward backprop style networks.  Also they must get above 
any "me vs. them" feelings they have with symbolic AI and look carefully
at the huge amount of machine learning theory which has already been
developed.  

Ref:  J. H. Schmidhuber.  Towards compositional learning with
       dynamic neural networks.  Technical Report FKI-129-90,
       Institut fuer Informatik, Technische Universitat Munchen, 1990.

-Thomas Edwards
 (P.S.: Looking for Ph.D. programs in robotics and electrical
        engineering for fall '91)