Controlled redundancy in interpolation-based neural nets

摘要

In this paper, we deal with the problem of associative memory synthesis via multivariate interpolation. We present an abstract yet simple formalism to address the possibility of detecting and eliminating redundant input data from the set of exemplars. The remaining pairs are then stored in a way so as to introduce controlled redundancy by replication of the corresponding neurons. The redundancy is detected via orthogonalization carried out in a Reproducing Kernel Hilbert Space setting.