No Subject

["Randall C. O'Reilly" <oreilly@grey.colorado.edu>]

If two events have the same outputs, then the samename stat will in
fact just select the first one (it just finds the first one that
matches).  The SE stat is not susceptible to this problem, and would
be a better indication of whether it's learning.  

				- Randy

rrobare@att.net writes:
> I'm working on a network that has a similar architecture to Rumelhart & Todd 
> (1990), with separate noun and verb inputs.  The goal is to produce 
> categorization at the output layer, i.e. with two (or more) different input 
> patterns to produce the same output pattern.  This should be a simple task, but 
> the network doesn't appear to learn it, and I'm hoping someone might have a 
> suggestion as to why.
> My only clue as to what is happening is in running a textlog with a samename 
> statistic.  This shows that when Event 1 and Event 2 have the same output, the 
> input event will be Event 2 but the output event will be called Event 1 and 
> counted as an error.  
> Thanks in advance for your assistance.
> 
> Rebecca Robare