Many problems in machine learning involve collecting high-dimensional multivariate observations or sequences of observations, and then fitting a compact model which explains these observations.  The predominant approaches used in machine learning for fitting models are based either on the principle of maximum likelihood or Bayesian inference.  However, the
algorithms used with these approaches (e.g., Expectation-Maximization) are known to suffer from slow convergence or poor quality local optima.

In the past several years, the machine learning and computer science communities have revisited a classical statistical approach called the method of moments, and designed computationally efficient algorithms based on this approach to tackle challenging learning problems.  Many of these algorithms have been based on spectral decompositions of moment matrices or other algebraic structures, and hence have also gone by the name of "spectral learning" algorithms.  In contrast to algorithms like E-M, these algorithms come with polynomial computational and sample complexity guarantees.  Moreover, they have been applied to learn the structure and parameters of many models including predictive state representations, finite state transducers, hidden Markov models, latent trees, latent junction trees, probabilistic context free grammars, and mixture/admixture models.  They have also been applied to a wide range of application domains including system identification, video modeling, speech modeling, robotics, and natural language processing.

The focus of this workshop will be on spectral learning algorithms and the application of the method of moments to machine learning problems.  We would like the workshop to be as inclusive as possible and encourage paper submissions and participation from a wide range of research related to this focus.

Important dates:
  • Submission deadline: April 10, 2014
  • Notification of acceptance: April 21, 2014
  • Workshop: June 25, 2014