Alternative Derivations of CCA Networks

Abstract

In the previous chapter, we discussed two new methods of performing Canonical Correlation Analysis (CCA) with artificial neural networks. In this chapter, we re-derive learning rules from a probabilistic perspective which then enables us, by use of a specific prior on the weights, to simplify the algorithm. We then derive CCA-type rules from Becker’s models (see Appendix D), though with a very different methodology from that used in [7]. We finally derive a robust version of the above rules from probability theory and compare the convergence of all the various rules on artificial data sets.