Jason Cawley
Wolfram Science Group
Phoenix, AZ USA
Registered: Aug 2003
Posts: 712 |
Actually, HTMs aren't really models of human intel, though the design is motivated by certain features ascribed to human intel by the designers (whether true of human intel or not). They are really a cross between Bayesian networks for machine learning and neural nets. As such, they are indeed general transforms or more exactly, classifiers, that one can imagine throwing at any kind of data, to see what the method manages to make of that data.
The obvious thing to try to do with them in NKS would be to make a CA rule recognizer, in the sense of trying to train an HTM to discriminate among different CA rules (the top level outputs that would count as success being, "this is a rule n pattern", "that is a rule m pattern"). Visual pattern recognition is what they are most developed at doing, really - they are largely an attempt to extend the success of NNs at static pattern recognition, to recognition of objects in various views, by incorporating a time domain and different scales. Well, you would just feed them data from simple CA rules, slightly more complex ones from simple initials (rule 90, rule 225, etc), then complex ones from simple, etc.
As for class 4s specifically, one might hope an HTM could recognize particles in relatively "thin" versions of them, mostly repeating background I mean. It would be a fair test of what is claimed for them - though not earth shattering in what it might show us about CAs.
Has it been worked on? I think not. Some have worked on CA methods for pattern recognition, and have fed CA images and textures to standard pattern recognition AI routines - that is about as close as it gets, in terms of stuff already done (as far as I know, obviously).
Report this post to a moderator | IP: Logged
|