🚀 go-pugleaf

davidmc@cpsc.ucalgary.ca (davidmc) writes:

>OK, a quick (possibly dumb) question: Can ALNs be used to implement
>recurrent neural nets? If so, how? If not, then how can they be used
>to implement a TM equivalent program?

Recurrent ALNs

There is no problem making a forest of ALNs recurrent.  A problem
of predicting dynamics of a simulated autonomous robot was solved
by Allen Supynuk by taking 1740 ALNs which fed their outputs to a
1740-bit input vector in a recurrent fashion.  He didn't train them
recurrently though.

There would be no problem in making an associative memory for binary
patterns based on recurrent ALNs.  As far as I know, no one has done it.

Turing machine

Not that I think it's the way to use ALNs, but you can form a finite
automaton that computes its next state and outputs based on a state
and input symbol.  An ALN could recognize when, for example, state 12
was the next appropriate state, conflicts being resolved by a priority
scheme; and similarly for outputs.  Now give the thing an infinite
tape, and another ALN for recognizing the direction to move it in, and
you have a learning Turing machine.  I doubt it would learn what you
wanted it to, but hey, what's research for?

I didn't understand the "If not..." part of your question.  All ALNs
can do is map from a finite set to a finite set, and this they can do
without restriction.  (They can do it.  Whether they can learn to do
it is a different question.)  Your question seems to boil down to what
you can do with that capability.  It's almost:"What can you do with
unlimited RAM?"  It takes more than RAM to make a computer, but
exactly what is flexible.
--
***************************************************
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