No subject

[Chris Cramer]

The following technical report is available. If you
would like to have copies do let me know.


           Pruning Hidden Neurons in the Kak Algorithm

                          Chris Cramer



                            ABSTRACT

               The Kak algorithm is an important new
          approach to training a feed-forward network. 
          Kak has shown that it is possible to compute
          weights for a single corner of an input space
          by inspection.  In this paper, the author will
          show that a facet classification algorithm,
          capable of mapping any input space using fewer
          hidden neurons, is also possible by combining
          a trial and error method with direct
          computation.  This facet algorithm allows for
          several input/output sequences to be covered
          by a single weight vector, thus pruning the
          necessary number of hidden neurons.  This is
          achieved by summing the weights, given by the
          Kak algorithm, for the various corner of the
          input space which are mapped to one.  Once the
          weights have been computed, the threshold
          weight may be determined.  This algorithm
          allows for the network to be trained during
          operation, after the initial training.  The
          author will demonstrate the superiority of the
          facet classification algorithm over the
          perceptron and backpropagation algorithms in
          computing a weight vector.



               Technical Report  ECE, 92-09, LSU.
                       September 22, 1992