Abstract
We develop a novel deep learning architecture for naturally complex valued data, which are 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 number, the general group that acts on 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 Frechet 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 rotations for the former and invariance properties for the latter. We demonstrate our method on widely used SAR dataset MSTAR and RadioML dataset. On MSTAR data, without any preprocessing, our network can achieve 98% classification accuracy on this highly imbalanced dataset using only 44,000 parameters, as opposed to 94% accuracy with more than 500,000 parameters with a baseline real-valued network on the two-channel real representation of the complex valued data. On RadioML data, we got comparable classification accuracy with the baseline while only using 10% of the parameters as the baseline model.
Introduction
Many critical real-world signals, including Synthetic Aperture Radar (SAR), Magnetic Resonance Imaging (MRI), and Radio Frequency (RF) signals, are inherently complex-valued. Traditionally, deep learning practitioners treat these signals as two-channel real-valued data (Real and Imaginary). However, this approach ignores the intrinsic geometry of the complex plane.
A primary challenge in complex-valued data is complex scaling ambiguity, where a signal's identity remains unchanged despite rotations or scaling in the complex plane. Standard Convolutional Neural Networks (CNNs) lack inherent invariance to these transformations, necessitating large parameter counts and extensive data augmentation to learn these symmetries. This paper introduces ComplexNet, an architecture designed to bake these geometric priors directly into the network structure.
Methods: Manifold Learning and wFM
ComplexNet operates by treating the space of complex numbers ($\mathbb{C}$) as a Riemannian manifold, specifically identifying it with the product of positive real numbers (magnitude) and the group of planar rotations ($SO(2)$).
Weighted Fréchet Mean (wFM) Convolution
In standard Euclidean space, a convolution is a weighted sum. On a non-Euclidean manifold, addition is not well-defined. Instead, the authors replace the standard convolution with the weighted Fréchet mean (wFM)—the point that minimizes the weighted sum of squared distances to all input points in a local neighborhood. This operator is mathematically equivariant to both scaling and rotation.
Architecture Components
- Tangent ReLU (tReLU): To introduce non-linearity without leaving the manifold, the authors utilize a "tangent ReLU" that operates in the log-domain of the manifold.
- Invariant Layer: The final layer calculates the distance between the output of the last convolution and learned cluster centers (wFMs). This ensures the final classification is invariant to global rotations or scaling of the input signal.
Experimental Results
The model was evaluated on the MSTAR (SAR imagery) and RadioML (RF signals) benchmarks.
MSTAR Performance
On the MSTAR dataset, which contains 10 classes of vehicles and is characterized by significant class imbalance, ComplexNet achieved 98.16% accuracy. Remarkably, it achieved this using only 44,826 parameters. In contrast, a baseline real-valued network reached only 94% accuracy despite utilizing over 500,000 parameters. This represents a 10x reduction in parameters with a significant gain in performance.
RadioML and Ablation Studies
For RadioML, ComplexNet matched state-of-the-art performance using only 10% of the baseline parameters (299k vs 2.8M). Ablation studies further demonstrated that the architecture's strength lies in its geometric handling of phase information. When tested on unit-norm data (phase only), ComplexNet significantly outperformed standard CNN baselines, proving that the benefit is derived from the architecture's design rather than just the input data format.
Conclusion
ComplexNet demonstrates that incorporating geometric priors into deep learning architectures can lead to massive improvements in parameter efficiency and accuracy for complex-valued data. By treating complex numbers as points on a manifold and utilizing the weighted Fréchet mean, the model naturally handles the symmetries inherent in SAR and RF signals. Future work will explore the application of this framework to other domains such as medical MRI and interferometry.
Summary updated