METHODOLOGY OF NETWORKS WITH SYMMETRICAL TRANSFORMATION
FUNCTIONS FOR NEURONS
A. V. Merkusheva, G. F. Malychina
Saint-Petersburg
The basic methodology of networks
with symmetrical transformation functions for neurons (STFN) is presented.
Such neural networks (NN) find application in the problems of approximation,
pattern recognition, systems (objects) identification, controller designing,
noise level lowering for signals in information-measurement systems. The
characteristic property of NN structure is the localization of hidden-layer
elements in the multidimensional vector space (whose dimension is identical
to the input information dimension) and the presence of STFN depending on
the (metric) norm of the difference between the hidden-layer element localization
vectors and the input signal-vector. The paper presents the applied theory
elements for pattern recognition using a network with STFN; NN learning criteria
on the basis of a functional regularized with the aid of A.N. Tikhonov’s
method; a general form of approximation and interpolation functions (obtained
on the basis of those criteria) using Green’s scheme for the inverse problem
generated by transformation with a differential operator; a method for regularization
parameter selection.