2pm, Tuesday, 18 Oct 2005:

<CENTER><B>
Dimensionality reduction in neural models: An
information-theoretic generalization of spike-
triggered average and covariance analysis
</B>

<P>
Jonathan Pillow
<BR>
NYU

</CENTER>

 We describe an information-theoretic framework for characterizing
  spiking neural responses using a Linear-Nonlinear-Poisson (LNP)
  cascade model.  The linear stage is recovered by sequentially
  searching for a set of linear filters that maximize the information
  between stimuli and spiking responses.  This optimization problem is
  made tractable by assuming that both the raw and spike-triggered
  stimulus ensembles are drawn from Gaussian distributions.  The
  resulting solution has several surprising properties: (1) it is
  equivalent to maximizing the likelihood of the spike train given the
  stimuli; (2) like many existing techniques, it is based on
  initial measurements of spike-triggered average and covariance, but
  it unifies these in a common framework that provides a set of
  filters sorted according to their informativeness about the neural
  response; (3) unlike unrestricted information-theoretic estimators,
  it is both computationally efficient and robust, and 
  the data requirements for recovery of 
  a linear stage of high dimensionality are relatively modest; 
  (4) it implicitly assumes that the nonlinear stage maps the
  linear filter responses to a Poisson spike rate using a
  ratio-of-Gaussians function, which is capable of mimicking neural
  response in the early visual system.  We demonstrate the
  effectiveness and advantages of the method by applying it to
  simulated responses of a Hodgkin-Huxley neuron, and the recorded
  extracelluar responses of macaque retinal ganglion cells and
  V1 cells.