Steve Dekorte wrote on Thu, 17 Jun 2010 16:06:45 -0700
> > [Self compiler technologies vs hardware]
> 
> It seems to me that those compiler tricks make assumptions about usage 
> patterns
> (they assume you're really not doing too much dynamic stuff) and aren't 
> generalizable.
> e.g. an associative memory can be used for generic databases (there is, after 
> all, a
> reason why routers use them) raising the possibility of a greater unification 
> of data
> and behavior.

Though Self and Io are often lumped together as "prototype based
languages", there are significant differences in the programming styles.
Self is very much a Smalltalk and it is normal for programmers to start
out programming in a Smalltalk style before evolving to something more
optimized for Self (doing more refactoring and exploration instead of up
front design). The compiler technology certainly was created to support
that more static style. The reverse was also true - programmers quickly
learnd to avoid what the compiler didn't do well (dynamic inheritance,
for example). 

> Also, as you mentioned, there are a number of levels on which associative 
> memory like
> functionality is currently used. Could a fully associative memory system 
> unify all these
> usage patterns?

The problem of the von Neuman architecture is that you have millions or
billions of transistors on one side (superscalar CPU) and more billions
on the other side (RAM) but just a tiny bridge connecting them. This is
not very efficient in terms of power, though it has been less than a
decade since we started worrying about this below the supercomputer
level. With the distributed memory parallel machines you have at least
lots of little bridges connecting smaller groups of transistors.

A single, huge associative memory would have the same problems as the
von Neuman architecture even though you have moved some of the smarts to
the memory side. Our brains are proof that you can have highly parallel
associative memories and my RNA proposal is an attempt to see what the
extreme case of that might be. Some hybrid scheme which replaced a
single 1 GB associative memory with 1000 such memories of 1 MB each
would be worth investigating. How would they collaborate?

While I mentioned Ian's paper, the "context" (or "worlds") papers from
VPRI are equally relevant for this discussion. There the normal
<objectID, selector> tuple used as the key for accessing the associative
memory is replaced by <objectID, selector, context> where you might be
interested in partial matches to reduce the duplication of information.

-- Jecel


_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc

Reply via email to