Recently, linear algebra techniques have given a fundamentally
different perspective for learning and inference in latent variable
models. Exploiting the underlying spectral properties of the model
parameters has led to fast, provably consistent methods for structure
and parameter learning that stand in contrast to previous approaches,
such as Expectation Maximization, which suffer from local optima and
slow convergence. Furthermore, these techniques have given insight
into the nature of latent variable models.
In this workshop, via a mix of contributed/invited talks, posters, and
discussion, we seek to explore the theoretical and applied aspects of
spectral methods including the following major themes:
- How can spectral techniques help us develop fast and local minima free solutions to real world problems involving latent variables in natural language processing, dynamical systems, computer vision etc. where existing methods such as Expectation Maximization are unsatisfactory?
- How can these approaches lead to a deeper understanding and
interpretation of the complexity of latent variable models?