Connections Per Second

[Marco.Zagha@B.GP.CS.CMU.EDU]

There are two problem with reporting just connections per second, even
if the "standard" method is used.

Connections from the input units require 4 floating point operations
per pattern, while connections in higher layers require 6 floating
operations per pattern (because of the recursive error computation).
Comparing just the CPS rating between a NetTalk benchmark and an
encoder is not quite fair, since the typical network for NetTalk has
many more connections from the input layer.

Another problem is that there are different ways of updating weights:
every pattern, every epoch, or every N patterns.

Unless someone can come up with a good way of capturing these effects
in a single number, I would suggest always giving the full details on
the network used and the frequency of weight updates.  It would also
be nice to report a performance model which estimates CPS in terms
of frequency of weight updates, number of weights, number of weights,
in the input layer, number of patterns, etc.

== Marco Zagha (marcoz at cs.cmu.edu)