No subject

>do not depend on the net's actually being implemented in parallel,
>rather than just being serially simulated? Is it only speed and
>capacity parameters, or something more?

I have a simple answer.  In practice, I've resisted using parallel machines to run 
backprop simulations because in deciding how best to parallelize the problem for a 
given machine, you tend to make choices you're less likely to make with a very fast serial 
machine.  So, for example, if you parallelize over patterns (one machine node processes the
entire net for a given pattern) you sacrifice the capability to update the weights after every
pattern.  These choices tend to be different for different parallel machines. 

This experience makes me suspect that, in modeling the brain, the specifics of the
parallel implementation (e.g., restriction to local connectivity) are likely to determine
the nature of information representation and learning algorithms, as well as of what types of
information processing the organism is capable.

Gale Martin
MCC