Re: Chapter 4: Vision

From: Harnad, Stevan (
Date: Sun Mar 09 1997 - 12:12:06 GMT

> From: Gill, Lynn <>
> > describing... what they term L-Grammer is not so understandable! Is it
> > saying that this is a computer program in which recognizing trees is
> > easy. And what is, "a global impression built up from the repetition of
> > microstructure?"
> According to Biederman, Geons can be used to explain local features
> (the microstructure) and global features ( the global impression ie.
> the whole object). I think it is possible that a computer program could
> be built to simulate and recognise natural objects based on repetition
> of microstructure. The reason I believe this is through a very limited
> understanding of fractals, (a pattern which can be described
> mathematically and which can be magnified at any point to reveal a
> repetition of the original pattern and so on - the global feature is
> repeated in the microstructure).

Lynn, that's absolutely correct -- and as far as I can tell, you got
this BEFORE I discussed fractals and self-similarity. Brava.

> Many natural objects, including trees, have been described as fractal.
> If I am right, and it is based on a mathematical equation or formula,
> ie. is symbolic, it should be possible to program a computer to
> recognise these sorts of irregular objects. Could it also be possible
> that we use geometric invariants as well as some sort of mathematical
> pattern recognition system in order to recognise whole (global)
> objects, like trees, from their microstucture. For example, using the
> pattern of branches, twigs and leaves to recognise the type of tree? Or
> am I completely off the track and in the trees!?

No, you're definitely on the right track. If an algorithm
(symbol-manipulation rule) can be used to recognise certain patterns
directly from the "pixel" (cone) shadow, then, apart from the shadow
itself, the recognition would all be done symbolically. But if the
analog of the sensory shadow and further internal analogs of it need
to be used, and perhaps even neural nets to detect features in it, then
it is not all being done with symbols and rules.

This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:51 GMT