Learning in Interpolation Networks for Irregular Sampling: Some Convergence Properties

Abstract

Recently, Ahumada and Yellott (1) and Maloney (5,6) have presented schemes for training networks designed to reconstruct irregularly sampled retinal images. In these schemes adjustable weighting networks provide compensation for the irregularities in the retinal array and the geometrical distortions in intermediate pathways. This paper presents some ideas relating to the convergence of the training algorithms.