Misgana and I had some further conversations on these ECAN experiments
(though these experiments have not yet been run, he's been doing some
refactoring)...
[and yes, we replaced the exterminator examples with peace, love and flowers...]
It seems we may be able to use this as a test case for concept
formation as well...
What I'm thinking is:
1)
Save, into an auxiliary Atomspace, the contents of the
AttentionalFocus every N seconds (one could make this auxiliary
Atomspace more compact by only saving events of an Atom exiting or
entering the AF, but this is a detail...)
2)
Do some pattern mining, using the pattern miner on this auxiliary
Atomspace, to figure out which combinations of Atoms tend to have been
in the AF at the same time
3)
As a scientific research point: it will be interesting to see the
extent to which combinations of Atoms that tend to be in the AF at the
same time are cross-connected with HebbianLinks
4)
We can also do some node and link building ("concept formation") based
on the results of 2 ...
4a) Build IntensionalInheritanceLinks between nodes A and B based on A
and B having HebbianLinks in common (e.g. if A and B both have strong
HebbianLinks to Y, then we can build an IntensionalInheritanceLInk
between A and B with a weight based on this common HebbianLink)
4b) Build a ConceptNode with MemberLinks to the nodes in the node-set
S observed to have common intervals of residence in the AF
5)
If we then do forward chaining inference (with premises selected based
on STI), and boost STI based on surprisingness of conclusions... we
can then get a nice feedback between ECAN, concept formation and
inference ...
...
This stuff has been documented in theoretical documents (eg.
Engineering General Intelligence) before but we have not previously
tried ECAN on a sufficiently large/complex Atomspace to support doing
these sorts of experiments...
Should be fun! though there are obviously a lot of steps to go through..
Any eager volunteers out there who want to help experiment with
concept formation along these lines, you can contact Misgana...
-- Ben
On Wed, Jan 4, 2017 at 4:17 PM, Ben Goertzel <[email protected]> wrote:
> Misgana etc.,
>
> Summarizing our discussion in the office today...
>
> 1)
> Load ConceptNet and WordNet into the Atomspace (this should take many
> GB but there are instances on AWS with loads of GB of RAM)
>
> 2)
> Experiment A)
> -- feed the system 10 articles on insects to read
> -- feed the system 5 articles on poisons to read [but not on
> insecticide -- other kinds of poisons]
> -- see if insecticide-related Atoms pop up in the Attentional Focus
> (they should)
>
> 3)
> Experiment B1)
> -- feed the system 10 articles on insects to read
> -- feed the system 5 articles on poisons to read [but not on
> insecticide -- other kinds of poisons]
> -- feed the system one article on insects
>
> Experiment B2)
> -- feed the system 10 articles on insects to read
> -- feed the system 5 articles on poisons to read [but not on
> insecticide -- other kinds of poisons]
> -- feed the system one article on cars
>
>
> Here what we want to observe is whether in B1, the switch of attention
> from poisons back to insects, is faster than in B2, the switch of
> attention from poisons to cars
>
> 4)
> Now, take this same Atomspace with ConceptNet and WordNet in it, and
> load in Simple English Wikipedia. The goal is not to have the system
> remember SEW, but rather to have it build HebbianLinks based on the
> SEW articles it is reading. We can have the Forgetting agent run, so
> that the Atoms read from prior SEW articles will be forgotten to make
> room for the Atoms from newly read SEW articles.... (i.e. the new
> sentences from SEW articles will have high STI but low LTI, whereas
> the Atoms from WordNet and ConceptNet will have high LTI and thus be
> unlikely to get forgotten...)
>
> Then, re-run experiments A and B on this Atomspace with all the
> HebbianLinks in it
>
> An interesting parameter to play with here, is the amount of STI
> spreading that goes along HebbianLinks versus other links
>
> This gives a chance to play with the role of weak links in stabilizing
> networks, as discussed e.g. in the excellent book
>
> https://www.amazon.com/Weak-Links-Universal-Stability-Collection/dp/3540311513
>
> A hypothesis is that the presence of the weak HebbianLinks in the
> Atomspace will cause the behavior on experiments A and B to be better
> (i.e. more insecticide stuff in the AF in experiment A; more rapid
> switch back to insects in experiment B) ...
>
> ....
>
> These experiments should help us tune ECAN to work sensibly on large,
> moderately messy Atomspaces ... and from here we should be able to
> move on to using ECAN to help provide guidance to PLN for common-sense
> inferences...
>
> -- Ben
>
>
> --
> Ben Goertzel, PhD
> http://goertzel.org
>
> “I tell my students, when you go to these meetings, see what direction
> everyone is headed, so you can go in the opposite direction. Don’t
> polish the brass on the bandwagon.” – V. S. Ramachandran
--
Ben Goertzel, PhD
http://goertzel.org
“I tell my students, when you go to these meetings, see what direction
everyone is headed, so you can go in the opposite direction. Don’t
polish the brass on the bandwagon.” – V. S. Ramachandran
--
You received this message because you are subscribed to the Google Groups
"opencog" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To post to this group, send email to [email protected].
Visit this group at https://groups.google.com/group/opencog.
To view this discussion on the web visit
https://groups.google.com/d/msgid/opencog/CACYTDBdcpnXLryO%2BKQbm8P8ZE_AUNioqjc5KYtOLUrmV5OS%3D-A%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.
