Continuous learning neural networks

Arthur T. Murray uj797 at victoria.tc.ca
Thu Jul 27 22:25:15 EST 2000


Troy Kelley, tkelley at erols.com, wrote on Thu, 27 Jul 2000:

> Can anyone out there point me toward research of NNs
> which learn and can respond continuously?  In other words
> they are capable of being trained and responding at the
> same time, not during two separate phases...

ATM:  
http://www.geocities.com/mentifex/mind4th.html Mind.Forth PD AI
is a non-traditional neural net which does not differentiate
between training phases and response phases.  Instead,
learning is going on all the time:  learning of new words,
and conceptualization of the new words.

Inasmuch as Mind.Forth creates a new node on a "concept fiber"
each time the concept is used, the nature of any given concept
can shift gradually over time, as the Forthmind "learns."

Mind.Forth is still so primitive that each syntactic structure
has to be hand-coded into the AI as a loop of parts of speech
to be called one-by-one in the generation of a sentence.

However, it is almost trivial to program a learning algorithm
which functions like the mind of a human baby, trying out
various novel syntax loops and seeing if they meet with
approval (i.e., are "ratified") by parents and others.

Do you know what you get when you extend a loop over time?
A spiral results, a kind of coil, with the same nodes
for insertion or deletion passing by repeatedly for review.
The insertion or deletion of a feature is called: learning.

> Thanks
> Troy Kelley
> Army Research Laboratory

--
http://www.geocities.com/mentifex/theory5.html : AI Theory
http://www.geocities.com/mentifex/acm.html : Step-by-step AI






More information about the Neur-sci mailing list