> On Thu, 2002-10-31 at 01:12, Tony Lofthouse wrote:
> >
> > In a biological brain closely associated neurons are 'generally' close
> > to each other physically. Whilst, artificial nodes and links are
> > generally not.
> >
> > For me this seems to lose a whole layer of information. The patterns
> > related to the interconnectedness are there but the information related
> > to the spatial layout of the nodes is lost.
> I'll give you a third way that breaks the dichotomy.  In our model we
> use a link/node structure which is not derived from NN models.  One
> property that was noticed early on is that if you built a natural 3D
> model of the network all the links are short, connecting only to other
> relatively local nodes.  This wasn't a designed feature, just something
> that was noticed later.  The network is very dense, but always local.
> It would be hard (impossible?) to implement our model without exhibiting
> this type of locality.

Yes, this is an interesting point.

I noticed something similar in early versions of Webmind, though we haven't
done such tests on Novamente.

What I did there was

a) assign a sparse n-dim vector to each node in the network

For instance if node 12 has only 2 links, to nodes 3 and 1, with weights

w_12,3 = .5
w_12,1 = .9

then node 12 gets assigned the vector

v_12 = (.9,0,.5,0,...)

b) map these vectors to 2D using a Kohonen network

What you get looks a lot nicer than what you'd get using a random graph...
not shockingly...

The (statistical) connection topology in a graph-based AI system interacting
with the real world, is generally different from the statistical connection
topology of a random graph.  Among other properties, you  do tend to get a
greater amount of "local linkage" in a 2D or 3D projection.

You get this even without explicit use of clustering algorithms or other
neighborhood-formation-inducing tools.

In Novamente our attention allocation methods tend to explicitly
create/reinforce this kind of connection topology anyway....

But, a connection topology that tends to have a lot of local connections is
not the same as one that really maps a 2D or 3D space, with the precision
desired for processing spatial input data...

-- Ben G

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/

Reply via email to