>Vlad>>>> Could you clarify what you mean by additional information here?
Ed>>>> Of a very general class that includes what you described in your
prior quote: "In my current model there are context-sensitive links between
nodes, AND and NOT combinators. Whenever one node is active ('link origin'),
another node ('link target') will be activated, but only if additionally yet
another node is active/inactive (it's a context-sensitivity part)."
Ed Porter
-----Original Message-----
From: Vladimir Nesov [mailto:[EMAIL PROTECTED]
Sent: Friday, November 30, 2007 7:13 PM
To: [email protected]
Subject: Re: Cortical Columns [WAS Re: [agi] Funding AGI research]
On Dec 1, 2007 1:59 AM, Ed Porter <[EMAIL PROTECTED]> wrote:
> Vladimir,
>
> I thought some additional information would be required to separate
somewhat
> similar, but different cases, as your system contains.
Edward,
Could you clarify what you mean by additional information here? Do you
mean that some additional information (experience) should be required
for system to actually learn rules like those I suggested for your
examples?
>
> I assume Valiant is not an idiot so he has some half-way reasonable
> explanation for his method of hebbian learning without full connection,
but
> in the AGI world of computer hardware and software we are lucky because
its
> is easy to fake full connection, in multiple different ways.
Yes, it should make things easier. I hope to get away with one column
for a proof-of-concept-level implementation.
I read a couple of Valiant's papers and saw his talk (
http://video.google.com/videoplay?docid=-3948805418537584481 ) but
didn't catch the part about justifying feasibility of learning in
those intermediate neurons in induction model. This question arose
independently of his model, and prior to actually reading his papers I
wondered how estimates based on such model can account for induction,
so I was dismayed to find 'intermediate neurons' as a solution without
explanation. In the talk he answers the question about this issue with
a reference to his book, saying that it has something to do with
attention and working memory mechanisms [as described in it].
>
> Ed Porter
>
> -----Original Message-----
> From: Vladimir Nesov [mailto:[EMAIL PROTECTED]
>
> Sent: Friday, November 30, 2007 3:32 PM
> To: [email protected]
> Subject: Re: Cortical Columns [WAS Re: [agi] Funding AGI research]
>
> On Nov 30, 2007 9:58 PM, Ed Porter <[EMAIL PROTECTED]> wrote:
> > Vladimir Nesov>>>> There are no well-articulated theories here. I guess
> that
> > columns are induction chips: they have potential all-to-all
connectivity,
> so
> > they can learn the rule in form 'after this signal comes that signal'
for
> > any two signals in column.
> >
> > Ed>>>> how does the induction chip avoid cross talk between "if ABCDEF
> then
> > G" and "if ABCDEG then H"?
>
> That's what I included "something not that
> much more intricate than" mantra in my message for :)
>
> In my current model there are context-sensitive links between nodes,
> AND and NOT combinators. Whenever one node is active ('link origin'),
> another node ('link target') will be activated, but only if
> additionally yet another node is active/inactive (it's a
> context-sensitivity part). Something like this is inevitable if only
> to provide sensible repertoire of representable functions.
>
> So to answer your example, let for simplicity assume that there's a
> node X that is active whenever ABCDE is encountered, then XF->G can be
> represented by a context-sensitive link from X to G sensitive to F, or
> G=AND(X,F). Learning of these context-sensitive links is still kind of
> inductive, since such link is learned when triplet of nodes shows
> regularities similar to ones imposed by link.
>
> I guess choice of context-specifying parameter can be allowed to be
> more limited, so there's a place for assemblies here (but only in the
> sense of 'similar nodes'). Advantage over model without potential
> all-to-all connectivity is that problem with learning induction as in
> Valiant's model no longer applies. He used intermediate nodes to
> account for induction, allowing inductive relation between A and B to
> be learned whenever there were neurons C _between_ A and B, so that
> activation would follow A->C->B path. I can't see how C will _locally_
> know to respond to A (above other things it hears from) if B listens.
> When there's a all-to-all connectivity, this magical step is not
> needed.
>
> Additionally, and it's the reason I used delays in that
> numeric-parameters discussion last month, it's nice to have the
> ability to include delays in parameters of links. It allows temporal
> patterns to be learned by the same rule. So, with XF->G, if "XF" is a
> sequence where X is active at tact n, F at n+1 and G needs to be
> activated at tact n+2, rule is G=AND(X:2,F:1), meaning that G will be
> activated if X was activated 2 tacts ago and F was activated 1 tact
> ago. It for example allows to use link G=NOT(X:2,G:1) instead of
> G=AND(X:2,F:1) to distinguish between XF->G and XG->H (rule
> G=NOT(X:2,G:1) means 'G will be activated whenever X was active 2
> tacts ago, but G was not active 1 tact ago').
>
>
> >
> >
> > -----Original Message-----
> > From: Vladimir Nesov [mailto:[EMAIL PROTECTED]
> > Sent: Wednesday, November 28, 2007 6:19 PM
> > To: [email protected]
> >
> > Subject: Re: Cortical Columns [WAS Re: [agi] Funding AGI research]
> >
> > Edward,
> >
> > There are no well-articulated theories here. I guess that columns are
> > induction chips: they have potential all-to-all connectivity, so they
> > can learn the rule in form 'after this signal comes that signal' for
> > any two signals in column. My current bet is that something not that
> > much more intricate than this single rule is sufficient to implement
> > cognition (and on von Neumann architecture you can implement one huge
> > column of billion 'neurons' that does roughly the same). So I see
> > columns as small AGI chips that operate within their narrow sensory
> > input/output and communicate with each other to form overall behavior
> > of the brain. Compartmentalization corresponds to character of
> > concepts with which columns in particular area mostly deal.
> >
> > In this view 'neuronal assembly' corresponds to either collection of
> > neurons (or dendrite fragments?) that represents very similar shades
> > of given concept, or to different neurons in different columns that
> > correspond to the same concept and implement inter-column interaction.
> >
> > For example, perception of an object can proceed starting from
> > different senses (e.g. vision/hearing), and particular object can be
> > detected locally and separately by subsystems involved in processing
> > of each kind of these senses. In this case separate groups of columns
> > work with information about the same object, but this information is
> > only integrated on a higher stage of perception. Does it make these
> > neurons that independently detect the same object/event, part of the
> > same 'assembly'? It just might.
>
>
> --
> Vladimir Nesov mailto:[EMAIL PROTECTED]
>
> -----
> This list is sponsored by AGIRI: http://www.agiri.org/email
> To unsubscribe or change your options, please go to:
> http://v2.listbox.com/member/?&;
>
> -----
> This list is sponsored by AGIRI: http://www.agiri.org/email
> To unsubscribe or change your options, please go to:
> http://v2.listbox.com/member/?&;
--
Vladimir Nesov mailto:[EMAIL PROTECTED]
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=8660244&id_secret=71098355-688809
