Abstract
We study complex-valued scaling as a type of symmetry natural and unique tocomplex-valued measurements and representations. Deep Complex Networks (DCN)extends real-valued algebra to the complex domain without addressingcomplex-valued scaling. SurReal takes a restrictive manifold view of complexnumbers, adopting a distance metric to achieve complex-scaling invariance whilelosing rich complex-valued information. We analyze complex-valued scaling as aco-domain transformation and design novel equivariant and invariant neuralnetwork layer functions for this special transformation. We also propose novelcomplex-valued representations of RGB images, where complex-valued scalingindicates hue shift or correlated changes across color channels. Benchmarked onMSTAR, CIFAR10, CIFAR100, and SVHN, our co-domain symmetric (CDS) classifiersdeliver higher accuracy, better generalization, robustness to co-domaintransformations, and lower model bias and variance than DCN and SurReal withfar fewer parameters.