HyperReal: Complex-Valued Layer Functions For Complex-Valued Scaling Invariance

Complex-valued measurements in MRI and SAR imaging often have complex-valued scaling ambiguity, calling for models that are invariant to complex-valued scaling of pixels. Deep Complex Networks (DCN) extends real-valued algebra to complex-valued algebra in neural networks, but it does not address the issue of complex-valued scaling. SurReal complex-valued networks adopt a manifold view of complex numbers and derive a distance metric that is invariant to complex scaling. With distance features, it achieves complex-scaling invariance. However, rich complex-valued information is lost in this representation, and additionally, SurReal is also prevented from using complex-valued non-linearity, limiting its expressive power. We simplify the manifold formulation of SurReal and propose a new layer function that achieves complex-scaling invariance within the complex domain. We can then build hierarchical complex-valued features with complex-scaling invariance. Our so-called HyperReal model results in a much leaner model with better generalization. Benchmarked on MSTAR, HyperReal beats DCN (and matches SurReal) with only 3%(40%) of their respective parameter counts.