Papers available on neural PCA and ICA

[Simone G.O. Fiori]

Dear Connectionists,
 the following two papers are now available:

     `Mechanical' Neural Learning for Blind Source Separation
     ========================================================
   
      by Simone Fiori - Dept. of Industrial Engineering
      University of Perugia, Perugia (Italy)

      Journal: Electronics Letters, Vol. 35, No. 22, Oct. 1999

                           Extended abstract

   In this Letter we suggest a new learning theory which ensures 
   that the weight-matrix of each layer of a neural network keeps 
   orthonormal during the whole learning phase. The learning theory 
   is based upon the study of the dynamics of an abstract rigid 
   mechanical system subject to a field of external forces deriving 
   from a potential energy function (PEF). We suggest that by properly 
   selecting the PEF it is possible to force the system to perform 
   different motions, hence the network to perform different tasks. 
   The proposed learning theory is then applied in order to solve 
   blind source separation problems.

                           Related papers
   
   The mentioned work is a part of a wider study about neural learning
   and weight flow on Stiefel-Grassman manifold. Interested readers
   can find more details on the following papers (and references therein):

   [1] S. Fiori et al., Orthonormal Strongly-Constrained Neural 
       Learning, Proc. IJCNN'98, pp. 1332 - 1337, 1998
   [2] --, A Second-Order Differential System for Orthonormal
       Optimization, Proc. of International Symposium on 
       Circuits and Systems, Vol. V, pp. 531 - 534, 1999
   [3] --, `Mechanical' Neural Learning and InfoMax Orthonormal 
       Independent Component Analysis, Proc. IJCNN'99, in press
   [4] --, Neural Learning and Weight Flow on Stiefel Manifold,
       Proc. X Italian Workshop on Neural Nets, pp. 325 -- 333,
       1998 (in English)
                                ~~~~~~~~~

    An Experimental Comparison of Three PCA Neural Networks
    =======================================================

    by Simone Fiori - Dept. of Industrial Engineering
    University of Perugia, Perugia (Italy)

    Journal: Neural Processing Letters, accepted for publication.

                                Abstract

   We present a numerical and structural comparison of three neural 
   PCA techniques: The GHA by Sanger, the APEX by Kung and Diamantaras, 
   and the $\psi$--APEX first proposed by the present author. Through 
   computer simulations we illustrate the performances of the algorithms 
   in terms of convergence speed and minimal attainable error; then an 
   evaluation of the computational efforts for the different algorithms 
   is presented and discussed. A close examination of the obtained 
   results shows that the members of the new class improve the numerical
   performances of the considered existing algorithms, and are also 
   easier to implement.

                                ~~~~~~~~~

 Comments and suggestions, especially about the first topic, would be
particularly welcome. Both comments and requests of reprints should 
be addressed to:

  Dr. Simone Fiori
  Dept. of Industrial Engineering, Univ. of Perugia
  Via Ischia, 131
  I-60026 Numana (An), Italy
  E-mail: simone at eealab.unian.it
  Fax: +39.0744.470188

 Best regards,
 S. Fiori