TR available - fault-tolerant recurrent neural networks

[Christian Omlin]

The following Technical Report is available via the NEC Research Institute 
archives:


__________________________________________________________________________________


            Fault-Tolerant Implementation of Finite-State Automata
                      in Recurrent Neural Networks

     RENSSELAER POLYTECHNIC INSTITUTE DEPT. OF COMPUTER SCIENCE TR CS 95-3

                        C.W. Omlin[1,2], C.L. Giles[1,3]

      [1]NEC Research Institute, 4 Independence Way, Princeton, NJ 08540
      [2]CS Department, Rensselaer Polytechnic Institute, Troy, NY 12180
      [3]UMIACS, University of Maryland, College Park, MD 20742} 

                     {omlinc,giles}@research.nj.nec.com
               
 
                                  ABSTRACT

Recently, we have proven that the dynamics of any deterministic finite-state 
automaton (DFA) with n states and m input symbols can be implemented in a sparse
second-order recurrent neural network (SORNN) with n+1 state neurons, O(mn) 
second-order weights and sigmoidal discriminant functions. We investigate how that
constructive algorithm can be extended to fault-tolerant neural DFA implementations
where faults in an analog implementation of neurons or weights do not affect the
desired network performance. We show that tolerance to weight perturbation can be
achieved easily; tolerance to weight and/or neuron stuck-at-zero faults, however,  
requires duplication of the network resources. This result has an impact on the 
construction of neural DFAs with a dense internal representation of DFA states.

__________________________________________________________________________________

http://www.neci.nj.nec.com/homepages/omlin/omlin.html

or

ftp://ftp.nj.nec.com/pub/omlinc/fault_tolerance.ps.Z

__________________________________________________________________________________