On 11/10/06, Ben Goertzel <[EMAIL PROTECTED]> wrote:
> > 2. Ben raised the issue of learning. I think we should divide learning
> > into 3 parts:
> >
> > (1) linguistic eg grammar
> > (2) semantic / concepts
> > (3) generic / factual.
>
> This leaves out a lot, for instance procedure learning and
> metalearning... and also perceptual learning ( e.g. object recognition)
> ... and a lot more...
> > 2. Ben raised the issue of learning. I think we should divide learning
> > into 3 parts:
> >
> > (1) linguistic eg grammar
> > (2) semantic / concepts
> > (3) generic / factual.
>
> This leaves out a lot, for instance procedure learning and
> metalearning... and also perceptual learning ( e.g. object recognition)
> ... and a lot more...
Yes, sorry I forgot procedural learning =). Perceptual learning is under concept learning, though it should not be called semantic. By concept learning I mean for example learning the definition of a "cube" (in vision) or the definition of the word "fighting" etc. They're the same mechanism.
> > I think the learning mechanisms in (1)-(3) are quite distinct, for example,
> > learning that "fish cat eat" is ungrammatical, and learning that cats like
> > to eat fish, are quite distinct matters.
>
> I disagree, I think that the same learning algorithms can be made to
> carry out all sorts of learning.... It's all just pattern recognition
> ;-) ... it's true that different domains of pat. rec. demand different
> types of inductive bias, but a sufficiently flexible general learning
> algorithm can be made to incorporate appropriate domain-specific
> biases ...
>
> In Novamente, the synthesis of probabilistic logical inference and
> probabilistic evolutionary learning is to be used to carry out all of
> the above kinds of learning you mention, and more....
Well, then your architecture would be monolithic and not modular. I think it's a good choice to break the AGI into n < 5 modules, with separate learning mechanisms. In the monolithic way, implementing algorithms would be unnecessarily complex. What is the advantage of a single learning mechanism?
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
