pith. sign in

arxiv: 1906.10048 · v1 · pith:GSKM2WRCnew · submitted 2019-06-24 · 💻 cs.CV

SurReal: Fr\'echet Mean and Distance Transform for Complex-Valued Deep Learning

Rudrasis Chakraborty , Jiayun Wang , Stella X. Yu This is my paper

Pith reviewed 2026-05-25 17:29 UTC · model grok-4.3

classification 💻 cs.CV
keywords complex-valued deep learningFréchet meanRiemannian manifoldequivariant convolutioninvariant layersSAR image classificationRF modulation classificationpolar form
0
0 comments X

The pith

A neural net for complex data uses weighted Fréchet means on a Riemannian manifold to create equivariant convolutions and invariant fully connected layers.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a deep learning architecture for complex-valued data that often carries scaling and rotation ambiguities. It models inputs as fields in the complex plane and uses the polar form to define the group action of planar rotation combined with non-zero scaling. From this it derives a convolution operator based on the weighted Fréchet mean on the associated Riemannian manifold, which respects the group action by equivariance, and a fully connected layer based on distance to that mean, which is invariant. On the MSTAR dataset the resulting model reaches 98 percent target classification accuracy with fewer than 45,000 parameters, improving on a 94 percent baseline while using only 8 percent of its size. On the RadioML dataset the same approach matches baseline modulation classification accuracy at 10 percent of the baseline model size.

Core claim

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.

What carries the argument

weighted Fréchet mean (wFM) on a Riemannian manifold for the convolution operator (equivariant to rotation and scaling) and distance to the wFM for the fully connected layer (invariant to those actions)

If this is right

  • On MSTAR SAR images the model reaches 98% accuracy with under 45,000 parameters, improving from the baseline 94% at 8% of its size.
  • On RadioML RF signals the model matches baseline accuracy at 10% of baseline model size.
  • Convolution layers are equivariant to non-zero scaling and planar rotation.
  • Fully connected layers are invariant to non-zero scaling and planar rotation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same manifold construction could be tested on other complex-valued domains such as MRI phase data to check for similar efficiency gains.
  • An ablation that removes the Riemannian metric while keeping the polar representation would isolate whether the manifold geometry itself drives the reported gains.
  • The approach suggests that explicitly encoding the multiplicative group structure of complex numbers can reduce the parameter count needed for rotation- and scale-robust signal classification.

Load-bearing premise

The polar-form group action of planar rotation and non-zero scaling supplies the symmetries needed to build equivariant and invariant layers that improve classification on the tested datasets.

What would settle it

Retraining both the proposed model and the real-valued baseline on the MSTAR dataset and observing no accuracy gain or parameter reduction would falsify the performance claim.

Figures

Figures reproduced from arXiv: 1906.10048 by Jiayun Wang, Rudrasis Chakraborty, Stella X. Yu.

Figure 1
Figure 1. Figure 1: Equivariance of weighted Fréchet mean filtering [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Sample architecture of our complex-valued CNN [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-valued CNN baseline model (top) and our [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: MSTAR has 10 imbalanced target classes, a sam [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: MSTAR classification accuracies by real-valued [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sample MSTAR filter responses of our model after the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: RadioML data samples. We plot one sample per [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative filter outputs after the first, seco [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Real-valued CNN model and our complex-valued [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
read the original 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 Fr\'echet 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes SurReal, a complex-valued neural network architecture that models data in the complex plane under the group action of planar rotations and non-zero scalings derived from the polar form. It introduces a weighted Fréchet mean (wFM) convolution operator claimed to be equivariant to this group action and a distance-to-wFM fully connected layer claimed to be invariant, both defined on a Riemannian manifold. The central empirical claim is that these operators yield improved or comparable classification accuracy on the MSTAR SAR dataset (94% to 98% at 8% model size, <45k parameters) and RadioML RF modulation dataset (comparable accuracy at 10% model size) relative to real-valued two-channel baselines.

Significance. If the performance gains can be rigorously attributed to the proposed equivariant/invariant operators rather than other factors, the work would offer a geometrically principled approach to handling scaling and phase ambiguities in complex data, with potential for parameter-efficient models in SAR imaging and communications. The Riemannian construction via Fréchet means is a distinctive technical element that could influence future complex-valued architectures if the symmetry assumptions align with the data.

major comments (2)
  1. [Abstract] Abstract and experimental sections: the reported accuracy gains (94%→98% on imbalanced MSTAR at <45k parameters; comparable RadioML accuracy at 10% baseline size) are presented without accompanying ablation studies, baseline architecture details, training protocols, or statistical significance tests that would isolate whether the wFM convolution and distance-to-wFM FC layers (and specifically their equivariance/invariance under the chosen C* group action) are responsible for the improvements versus the two-channel real baseline.
  2. [Experiments] The manuscript does not provide analysis or experiments testing whether the planar rotation × non-zero scaling group action matches the dominant ambiguities in MSTAR or RadioML data (as opposed to global phase, translation, or amplitude normalization), which is required to attribute the performance delta to the Riemannian construction rather than generic complex-valued processing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to strengthen the experimental validation and discussion of the group action.

read point-by-point responses

  1. Referee: [Abstract] Abstract and experimental sections: the reported accuracy gains (94%→98% on imbalanced MSTAR at <45k parameters; comparable RadioML accuracy at 10% baseline size) are presented without accompanying ablation studies, baseline architecture details, training protocols, or statistical significance tests that would isolate whether the wFM convolution and distance-to-wFM FC layers (and specifically their equivariance/invariance under the chosen C* group action) are responsible for the improvements versus the two-channel real baseline.

    Authors: We agree that the manuscript would benefit from more detailed experimental support to isolate the contribution of the proposed operators. In the revision we will add ablation studies that replace the wFM convolution and distance-to-wFM layers with standard complex-valued counterparts while keeping all other factors fixed, provide complete specifications of the baseline architectures and training protocols, and report statistical significance of the accuracy differences. These additions will allow readers to better attribute performance to the equivariant and invariant properties. revision: yes

  2. Referee: [Experiments] The manuscript does not provide analysis or experiments testing whether the planar rotation × non-zero scaling group action matches the dominant ambiguities in MSTAR or RadioML data (as opposed to global phase, translation, or amplitude normalization), which is required to attribute the performance delta to the Riemannian construction rather than generic complex-valued processing.

    Authors: The group action follows directly from the polar decomposition of complex numbers, which encodes the scaling and rotation ambiguities that arise in SAR imaging and RF signal acquisition. We will expand the manuscript with a dedicated discussion subsection that justifies this choice using the physical characteristics of each dataset. While we do not currently include side-by-side experiments with alternative group actions, the consistent gains across two distinct domains support the relevance of the chosen symmetries. revision: partial

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper selects the polar-form group action (planar rotation × non-zero scaling) as the symmetry for complex-valued data and constructs wFM convolution (equivariant) and distance-to-wFM FC (invariant) layers by design from that choice. Reported gains (94%→98% on MSTAR at <45k params; comparable RadioML accuracy at 10% size) are presented as empirical outcomes on external datasets, not as quantities derived or forced by the layer equations themselves. No self-citations, fitted parameters renamed as predictions, self-definitional reductions, or ansatz smuggling appear in the abstract or described chain. The architecture is self-contained against the chosen symmetry; performance claims rest on dataset evaluation rather than reducing to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the standard identification of complex numbers with a Riemannian manifold equipped with the polar-form group action; no free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption Complex numbers in polar form admit a Riemannian manifold structure whose isometries include planar rotation and non-zero scaling.
    This symmetry is invoked to justify the equivariance and invariance properties of the new layers.

pith-pipeline@v0.9.0 · 5761 in / 1300 out tokens · 36075 ms · 2026-05-25T17:29:58.650885+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 8 internal anchors

  1. [1]

    An introduction to differen- tiable manifolds and Riemannian geometry , volume

    William M Boothby. An introduction to differen- tiable manifolds and Riemannian geometry , volume

    work page

  2. [2]

    Academic press, 1986

    work page 1986

  3. [3]

    Adaptive learning for complex-valued data

    Kerstin Bunte, Frank-Michael Schleif, and Michael Biehl. Adaptive learning for complex-valued data. In ESANN. Citeseer, 2012

    work page 2012

  4. [4]

    A CNN for homogneous Riemannian manifolds with applications to Neuroimaging

    Rudrasis Chakraborty, Monami Banerjee, and Baba C V emuri. H-cnns: Convolutional neural networks for riemannian homogeneous spaces. arXiv preprint arXiv:1805.05487, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    ManifoldNet: A Deep Network Framework for Manifold-valued Data

    Rudrasis Chakraborty, Jose Bouza, Jonathan Manton, and Baba C V emuri. Manifoldnet: A deep network framework for manifold-valued data. arXiv preprint arXiv:1809.06211, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    Group equivariant con- volutional networks

    Taco Cohen and Max Welling. Group equivariant con- volutional networks. In International conference on machine learning, pages 2990–2999, 2016

    work page 2016

  7. [7]

    Spherical CNNs

    Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical CNNs. arXiv preprint arXiv:1801.10130, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  8. [8]

    Willett, and Joni Dambre

    Sander Dieleman, Kyle W . Willett, and Joni Dambre. Rotation-invariant convolutional neural networks for galaxy morphology prediction. Monthly Notices of the Royal Astronomical Society, 2015

    work page 2015

  9. [9]

    Abstract algebra, volume 3

    David Steven Dummit and Richard M Foote. Abstract algebra, volume 3. Wiley Hoboken, 2004

    work page 2004

  10. [10]

    Polar Transformer Networks

    Carlos Esteves, Christine Allen-Blanchette, Xiaowei Zhou, and Kostas Daniilidis. Polar Transformer Net- works. arXiv preprint arXiv:1709.01889 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    Freeman and Edward H Adelson

    William T. Freeman and Edward H Adelson. The de- sign and use of steerable filters. IEEE Transactions on Pattern Analysis & Machine Intelligence, (9):891–906, 1991

    work page 1991

  12. [12]

    Differential geometry and symmet- ric spaces, volume 12

    Sigurdur Helgason. Differential geometry and symmet- ric spaces, volume 12. Academic press, 1962

    work page 1962

  13. [13]

    Mstar extended operating conditions: A tutorial

    Eric R Keydel, Shung Wu Lee, and John T Moore. Mstar extended operating conditions: A tutorial. In Algorithms for Synthetic Aperture Radar Imagery III , volume 2757, pages 228–243. International Society for Optics and Photonics, 1996

    work page 1996

  14. [14]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  15. [15]

    On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups

    Risi Kondor and Shubhendu Trivedi. On the general- ization of equivariance and convolution in neural net- works to the action of compact groups. arXiv preprint arXiv:1802.03690, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  16. [16]

    ImageNet Classification with Deep Convolutional Neural Networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hin- ton. ImageNet Classification with Deep Convolutional Neural Networks. Advances In Neural Information Processing Systems, 2012

    work page 2012

  17. [17]

    Affinity cnn: Learning pixel-centric pairwise relations for figure/ground embedding

    Michael Maire, Takuya Narihira, and Stella X Y u. Affinity cnn: Learning pixel-centric pairwise relations for figure/ground embedding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 174–182, 2016

    work page 2016

  18. [18]

    Understanding Deep Convolu- tional Networks

    Stéphane Mallat. Understanding Deep Convolu- tional Networks. Philosophical Transactions A , 374:20150203, 2016

    work page 2016

  19. [19]

    Les éléments aléatoires de nature quelconque dans un espace distancié

    Maurice Fréchet. Les éléments aléatoires de nature quelconque dans un espace distancié. Annales de l’I. H. P .,, 10(4):215–310, 1948

    work page 1948

  20. [20]

    Convolutional Radio Modulation Recognition Networks

    Timothy J O’Shea, Johnathan Corgan, and T. Charles Clancy. Convolutional radio modulation recognition networks. arXiv preprint arXiv:1602.04105 , 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  21. [21]

    Radio machine learning dataset generation with gnu radio

    Timothy J O’Shea and Nathan West. Radio machine learning dataset generation with gnu radio. Proceed- ings of the 6th GNU Radio Conference , 2016

    work page 2016

  22. [22]

    Deep Complex Networks

    Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy Serdyuk, Sandeep Subramanian, João Felipe Santos, Soroush Mehri, Negar Rostamzadeh, Y oshua Bengio, and Christopher J Pal. Deep complex networks. arXiv preprint arXiv:1705.09792, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Joint embedding and classification for sar target recognition

    Jiayun Wang, Patrick Virtue, and Stella X Y u. Joint embedding and classification for sar target recognition. arXiv preprint arXiv:1712.01511 , 2017

    work page arXiv 2017

  24. [24]

    Harmonic net- works: Deep translation and rotation equivariance

    Daniel E Worrall, Stephan J Garbin, Daniyar Tur- mukhambetov, and Gabriel J Brostow. Harmonic net- works: Deep translation and rotation equivariance. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), volume 2, 2017

    work page 2017

  25. [25]

    Angular embedding: A robust quadratic criterion

    Stella Y u. Angular embedding: A robust quadratic criterion. IEEE transactions on pattern analysis and machine intelligence, 34(1):158–173, 2012

    work page 2012