Purely Logical Neural Principal Component and Independent Component Learning

Aapo Hyvärinen

Abstract

New algorithms for neural Principal Component and Independent Component Analysis (PCA and ICA) are introduced. Special emphasis is laid on the locality of learning. This enables a simpler hardware implementation, and may provide a more plausible model of biological neurons. To achieve this, the algorithms feature a new kind of feedback which is multiplicative and anti-Hebbian. The convergence of the ICA algorithm is proven analytically in the general case; the convergence of the PCA algorithm is proven for Gaussian data.
BibTeX
@inproceedings{hyvarinen1996purely,
  author = {Aapo Hyvärinen},
  title = {{Purely Logical Neural Principal Component and Independent Component Learning}},
  booktitle = {ICANN},
  year = {1996},
  doi = {10.1007/3-540-61510-5\_27},
}

Loading related content...

Figure