On 11/10/06, Ben Goertzel <[EMAIL PROTECTED]> wrote:
> The "monolithic vs. modular" distinction, as you pose it, is
> insufficiently refined...
>
> The Novamente design is modular, in two senses:
>
> 1) there is a high-level architecture consisting of a network of
> functionally specialized lobes -- a lobe for language processing, a
> lobe for visual perception, a lobe for general cognition etc.
>
> 2) each lobe contains a set of agents, each one carrying out a
> particular cognitive process on the knowledge store (e.g. reasoning,
> declarative-to-procedural knowledge conversion, pattern mining, etc.)
>
> But, the actual cognition carried out by each agent is done using a
> particular customization of the same generalized learning algorithm.
> The "monolithic vs. modular" distinction, as you pose it, is
> insufficiently refined...
>
> The Novamente design is modular, in two senses:
>
> 1) there is a high-level architecture consisting of a network of
> functionally specialized lobes -- a lobe for language processing, a
> lobe for visual perception, a lobe for general cognition etc.
>
> 2) each lobe contains a set of agents, each one carrying out a
> particular cognitive process on the knowledge store (e.g. reasoning,
> declarative-to-procedural knowledge conversion, pattern mining, etc.)
>
> But, the actual cognition carried out by each agent is done using a
> particular customization of the same generalized learning algorithm.
This architecture seems overly complex, but it depends on the details. The word "agent" suggests something autonomous and person-like, maybe you mean many subprocesses? Cause I think it is strange to have agents within an intelligent agent.
> Actually, I believe it is simpler to have a single generalized
> learning approach underlying the various cognitive agents.
>
> >What is the advantage of a single learning
> > mechanism?
>
> The primary advantage is that it allows the different agents to
> interact and interoperate with each other much more sensitively than
> if they all used qualitatively different internal methods. This is
> critical because the most essential aspects of intelligence are those
> that emerge from the appropriate interaction of multiple cognitive
> agents.
I think the issue here is "emergence", which is also central to the neural vs logic debate. In my approach I have so far neglected emergence, and the logic is hard-wired. I think it is possible to build a fairly intelligent AGI using hard-wired logic. The question is whether the lack of emergence / flexibility would be a severe shortcoming. I don't have the answers yet, but I think this is where we should think about.
[ By the way, I didn't mean to insinuate that anyone is stupid, it was unintended. Also I didn't mean that NN is a trivial issue. Sorry about the confusion. ]
YKY
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/?list_id=303
