John,

Thanks for your reply to my questions about your project "Tommy" in your
previous post. I'm very interested about the details but please forgive my
relative freshness to this field (CS graduate heading to an AI master :)
I'm particularly interested in the types of pattern mining you're planning
to perform, the types of patterns you think are necessary to search for
(spatial, temporal, causal, etc), and which search techniques you're going
to use given time constraints. But I guess you will release more information
as the project continues.

Mike Tintner wrote on May 16, 2007 3:13 PM:

Good example of how analogy/ blending works. But then you seem to be going
back (in my terms) to the idea that analogy works by symbolic mappings.


As far as I can see, John is not referring to any particular symbolic
implementation, but to analogy in general. You can go both ways: NARS and
Copycat both find analogies, the former symbolically but the latter is
somewhere between symbolic and connectionist.



The obvious term that I have been searching for in these discussions is
MORPHING. The way analogy actually works, I suggest, both in your example
and most of the time, is by the brain morphing one graphic/image into
another or into a composite..


The brain does not use  images for representation, except tiny patches in
the very 'lowest' regions in the visual cortex. Representation is abstract,
distributed. You could read "Seeing and Visualizing: It's Not What You
Think" by Zenon W. Pylyshyn for a comprehensive synthesis of research and
theory.

Your idea that some kind of 'morphing' happens in the brain is not new. An
interesting technique (imho) is Geoffrey Hinton's RBM (Restricted Boltzmann
Machine) which is a form of generative neural network. After training on
handwriting digits it can perform 'confabulation' which means that the
network wanders between different consistent constraint states. This is
results in interesting 'movies', which look like simplified versions of
human imagination and dream. You should see them, it's fairly consistent
with your view of 'morphing'.
Despite the fairly limited amount of artificial neurons and the networks'
generative nature, they perform very well (last time I checked, the best) on
the MNIST handwritten digit database set. It is computionally expensive
though.

-----
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=231415&user_secret=fabd7936

Reply via email to