Kernel and Range Approach to Analytic Network Learning

Abstract

A novel learning approach for a composite function that can be written in the form of a matrix system of linear equations is introduced in this paper. This learning approach, which is gradient-free, is grounded upon the observation that solving the system of linear equations by manipulating the kernel and the range projection spaces using the Moore–Penrose inversion boils down to an approximation in the least squares error sense. In view of the heavy dependence on computation of the pseudoinverse, a simplification method is proposed. The learning approach is applied to learn a multilayer feedforward neural network with full weight connections. The numerical experiments on learning both synthetic and benchmark data sets not only validate the feasibility but also depict the performance of the proposed formulation.

Keywords

• Least squares error
• linear algebra
• multilayer neural networks