Liam Paninski, 7/22/03
<br>
<br>
Classification via integrate-and-fire cells
<br>
<br>

The title is a little misleading.  I'll talk a little about Gaussian
process classifiers (GPCs), a nonparametric approach to pattern
recognition that has become popular in the machine learning literature
recently.  My interest in this field was piqued by Zoubin Ghahramani,
who pointed out that some of the work I had done with Jonathan Pillow
on the estimation of neural encoding models based on <a
href="pillow2_4.html">integrate-and-fire</a> cells might be applicable
to GPCs.  <br> <br> 

A good source for further information on GPCs (especially on the
generalization bounds I mentioned) is <a
href="http://www.dai.ed.ac.uk/homes/seeger/papers/pacbgp.html">here</a>. 
<br><br>

I'm expecting the discussion to be brief, since it will be our first
after a little midsummer vacation.  If there is time/interest left
over, I'll say a few words about an interesting statistical trick for
showing that no estimator can possibly perform better than the one you
have developed.  (Sadly, this trick only applies to certain problems.)
This idea turns out to have an interesting and unexpected application
to the entropy estimation problem.