From: Anders Sandberg (asa@nada.kth.se)
Date: Sat Apr 14 2001 - 10:56:12 MDT
On Thu, Apr 12, 2001 at 11:18:45PM -0400, Eliezer S. Yudkowsky wrote:
> I'm already familiar with Hopfield networks, the analogy to spin glasses
> in solid-state physics, and so on. I think that they're fine for a
> monolayer problem but lack the holistic structure needed to, e.g.,
> visualize a triangular lightbulb. See CaTAI section 2.3.4, "Concept
> application and combination".
>
> Also, AFAIK, current Hopfield models use a single state which iteratively
> relaxes, rather than a set of superposed states; the actual physics for
> most phenomena resembles a probability fluid forming puddles in local
> minima of a potential energy surface.
Well, a single Hopfield network is rather boring. Making the neurons
more realistic (continous activations rather than binary states,
different kinds of connections, even compartment models) spices up
things a bit, but deep down autoassociative networks of this kind are
rather similar. Not uninteresting, but in themselves as you say a
passive system that relaxes towards certain attractor states.
The interesting things happen when you start connecting these
autoassociators with each other and other neural networks into more
complex structures. As I might have mentioned, I work on memory
consolidation models where two or more networks learn from each other,
and another student has made a reinforcement learning agent; in related
work on working memory we have come up with various schemes for networks
representing combinations of features. This is of course where our work
starts getting into this thread's representation issues. Making a
network for "imagining" a triangular lightbulb doesn't see that
impossible, although the binding issues are complex of course.
I can't say I have much to contribute to this philosophy infested area
of representation, other than a general feeling that due to our old
philosophical legacy we might make the issue thornier than it really
needs to be.
-- ----------------------------------------------------------------------- Anders Sandberg Towards Ascension! asa@nada.kth.se http://www.nada.kth.se/~asa/ GCS/M/S/O d++ -p+ c++++ !l u+ e++ m++ s+/+ n--- h+/* f+ g+ w++ t+ r+ !y
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:06:58 MST