One of the interesting but disappointing things about groups like this is that
there is pretty serious lack of insightful criticism and comment. Part of
that, of course, is just based on simple lack of understanding, but most of it
seems based on total disbelief that the premises of the approaches that have
been talked about are being taken seriously. I really cannot understand how
anyone can take AIXI seriously and I cannot understand how someone can take the
neuroscience approach seriously. It would be as if I were to say that I am
working on an AGI program that is based on a polynomial time solution to
Boolean SAT (which I will one day solve.) I mean, can you take something like
that seriously without first seeing the polynomial time solution (or an amazing
AGI program which was based on it)? There is nothing wrong with looking at
these ideas and seeing how far you can go with them, but I think there is a lot
wrong with believing that one of these methods are currently viable. Using
Bayesian methods, compression methods and information theory, trying to create
algorithms that emulate observed and theoretical neural processes, and trying
to come up with creative logical methods all make perfect sense to me. I just
think that the fringe science that is based on taking some sound methods to an
absurd theoretical extreme looks like a pretty terrible place to start.
I went through some severe cold feet with my own would-be AGI project. The
massive inefficiency of a practical method of representing the possibilities is
really unacceptable. I just could not get myself going with such inept
representational methods that we seem to be stuck with. However, after going
through a couple of days of talking myself into acceptance I finally came up
with an elementary system that would create some efficiencies without making
the look-ups too deep. I am thinking of the problem of initial recognition but
it is the same through out all the stages of analysis and response. I am just
going to come up with a simple intuitive method to reduce the grossest
inefficiencies that a simpler implementation of my ideas would create.
I thought about using numerous Neural Networks or a Bayesian Networks for the
initial recognition lookup problem, but then I started wondering about a more
definitive network that would use a few of the characteristics of the neural
network (it would become more extensive to represent more inputs or to be more
precise in determining the outputs for a particular kind of input) but it would
also have the characteristics of the network that I have been thinking of (it
might utilize a greater variety of specific markers to represent syntactic
characteristics of input and output, it could use reason based reasoning and so
on. The only problem with this plan is that I haven't figured it out yet. The
whole project is supposed to help me discover how I might create such a thing,
so I really don't know if I could start the project with it. I might try a
simplified version of it. I might use weighted evaluations but they could, for
example, be used to represent approximations to multiple output values
(representing indexes to data). It would not be a conventional Neural or
Bayesian Network. For example, it could be designed to represent
approximations to values in some non-conventional ways. This is an interesting
idea. I have to think about it for a few days.
Jim Bromer
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
