Learning of robust principal component subspace
Juha Karhunen, J. Joutsensalo · 2005
We study various neural algorithms for learning so-called robust principal component subspace. Standard principal components and the corresponding subspace are defined in terms of quadratic optimization criteria, leading to algorithms having linear learning term. The robust algorithms are derived by optimizing a similar criterion that grows less that quadratically. This introduces a nonlinearity into the gradient algorithms, but makes the results more robust against strong noise and outliers.