In this tutorial by Dr. Liam Paninski, the Expectation-Maximization (EM) algorithm is discussed and illustrated in a variety of neural examples.

Key topics addressed:

  • Example: Mixture models and spike sorting
  • The method of bound optimization via auxiliary functions provides a useful alternative optimization technique
  • The EM algorithm for maximizing the likelihood given hidden data may be derived as a bound optimization algorithm
  • EM may easily be adapted to optimize the log-posterior instead of the log-likelihood
  • Example: Deriving the EM algorithm for the mixture model (spike sorting) case
  • Example: Spike sorting given stimulus observations
  • Example: Generalized linear point-process models with spike-timing jitter
  • Example: Fitting hierarchical generalized linear models for spike trains
  • Example: Latent-variable models of overdispersion and common-input correlations in spike counts
  • Example: Iterative proportional fitting
  • The E-step may be used to compute the gradients of the marginal likelihood
  • The convergence rate of the EM algorithm depends on the “ratio of missing information”.
You may access the tutorial here or visit the author’s home page here.

Post to Twitter Post to Yahoo Buzz Post to Delicious Post to Digg Post to Facebook Post to Google Buzz Post to LinkedIn Post to Slashdot Post to StumbleUpon Post to Technorati