Neural networks for invariant pattern recognition

In this paper, we discuss a methodology for applying feedforward networks to problems of invariant pattern recognition. We present the Group Representation Network (GRN), a type of feedforward network with the property that its output is invariant under a group of transformations of its input. Since the invariance of such a network is inbuilt, it does not need to be learned. Consequently it is capable of a better generalization performance than a conventional network for solving the same symmetric problem. In addition, the GRN has fewer free parameters than connections and we can hence expect it to train faster than an ordinary network of the same connectivity.