On 9/5/06, Kingma, D.P. <[EMAIL PROTECTED]> wrote:
> YKY, I agree with your views, predicate logic is much more
> straightforward to work with, and I absolutely respect all the work
> and thoughts put into it.
>
> A problem within the AI domain is that Vision has not been solved yet.
> The existing and functioning algorithms are mainly specialised into
> sub domains like face recognition etc. These are very nice but are not
> general enough to use in an arbitrary environment. There does not
> exist an elegant solution to general vision theory but I suspect that
> the paper of Hawkins et al does point in the right direction.
> My wish is to have a general perceptual unsupervised hierarchical
> learning algorithm with enough generality to be used for any domain
> with N-dimensional space plus time (if applyable). That is generally
> what Hawkins et al are talking about, but I think they're just
> pioneering a field that offers a wide collection of possible
> implementations.
> Hawkin's concept is similar to Hebbian learning but it makes use of
> the hierarchical structure of some perceptual (e.g. visual) data. This
> makes it quite fast. My idea is to use hierarchical lower visual
> perceptions, which are connected to a Hebbian learning module higher
> in hierarchy to connect/combine different sensory inputs to create
> multisensory perceptual experiences. These could then be brought to
> even higher level percepts such as "Peter hits Ann" or even "Peter is
> about to hit Ann". But if even moderate-level abstractions are
> achieved, that would be a major breakthrough and useful for any AI or
> AGI.
> YKY, I agree with your views, predicate logic is much more
> straightforward to work with, and I absolutely respect all the work
> and thoughts put into it.
>
> A problem within the AI domain is that Vision has not been solved yet.
> The existing and functioning algorithms are mainly specialised into
> sub domains like face recognition etc. These are very nice but are not
> general enough to use in an arbitrary environment. There does not
> exist an elegant solution to general vision theory but I suspect that
> the paper of Hawkins et al does point in the right direction.
> My wish is to have a general perceptual unsupervised hierarchical
> learning algorithm with enough generality to be used for any domain
> with N-dimensional space plus time (if applyable). That is generally
> what Hawkins et al are talking about, but I think they're just
> pioneering a field that offers a wide collection of possible
> implementations.
> Hawkin's concept is similar to Hebbian learning but it makes use of
> the hierarchical structure of some perceptual (e.g. visual) data. This
> makes it quite fast. My idea is to use hierarchical lower visual
> perceptions, which are connected to a Hebbian learning module higher
> in hierarchy to connect/combine different sensory inputs to create
> multisensory perceptual experiences. These could then be brought to
> even higher level percepts such as "Peter hits Ann" or even "Peter is
> about to hit Ann". But if even moderate-level abstractions are
> achieved, that would be a major breakthrough and useful for any AI or
> AGI.
"Been there, failed to do that." =)
The predicate Hit(x,y) can be used to recognize many events, such as "John hits Mary", "the robot hits the dog", or "King Kong hits Tyrannosaurus"; this is the flexibility of using variables. With NNs, variables are still a very tricky problem.
Hawkin's theory is essentially a variation of MLP (multi-layer perceptron). The problem of variables does not go away. Once you try implementing NN on vision you'll see why...
YKY
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
