Provable ICA with unknown Gaussian noise, and implications for Gaussian mixtures and autoencoders (Q2345948)

Independent component analysis (ICA) is a computational method for separating a multivariate signal into additive sub-components. ICA defines a generative model for the observed multivariate data belonging to large databases. In this context, the paper presents a new algorithm with provable performance guarantees for unknown Gaussian noise. Formally, this algorithm can be seen as a nonlinear auto-encoding scheme analogous to a restricted Boltzmann machine, but with uniform mixing weights.