Global and local weight sharing

We provide a unified view and comparison of different methods in the invariance problem. We show that the local symmetry network can reduce the number of weights, that the hardware implementation can be localized and that the number of multiplications is far less than in the global symmetry network. The geometric interpretations of global and local networks are also given. Finally, we examine the performance differences of the two network configurations and discuss the improvement on generalization for the local symmetry method