Adaptive averaging in higher order neural networks for invariant pattern recognition

For the task of position-, scale-, and rotation-invariant pattern recognition higher order neural networks have shown good separation results for different object classes. However, their use is limited by the large number of monomials that have to be calculated to get the invariant features. Even with the application of methods like coarse coding to reduce the number of monomials the computational requirements are still large. Here a new method is proposed where the averaging of the monomials to form the invariant features is performed adaptively with a network architecture that uses coupled nodes. Thereby the features are adapted to a specific application. This reduces significantly the number and order of the monomials necessary for the invariant recognition of patterns. Moreover, the classification part of the network becomes superfluous since the output of the network is already one-dimensional and binary. The properties of the new method in comparison to higher order neural networks are shown for position and 90-degree-rotation invariance and for invariance with respect to different object backgrounds. The application to rotations of smaller angles is outlined