repetitive training

[Boris 'pi' Piwinger]

Greg Louis wrote:

>> ("sit"->cookie, "sit"->cookie, etc ;).  If the first registration does
>> not move the classification into the cutoff zone, then it registers
>> again and again either until it does classify correctly, or until an
>> arbitrary maximum is reached (default 10 recursions) in case it never
>> converges.
> 
> Great!  A procedure that is easy to reproduce "in the lab" for testing.
> I may very likely run such an experiment.

Indeed. I don't expect that to work out too well, since it
will create the need to do something to the other side, but
that will happen only after errors occur. It would be nice
to see if that happens in practice.

> (BTW I think I should point out that up to now, pi and I have _not_
> been testing the same procedure and nothing I've written was intended
> to claim otherwise.  The closest I came to doing things his way was in
> the small experiment reported at http://www.bgl.nu/training.html; after
> that I was, and still am, more interested in seeing whether repetitive
> training has any value _in_my_working_setup_.  So when pi says my
> experiments up to now don't apply to his technique I don't disagree at
> all. 

Great. It was my impression that this was the idea.

> The most important disagreement we have is over the value of
> reoptimizing parameters after training; I claim it's moderately useful
> to do so, pi seems to feel it's too far from what happens in
> production

Even worse, I feel it will break the training so far which
is made to fit with the given parameters.

So what I would like to see in you latest experiment are the
actual parameters used for each round. And possibly for each
round the error rate before and after the tuning.

>> So far, results have been great.  Unsures have been reduced
>> substantially.  It does not seem to have contributed to any false
>> positives, as I haven't received any myself, nor have had any reported. 

That surprises me.

pi