12:00, Wed 31 Jan, WeH 7220

	   Using EM to analyze search behavior in gerbils:
		  a case study with very noisy data

			   A. David Redish*


I will present a case study of how the EM algorithm had to be modified
in order to handle noisy data (commonly seen in neuroscience
experiments).  

We are interested in how rodents use landmarks to navigate.  To
explore this issue, we trained gerbils to find a sunflower seed buried
in wood chips.  The location of the seed was held constant relative to
three landmarks (small white cylinders), which were moved (as a group)
from trial to trial.  This forced the gerbils to use the landmarks as
cues to find the food.

We generated histograms of where the gerbils spent their time when
there was no seed and the configuration of landmarks was changed.
Gerbils do not go to the "goal" location and sit, but they do tend to
spend more time at "goal" locations than other parts of the arena.  We
used the EM algorithm to fit gaussians to the histograms to determine
where the gerbils spent most of their time.  

In order to handle the noisy data, we had to modify the EM algorithm
to accomodate outliers (throwing out more than 40% of the data).  I
will show how we did this and still produced statistically valid
answers (some of them quite surprising) and how we were able to use
these results to differentiate competing hypotheses of how gerbils
navigate using landmarks.

PS. I promise a very cute video if the AV equipment is available.

* Work with 
	David S. Touretzky, Steven J. Gaulin, Chris Reiber-Milberg, 
	Lisa M. Saksida, and David Banks