Abstract
We develop a novel deep learning architecture for naturally complex-valued data, which is often subject to complex scaling ambiguity. We treat each sample as a field in the space of complex numbers. With the polar form of a complex-valued number, the general group that acts in this space is the product of planar rotation and non-zero scaling. This perspective allows us to develop not only a novel convolution operator using weighted Fréchet mean (wFM) on a Riemannian manifold, but also a novel fully connected layer operator using the distance to the wFM, with natural equivariant properties to non-zero scaling and planar rotation for the former and invariance properties for the latter. Compared to the baseline approach of learning real-valued neural network models on the two-channel real-valued representation of complex-valued data, our method achieves surreal performance on two publicly available complex-valued datasets: MSTAR on SAR images and RadioML on radio frequency signals. On MSTAR, at 8% of the baseline model size and with fewer than 45,000 parameters, our model improves the target classification accuracy from 94% to 98% on this highly imbalanced dataset. On RadioML, our model achieves comparable RF modulation classification accuracy at 10% of the baseline model size.
Introduction
In modern deep learning, data is typically treated as points in Euclidean space. However, many critical signal types—such as Synthetic Aperture Radar (SAR), Magnetic Resonance (MR) images, and Radio Frequency (RF) signals—are naturally complex-valued. Standard practice involves splitting these complex numbers ($a + ib$) into two real-valued channels, effectively treating the data as living in a flat, 2D Euclidean plane.
This conventional approach ignores the intrinsic geometry of complex numbers and fails to account for complex scaling ambiguity. In many applications, a signal may undergo rotation in the complex plane or be scaled by a non-zero factor without altering its fundamental information. Standard Convolutional Neural Networks (CNNs) are not inherently aware of these transformations and often require extensive data augmentation to learn such invariances. The SurReal framework addresses this by rethinking neural network layers through the lens of Riemannian manifolds, baking equivariance and invariance directly into the mathematical foundations of the architecture.
Methodology: Deep Learning on Riemannian Manifolds
The core innovation of SurReal is mapping complex numbers to a product manifold: $\mathbb{R}^+ \times SO(2)$, where the positive reals represent magnitude and the group of 2D rotations represents phase. This geometric perspective enables the creation of layers that respect the polar geometry of the data.
Weighted Fréchet Mean (wFM) Convolution
In standard CNNs, convolution is a linear weighted sum. On a manifold, a simple sum can pull points off the manifold surface. SurReal replaces this with the weighted Fréchet mean (wFM). The wFM is defined as the point on the manifold that minimizes the weighted distance to the input points. This operation is mathematically guaranteed to be equivariant to both scaling and rotation, meaning the network's internal representations transform predictably when the input is rotated or scaled.
Tangent ReLU (tReLU) and Invariant Layers
Standard activation functions like ReLU do not have a natural definition on a manifold. SurReal introduces the Tangent ReLU (tReLU), which projects manifold data onto a flat tangent space using a logarithmic map, applies a standard ReLU, and then projects the result back onto the manifold via an exponential map.
To achieve final classification, the network must be invariant to scaling. The authors replace traditional fully connected (FC) layers with a distance transform. This layer measures the distance from an input to the learned Fréchet mean of each class. Since the distance between points on the manifold remains constant under global rotation or scaling, the resulting classification is naturally invariant to these transformations.
Experimental Evaluation
The SurReal architecture was evaluated on two primary complex-valued benchmarks: the MSTAR dataset (SAR images of vehicles) and the RadioML dataset (RF modulation classification).
MSTAR and RadioML Results
On the MSTAR dataset, which is characterized by high noise and class imbalance, SurReal improved classification accuracy from 94% to 98%. Remarkably, it achieved this while using only 8% of the parameters required by the baseline real-valued CNN (fewer than 45,000 parameters compared to over 530,000). Similar results were observed on the RadioML dataset, where SurReal matched baseline performance with only 10% of the model size.
The Power of Phase
A key insight from the study was the importance of phase information. When magnitudes were normalized, the SurReal model maintained a 97% accuracy using phase alone. In contrast, standard CNNs failed this test, dropping to approximately 46% accuracy. This demonstrates that the manifold-aware approach captures structural information that Euclidean models overlook.
Conclusion
SurReal demonstrates that respecting the intrinsic geometry of complex-valued data leads to models that are both more accurate and significantly more efficient. By replacing Euclidean operations with manifold-valued equivalents like the weighted Fréchet mean, the architecture achieves state-of-the-art results with a fraction of the parameters. While the iterative nature of the wFM solver introduces some computational overhead, the dramatic reduction in parameter count and improved robustness to scaling and rotation make it a powerful framework for complex-valued signal processing. Future work may extend these principles to other non-Euclidean data types in fields such as robotics and medical imaging.
Summary updated