Scale Shrinking Transformation and Applications

Abstract

This paper presents a novel information processing technique called scale shrinking transformation (SST). SST comprises three steps: initialization, matrix transformation, and using the column vectors of the transformed matrix as the new input vectors. The essence of SST is that the structural correlation between original inputs can be obtained. More significantly, the transformed matrix contains elements with much smaller scale variation. When applied to existing feedforward neural networks, it can alleviate problems commonly encountered in tasks of function approximation, separating nonlinearly classes, and noise filtering. When the column vectors are used as the new input to a feedforward network that comprises hidden layers, training speed can be reduced. The input scale divergence problem that plagues higher-order neural networks can also be alleviated with SST.