多層競合ネットによる座標変換に不変なパターン認識

抄録

A multi-layered competitive neural network is presented for learning to achieve pattern recognition in a manner invariant to linear and/or nonlinear coordinate transformations such as projection, shift, rotation, magnification and so on. The transformed input patterns stored in the network are multiplied by the Jacobian of the transformation, an aspect shown to be essential for the transformation invariance. The network also has excellent generalization ability as has been verified by computer simulation.