[pdp-discuss] LEABRA++ ev_nm algorithm

[Pradeep Ramanathan]

Dear List,

I am encountering a problem and couldn't find mention of it in the
archives.

I have two sets of input patterns which originally derive from picture
stimuli but have undergone transformation.  Let's refer to these sets of
events as A1 - A100, and B1 - B100.  I have only one set of output
patterns, let's call them C1 - C100.  The network gets trained to map both
A1 and B1 to the output pattern C1, both A2 and B2 to C2, and so on down
the line.

During testing, I have noticed that whenever the network makes an error in
identifying an input pattern, for a large number of these errors, it
reports A1 in the ev_nm field of ClosestEventStat.  By "large number" I
mean sometimes as many as 60% of the erroneous ev_nm entries are listed as
A1.  (By the way, most of the other errors are understandable errors, in
that they can be confused with the correct output event.)

The occurence of such a large number of this kind of error made me suspect
that under certain circumstances, the network just "looks up" the first
event in the list of input events and reports that in the ev_nm field,
hence A1 since it is alphanumerically first in the list of input events.  I
wondered if when several events are within the same, equally low, distance
value the network just goes to the list of events in the environment and
picks the first one, which would be A1.

I tried downloading the source code files for ClosestEventStat, and perused
the algorithm for reporting ev_nm.  It doesn't look fishy at all, and I
don't see anything that appears like it might cause this problem. 
(Although, admittedly, I had some trouble following the coding language.) 
I also tried using different functions for calculating the distance,
thinking that maybe using SSE was not a good idea.  However, I have tried
all 7 distance functions, and they all demonstrate the same problem. 
(Incidentally, there is nothing unique about the picture that A1 represents
that makes it highly confusable with a large number of other pictures.)

Any ideas?

Thanks,
Pradeep



Pradeep Ramanathan, M.S., M.A., CCC-SLP
Speech-Language Pathologist/Doctoral Candidate
University of Minnesota