Mixture models are a staple in machine learning and applied statistics for treating data taken from multiple sub-populations. For many classes of mixture models, parameter estimation is computationally and/or information-theoretically hard in general. However, much progress has been made over the past decade or so to overcome these hardness barriers by focusing on sub-classes that rule out the intractable cases.

One very powerful and general sub-class is the multi-view setting, where one can take advantage of several non-redundant sources of information to help distinguish different sub-populations. In this talk, I'll describe a general technique that is applicable even in semi-parametric settings, where one may not have a parametric model for individual mixture components. This technique also yields a number of new unsupervised learning results for well-studied problems, as well as very practical and scalable learning algorithms.