SurReal: Complex-Valued Learning as Principled Transformations on a Scaling and Rotation Manifold

Rudrasis Chakraborty, Yifei Xing, Stella X. Yu

TL;DR

This paper introduces SurReal, a novel deep learning framework for complex-valued data that treats complex numbers as points on a geometric manifold. By modeling complex scaling and rotation as inherent group actions, SurReal achieves principled invariance and equivariance, leading to significantly smaller and more performant models compared to existing methods.

What Problem Does This Paper Address?

Deep learning has seen tremendous success, but most architectures are designed for data in vector spaces. Many real-world signals, like Synthetic Aperture Radar (SAR) and Magnetic Resonance (MR) images, are inherently complex-valued. While complex numbers offer richer information than real numbers, current deep learning approaches often treat their real and imaginary parts as independent channels, ignoring their underlying geometric relationships. This is problematic because complex-valued data can be subject to arbitrary scaling and rotation (e.g., due to sensor calibration or acquisition parameters), which affects both components simultaneously. Existing methods to handle this often resort to data augmentation, which is inefficient and doesn't guarantee true invariance. This paper aims to develop a deep learning framework that intrinsically handles the geometry of complex numbers, making it robust to complex-valued scaling and rotation.

Methods

The core of SurReal lies in its geometric interpretation of complex numbers. Instead of a 2D Euclidean space, the authors model the non-zero complex plane as a product manifold of positive scaling ($R^+$) and planar rotations ($SO(2)$). This allows them to define new layer functions:

  • Weighted Fréchet Mean (wFM) Convolution: This is a novel convolution operator that computes the weighted average of complex numbers on the defined manifold. It's designed to be equivariant to complex-valued scaling and rotation, meaning if the input data is scaled/rotated, the output is also scaled/rotated in the same way.
  • Tangent ReLU (tReLU) and G-transport: These are new non-linear activation functions that operate within the manifold's tangent space or by transporting feature values via group actions, providing non-linearity in a geometrically principled way.
  • Distance Transform Layer: This layer computes the manifold distance between feature maps and their wFM, acting as a fully-connected layer that is invariant to complex-valued scaling and rotation. This layer effectively converts complex-valued representations into real-valued ones, suitable for standard downstream classifiers.
  • Complex-Valued Residual Connections: The authors extend their framework to incorporate residual connections, which are crucial for training deep networks, by using channel-wise concatenation instead of simple addition (which is not well-defined on manifolds).

Datasets

The proposed SurReal model was evaluated on two complex-valued datasets:

  • MSTAR Dataset: A dataset of complex-valued X-band SAR images for target recognition, featuring 11 classes of vehicles and background clutter.
  • RadioML Dataset: A dataset of synthetically generated radio signals with various modulation modes, designed for classification tasks.

Key Insights and Ideas

  • Complex Numbers as Manifolds: The most significant insight is reframing complex numbers not as simple 2D vectors but as points on a $R^+ \times SO(2)$ manifold. This geometric perspective unlocks principled ways to handle their inherent properties.
  • Equivariance vs. Invariance: The paper clearly distinguishes between equivariance (transformations in input lead to corresponding transformations in output, handled by wFM convolution) and invariance (transformations in input do not affect the output, handled by the distance transform layer). This principled approach is key to their model's robustness.
  • Intrinsic Handling of Scaling: Instead of extrinsic data augmentation, SurReal builds invariance to complex-valued scaling directly into its layer functions. This is a more elegant and efficient solution.
  • Leaner Models, Better Performance: The geometric approach leads to significantly smaller models (fewer parameters) that still outperform both real-valued and existing complex-valued deep learning baselines, especially in low-data regimes.

Key Results

  • MSTAR Classification: SurReal models (both basic and residual) consistently outperformed real-valued (ResNet50) and complex-valued (DCN) baselines. Notably, SurReal achieved significantly higher accuracy (90%) than DCN (60%) and ResNet50 (45%) when trained with only 5% of the data.
  • Model Size: SurReal models are remarkably lean. The basic SurReal CNN has 67K parameters, and the residual version has 109K. This is a fraction of the parameters of ResNet50 (23.5M) and DCN (154K).
  • RadioML Classification: SurReal also showed strong performance on the RadioML dataset, outperforming baselines at higher SNRs and achieving comparable results at lower SNRs with a fraction of the model size.

Limitations and Future Work

  • Computational Cost of wFM: The paper notes that the weighted Fréchet mean (wFM) computation does not have a closed-form solution and requires an iterative solution, which can be computationally intensive.
  • Generalization to Other Manifolds: While the paper focuses on the complex plane manifold, the underlying principles of geometric deep learning could potentially be extended to other non-Euclidean spaces.
  • Further Parameter Reduction: The authors mention that tensor ring decomposition can be used to further reduce the parameter count of their convolutional layers, suggesting avenues for even leaner models.
  • Exploration of Activation Functions: While tReLU and G-transport are proposed, further exploration of other manifold-based activation functions could be beneficial.

Summary updated

Abstract

Complex-valued data are ubiquitous in signal and image processing applications, and complex-valued representations in deep learning have appealing theoretical properties. While these aspects have long been recognized, complex-valued deep learning continues to lag far behind its real-valued counterpart. We propose a principled geometric approach to complex-valued deep learning. Complex-valued data could often be subject to arbitrary complex-valued scaling; as a result, real and imaginary components could covary. Instead of treating complex values as two independent channels of real values, we recognize their underlying geometry: we model the space of complex numbers as a product manifold of nonzero scaling and planar rotations. Arbitrary complex-valued scaling naturally becomes a group of transitive actions on this manifold. We propose to extend the property instead of the form of real-valued functions to the complex domain. We define convolution as the weighted Fréchet mean on the manifold that is equivariant to the group of scaling/rotation actions and define distance transform on the manifold that is invariant to the action group. The manifold perspective also allows us to define nonlinear activation functions, such as tangent ReLU and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$G$ </tex-math></inline-formula> -transport, as well as residual connections on the manifold-valued data. We dub our model SurReal, as our experiments on MSTAR and RadioML deliver high performance with only a fractional size of real- and complex-valued baseline models.
BibTeX
@article{chakraborty2022surreal,
  author = {Rudrasis Chakraborty and Yifei Xing and Stella X. Yu},
  title = {{SurReal: Complex-Valued Learning as Principled Transformations on a Scaling and Rotation Manifold}},
  journal = {IEEE Trans. Neural Networks Learn. Syst.},
  year = {2022},
  url = {https://openreview.net/forum?id=oN_-3-phB52},
}
Figure