Dimension reduction in unsupervised learning via Non-Gaussian Component Analysis

The report will be focused on the dimension reduction topic. In this talk, we will discuss methods based on Non-gaussian component analysis (NGCA). It can be formulated as a problem of identifying a low-dimensional non-Gaussian component of the whole distribution in an iterative and structure adaptive way. In more details, we will mainly consider NGCA procedure of identification of the non-Gaussian subspace using Principle Component Analysis (PCA) method, sparse NGCA (SNGCA) which replaces the PCA-based procedure with an algorithm based on convex projection and approach for direct estimation of the projector on the target subspace based on semidefinite programming. Recovering the structure when its effective dimension is unknown will be also discussed.