We consider identifiability and subspace selection for the ICA model with additive Gaussian noise. We discuss a canonical decomposition that allows us to decompose the system into a signal and a noise subspace and show that an unbiased estimate of these can be obtained using a standard ICA algorithm. This can also be used to estimate the relevant subspace dimensions and may often be preferable to PCA dimension reduction. Finally we discuss the identifiability issues for the subsequent 'square' noisy ICA model after projection. © Springer-Verlag 2004.
CITATION STYLE
Davies, M. (2004). Identifiability, subspace selection and noisy ICA. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 3195, 152–159. https://doi.org/10.1007/978-3-540-30110-3_20
Mendeley helps you to discover research relevant for your work.