A New Architecture for Multilayer Perceptrons as Function Approxi mators
Changlin Cai · Journal of Sichuan University · 2006
Through the construction of a specific kind of three-layer feedforward neural networkcalled“Selective Activation Network(SAN)”,the authorsintroduce a newarchitecture for mul-tilayer perceptrons(MLPs) as function approxi mators of continuous mappings between finite di-mensional Euclidean spaces whose domain is compact.Furthermore,they showthat neural net-works with such architecture are capable of approxi mating arbitrarily well any such mapping,yetwithout possibly causing the destruction or loss of the previouslylearnt knowledge whenthey aretrained with newpatterns.