Liam Paninski
<br><br>
Some recent results on the neural coding problem
<br><br>

A central problem in neuroscience is to estimate the conditional
probabilities P(response | stimulus) from neurophysiological data.
Stated in this generality, this problem is unsolvable; we will never
be able to record enough responses to specify these distributions for
all possible stimuli.  (This theme is not restricted to the neural
context, of course.)  Thus, we have to try to solve some kind of
easier version of this problem.  As time permits, I'd like to talk
about some recent and ongoing work on three such approaches:
<br><br>
1) Instead of trying to estimate the full distribution, just estimate a
few of the most important functionals of the distribution.  In
particular, I will discuss the problem of estimating certain
information-theoretic quantities from data.
<br><br>
2) Fit some low-dimensional model to the distributions, instead of
trying to estimate them by "brute force" (nonparametrically).  Models
of ``cascade'' form --- with a simple linear prefiltering stage
followed by a stochastic, nonlinear spiking step --- have turned out
to be especially interesting and useful.  I'll focus on applications
to the analysis of population coding of complex hand movements in
primary motor cortex.
<br><br>
3) Design the experiment to be as efficient as possible, so we don't
waste our time recording the responses to "uninformative" stimuli.
This leads naturally to a simple Bayesian adaptive experimental-design
procedure (for which, in turn, we can prove some interesting
convergence theorems).