pith. sign in

arxiv: 2406.09588 · v7 · submitted 2024-06-13 · 💻 cs.CV · cs.LG

Learning Color Equivariant Representations

Yulong Yang , Felix O'Mahony , Christine Allen-Blanchette This is my paper

Pith reviewed 2026-05-24 00:12 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords color equivariancegroup convolutional networkslifting layerhue saturation luminanceperceptual transformationsgeneralizationsample efficiency
0
0 comments X

The pith

A lifting layer lets group convolutional networks achieve color equivariance by transforming the input image directly.

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

The paper establishes that group convolutional neural networks can be made equivariant to hue, saturation, and luminance shifts. It does so by replacing filter rotation with a lifting layer applied to the input image itself. This change avoids invalid RGB values and cuts equivariance error by more than three orders of magnitude. The resulting networks generalize more reliably to unseen color conditions and require fewer training examples than standard models on both synthetic and real datasets.

Core claim

Group convolutional neural networks achieve equivariance to color variation through a lifting layer that transforms the input image directly rather than the convolutional filters. This construction extends hue equivariance to saturation and luminance shifts, eliminates invalid RGB values, and produces networks with strong generalization to out-of-distribution perceptual changes plus improved sample efficiency over conventional architectures.

What carries the argument

The lifting layer, which transforms the input image directly to support color group convolutions while preserving valid RGB values.

If this is right

  • The networks generalize strongly to out-of-distribution perceptual variations in hue, saturation, and luminance.
  • Sample efficiency improves relative to conventional convolutional architectures.
  • Performance exceeds competitive baselines on both synthetic and real-world image datasets.

Where Pith is reading between the lines

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

  • The same lifting approach could be tested on other continuous perceptual attributes such as contrast or white balance.
  • Color-equivariant layers might combine with geometric equivariance layers for joint robustness to lighting and viewpoint changes.
  • The method could be evaluated on video sequences to check whether temporal consistency of color transformations holds.

Load-bearing premise

That applying the lifting layer to the input image produces a representation that remains valid and equivariant under color group actions without introducing new artifacts.

What would settle it

Run the same set of color-transformed test images through the network with and without the lifting layer; if the measured equivariance error does not fall by at least three orders of magnitude with the lifting layer, the central claim is false.

Figures

Figures reproduced from arXiv: 2406.09588 by Christine Allen-Blanchette, Felix O'Mahony, Yulong Yang.

Figure 1
Figure 1. Figure 1: Color-equivariant network. (a) The equivariance of our hue-equivariant model is illustrated by the commutativity of the (hue) rotation and neural network mapping. A hue rotation of 90◦ in the input image space (top-left to bottom-left), results in a feature map rotation at each layer of the network (top-right to bottom-right). Corresponding feature maps are highlighted with a blue border. (b) An input imag… view at source ↗
Figure 2
Figure 2. Figure 2: Saturation-equivariant feature maps. We illustrate the equivariance of our saturation￾equivariant model. A saturation shift in the input image space (top-left to bottom-left), results in a feature map translation at each layer of the network (top-right to bottom-right). Corresponding feature maps are highlighted with a blue border. saturation group acts on an HSL image by the group action φs : SN × X → X, … view at source ↗
Figure 3
Figure 3. Figure 3: Impact of order on hue rotation invertibility. Our lifting layer (blue) operates on HSL input images where each hue rotation is invertible. The lifting layer proposed in CEConv (orange) operates on RGB filters and suffers from invalid hue rotations for all discretizations of the hue group except for N = 1 and N = 3 (i.e., symmetries of the axis-aligned RGB cube). (Left) We show the impact of invalid hue ro… view at source ↗
Figure 4
Figure 4. Figure 4: Model sample efficiency. We show the error improvement (higher is better) over the Z2CNN baseline as a function of the per￾centage of training examples used. The advan￾tage of our Hue-N models increases as the per￾centage of training examples used decreases [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hue-shift MNIST feature map visualization. We compare the feature map trajectories of MNIST digits as their hue is varied from 1◦ to 360◦ . The color of the trajectory corresponds to the class label. (a) tSNE projection of hue shifted feature map trajectories in the Z2CNN baseline model. As the hue of the input changes, the location of the digit in the feature space changes significantly. (b) tSNE projecti… view at source ↗
Figure 6
Figure 6. Figure 6: Hue-equivariant feature maps. (a) We qualitatively compare hue-shifted feature maps produced using our lifting layer and the CEConv lifting layer. Highlighted feature maps obtained using our lifting layer (Hue-4) are qualitatively indistinguishable, while there are visible discrepancies in the feature maps of CEConv-4. (b) We quantitatively compare the equivariance-error of the hue￾shifted feature maps for… view at source ↗
Figure 7
Figure 7. Figure 7: Color sorting on CIFAR-10. We show images from CIFAR-10 automobile class ordered by pairwise distance using Hue-4 and ResNet44 feature maps. The structure of the Hue-4 feature maps naturally allow for color based sorting, whereas the ResNet44 feature maps do not. We include visualizations of the entire automobile class sorted using Hue-N and CEConv-N in Appendix C.3. Cars (Krause et al., 2013), STL-10 (Coa… view at source ↗
Figure 8
Figure 8. Figure 8: Lifting layer. (a) In Hue-N, an input image (left) is lifted to the hue-saturation group (right) by shifting its hue and saturation values. (b) CEConv shifts the hue of a filter in the RGB space by rotating its values about an axis passing through the point p = (1, 1, 1). This approach results in invalid hue rotations for all discretizations of the hue group that are not symmetries of the axis-aligned RGB … view at source ↗
Figure 9
Figure 9. Figure 9: Hue shift MNIST feature map visualization. tSNE projection of hue shifted feature map trajectories in Hue-3 and CEConv-3. 0 1 2 3 4 5 6 7 8 9 (a) Hue-4 0 1 2 3 4 5 6 7 8 9 (b) CEConv-4 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Hue shift MNIST feature map visualization. tSNE projection of hue shifted feature map trajectories in Hue-4 and CEConv-4. In contrast Hue-4, there is more ambiguous scatter in the feature maps of CEConv-4. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Color sorting on CIFAR-10 using Hue-3. We show images from CIFAR-10 automobile class ordered by pairwise distance using Hue-3 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Color sorting on CIFAR-10 using CEConv-3. We show images from CIFAR-10 automobile class ordered by pairwise distance using CEConv-3 [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Color sorting on CIFAR-10 using Hue-4. We show images from CIFAR-10 automobile class ordered by pairwise distance using Hue-4 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Color sorting on CIFAR-10 using CEConv-4. We show images from CIFAR-10 automobile class ordered by pairwise distance using CEConv-4. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Hue-shift MNIST dataset. (a) Examples from the training dataset and in-distribution testing dataset A are colored with a randomly selected hue between 0◦ and 240◦ . (b) Examples from out-of-distribution testing dataset B are colored with a randomly selected hue between 240◦ and 360◦ . D.2 HUE-SHIFT 3D SHAPES We introduce the Hue-shift 3D Shapes dataset to evaluate the performance of our equivariant archi￾… view at source ↗
Figure 16
Figure 16. Figure 16: Hue-shift 3D Shapes dataset. (a) Examples from the training dataset and in-distribution testing dataset A. The color of the wall, the floor and the shape are randomly selected from the first half of the color space (colors 0-4). (b) Examples from the out-of-distribution testing dataset B. The color of the wall, the floor and the shape are randomly selected from the second half of the color space (colors 5… view at source ↗
Figure 17
Figure 17. Figure 17: Small NORB dataset. (a) Examples from the training dataset and in-distribution testing dataset A. (b) Examples from the out-of-distribution testing dataset B with lower lighting conditions. (c) Examples from the out-of-distribution testing dataset C with higher lighting conditions. D.4 TINY-IMAGENET We use the Tiny-ImageNet dataset to evaluate the performance of our equivariant architectures in the presen… view at source ↗
Figure 18
Figure 18. Figure 18: Camelyon-17 dataset saturation statistics. Example images and saturation statistics by node (hospital). We present experimental results for four discretizations of the saturation space. Our choice of discretizations, i.e., d ∈ {1/20, 1/10, 3/20, 1/2}, are determined with consideration of the saturation range of the training and validation set (see [PITH_FULL_IMAGE:figures/full_fig_p022_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Luminance equivariant feature maps. We illustrate the equivariance of our luminance￾equivariant model. A luminance shift in the input image space (top-left to bottom-left), results in a feature map translation at each layer of the network (top-right to bottom-right). Corresponding feature maps are highlighted with a blue border [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
Figure 20
Figure 20. Figure 20 [PITH_FULL_IMAGE:figures/full_fig_p027_20.png] view at source ↗
read the original abstract

In this paper, we introduce group convolutional neural networks (GCNNs) equivariant to color variation. GCNNs have been designed for a variety of geometric transformations from 2D and 3D rotation groups, to semi-groups such as scale. Despite the improved interpretability, accuracy and generalizability of these architectures, GCNNs have seen limited application in the context of perceptual quantities. Notably, the recent CEConv network uses a GCNN to achieve equivariance to hue transformations by convolving input images with a hue rotated RGB filter. However, this approach leads to invalid RGB values which break equivariance and degrade performance. We resolve these issues with a lifting layer that transforms the input image directly, thereby circumventing the issue of invalid RGB values and improving equivariance error by over three orders of magnitude. Moreover, we extend the notion of color equivariance to include equivariance to saturation and luminance shift. Our hue-, saturation-, luminance- and color-equivariant networks achieve strong generalization to out-of-distribution perceptual variations and improved sample efficiency over conventional architectures. We demonstrate the utility of our approach on synthetic and real world datasets where we consistently outperform competitive baselines.

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 / 2 minor

Summary. The paper introduces group convolutional neural networks (GCNNs) that are equivariant to color variations, specifically extending to hue, saturation, and luminance shifts. It proposes a lifting layer that transforms the input image directly (rather than convolving with rotated filters as in CEConv) to avoid invalid RGB values. The authors claim this yields an equivariance error improvement of over three orders of magnitude, leading to stronger out-of-distribution generalization and better sample efficiency than conventional architectures, with demonstrations on synthetic and real-world datasets.

Significance. If the equivariance error reduction and fair baseline comparisons hold under the stated conditions, the work would provide a concrete architectural improvement for perceptual equivariance in GCNNs, potentially aiding generalization in color-sensitive vision tasks. The explicit handling of saturation and luminance beyond hue is a useful extension.

major comments (2)
  1. [Method (lifting layer description) and Experiments (equivariance error table/figure)] The central claim of the lifting layer improving equivariance error by over three orders of magnitude (abstract and method) is load-bearing for the generalization and sample-efficiency results, yet the manuscript provides no explicit definition of the equivariance error metric (e.g., L2 norm over which group elements, normalization, test distribution), no equation for the lifting operator, and no confirmation that CEConv baselines were reimplemented with identical hyperparameters or color-space handling.
  2. [Preliminaries / Color Equivariance section] § on color group definitions: the extension to saturation and luminance equivariance is presented as a direct generalization, but the paper does not specify whether the combined color group is a direct product or semidirect product and how the lifting layer composes the individual transformations without introducing additional parameters or breaking the claimed parameter-free property.
minor comments (2)
  1. [Experiments] Figure captions for the synthetic dataset experiments should explicitly state the number of training samples used for the sample-efficiency curves and whether error bars represent standard deviation over multiple seeds.
  2. [Preliminaries] Notation for the hue/saturation/luminance transformations is introduced without a compact group-theoretic summary (e.g., explicit generators or parametrization), which would aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address the major comments below and will make revisions to improve the clarity of the manuscript.

read point-by-point responses

  1. Referee: [Method (lifting layer description) and Experiments (equivariance error table/figure)] The central claim of the lifting layer improving equivariance error by over three orders of magnitude (abstract and method) is load-bearing for the generalization and sample-efficiency results, yet the manuscript provides no explicit definition of the equivariance error metric (e.g., L2 norm over which group elements, normalization, test distribution), no equation for the lifting operator, and no confirmation that CEConv baselines were reimplemented with identical hyperparameters or color-space handling.

    Authors: We agree that these details should be made explicit to support the central claims. The equivariance error is measured as the average L2 distance between f(g·x) and g·f(x) over sampled group elements g in the color group, normalized by the input magnitude, evaluated on a held-out test distribution. We will add the mathematical definition and the equation for the lifting operator L(x) = x transformed in color space directly. For the baselines, CEConv was reimplemented using the same hyperparameters and RGB color space handling as in the original work; we will include a statement confirming the reimplementation details in the experiments section. revision: yes

  2. Referee: [Preliminaries / Color Equivariance section] § on color group definitions: the extension to saturation and luminance equivariance is presented as a direct generalization, but the paper does not specify whether the combined color group is a direct product or semidirect product and how the lifting layer composes the individual transformations without introducing additional parameters or breaking the claimed parameter-free property.

    Authors: The color group is the direct product of the hue, saturation, and luminance groups since these transformations act independently on the color channels and commute. The lifting layer composes them by applying the transformations sequentially to the input image in HSL color space before lifting to the group, which remains parameter-free as no learned weights are involved in the lifting. We will add this specification to the preliminaries section to clarify the group structure and composition. revision: yes

Circularity Check

0 steps flagged

No circularity; architectural claim rests on new component and experiments

full rationale

The paper introduces a lifting layer as a novel fix for invalid RGB values in prior hue-equivariant GCNNs (CEConv) and reports empirical gains in equivariance error plus OOD generalization. No equations, derivations, or self-citations appear in the provided text that would reduce any claimed prediction or result to its own inputs by construction. The central claims are presented as consequences of the architectural design choice and benchmark results rather than tautological redefinitions or fitted quantities renamed as predictions. This is a standard non-circular case of an empirical architecture paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that a lifting layer can be defined to act directly on input images while preserving group equivariance properties for color transformations; no free parameters or invented physical entities are mentioned.

axioms (1)
  • domain assumption Group convolutional neural networks can be extended to perceptual color transformations while maintaining equivariance
    Invoked when introducing color-equivariant GCNNs as a direct extension of geometric GCNNs.
invented entities (1)
  • lifting layer for direct image transformation no independent evidence
    purpose: Transforms input images to achieve color equivariance without producing invalid RGB values
    New architectural component introduced to resolve the invalid-value problem in prior hue-equivariant networks.

pith-pipeline@v0.9.0 · 5732 in / 1154 out tokens · 21586 ms · 2026-05-24T00:12:51.676658+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

36 extracted references · 36 canonical work pages · 4 internal anchors

  1. [1]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012

    work page 2012

  2. [2]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  3. [3]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770--778, 2016

    work page 2016

  4. [4]

    Understanding image representations by measuring their equivariance and equivalence

    Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 991--999, 2015

    work page 2015

  5. [5]

    Invariant scattering convolution networks

    Joan Bruna and St \'e phane Mallat. Invariant scattering convolution networks. IEEE transactions on pattern analysis and machine intelligence, 35 0 (8): 0 1872--1886, 2013

    work page 2013

  6. [6]

    Group equivariant convolutional networks

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

    work page 2016

  7. [7]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263--1272. PMLR, 2017

    work page 2017

  8. [8]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652--660, 2017

    work page 2017

  9. [9]

    Matrix capsules with em routing

    Geoffrey E Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with em routing. In International conference on learning representations, 2018

    work page 2018

  10. [10]

    Hamiltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. Advances in neural information processing systems, 32, 2019

    work page 2019

  11. [11]

    Mitosis counting in breast cancer: Object-level interobserver agreement and comparison to an automatic method

    Mitko Veta, Paul J Van Diest, Mehdi Jiwa, Shaimaa Al-Janabi, and Josien PW Pluim. Mitosis counting in breast cancer: Object-level interobserver agreement and comparison to an automatic method. PloS one, 11 0 (8): 0 e0161286, 2016

    work page 2016

  12. [12]

    Color representation in deep neural networks

    Martin Engilberge, Edo Collins, and Sabine S \"u sstrunk. Color representation in deep neural networks. In 2017 IEEE International Conference on Image Processing (ICIP), pages 2786--2790. IEEE, 2017

    work page 2017

  13. [13]

    Impact of colour on robustness of deep neural networks

    Kanjar De and Marius Pedersen. Impact of colour on robustness of deep neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21--30, 2021

    work page 2021

  14. [14]

    Circle: Color invariant representation learning for unbiased classification of skin lesions

    Arezou Pakzad, Kumar Abhishek, and Ghassan Hamarneh. Circle: Color invariant representation learning for unbiased classification of skin lesions. In European Conference on Computer Vision, pages 203--219. Springer, 2022

    work page 2022

  15. [15]

    Learning invariances in neural networks from training data

    Gregory Benton, Marc Finzi, Pavel Izmailov, and Andrew G Wilson. Learning invariances in neural networks from training data. Advances in neural information processing systems, 33: 0 17605--17616, 2020

    work page 2020

  16. [16]

    On the generalization of equivariance and convolution in neural networks to the action of compact groups

    Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning, pages 2747--2755. PMLR, 2018

    work page 2018

  17. [17]

    A general theory of equivariant cnns on homogeneous spaces

    Taco S Cohen, Mario Geiger, and Maurice Weiler. A general theory of equivariant cnns on homogeneous spaces. Advances in neural information processing systems, 32, 2019 a

    work page 2019

  18. [18]

    Harmonic networks: Deep translation and rotation equivariance

    Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5028--5037, 2017

    work page 2017

  19. [19]

    Polar Transformer Networks

    Carlos Esteves, Christine Allen-Blanchette, Xiaowei Zhou, and Kostas Daniilidis. Polar transformer networks. arXiv preprint arXiv:1709.01889, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  20. [20]

    Deep scale-spaces: Equivariance over scale

    Daniel Worrall and Max Welling. Deep scale-spaces: Equivariance over scale. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  21. [21]

    Empowering networks with scale and rotation equivariance using a similarity convolution

    Zikai Sun and Thierry Blu. Empowering networks with scale and rotation equivariance using a similarity convolution. In The Eleventh International Conference on Learning Representations, 2022

    work page 2022

  22. [22]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  23. [23]

    Equivariant multi-view networks

    Carlos Esteves, Yinshuang Xu, Christine Allen-Blanchette, and Kostas Daniilidis. Equivariant multi-view networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1568--1577, 2019

    work page 2019

  24. [24]

    E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E Smidt, and Boris Kozinsky. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13 0 (1): 0 2453, 2022

    work page 2022

  25. [25]

    Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data

    Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In International Conference on Machine Learning, pages 3165--3176. PMLR, 2020

    work page 2020

  26. [26]

    Gauge equivariant convolutional networks and the icosahedral cnn

    Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolutional networks and the icosahedral cnn. In International conference on Machine learning, pages 1321--1330. PMLR, 2019 b

    work page 2019

  27. [27]

    Structuring representation geometry with rotationally equivariant contrastive learning

    Sharut Gupta, Joshua Robinson, Derek Lim, Soledad Villar, and Stefanie Jegelka. Structuring representation geometry with rotationally equivariant contrastive learning. In The Twelfth International Conference on Learning Representations, 2023

    work page 2023

  28. [28]

    A perception-based color space for illumination-invariant image processing

    Hamilton Y Chong, Steven J Gortler, and Todd Zickler. A perception-based color space for illumination-invariant image processing. ACM Transactions on Graphics (TOG), 27 0 (3): 0 1--7, 2008

    work page 2008

  29. [29]

    Color equivariant convolutional networks

    Attila Lengyel, Ombretta Strafforello, Robert-Jan Bruintjes, Alexander Gielisse, and Jan van Gemert. Color equivariant convolutional networks. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  30. [30]

    Differential geometry and lie groups, volume 12

    Jean Gallier and Jocelyn Quaintance. Differential geometry and lie groups, volume 12. Springer, 2020

    work page 2020

  31. [31]

    Mnist handwritten digit database

    Yann LeCun, Corinna Cortes, and CJ Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010

    work page 2010

  32. [32]

    3d shapes dataset

    Chris Burgess and Hyunjik Kim. 3d shapes dataset. https://github.com/deepmind/3dshapes-dataset/, 2018

    work page 2018

  33. [33]

    From detection of individual metastases to classification of lymph node status at the patient level: the camelyon17 challenge

    Peter Bandi, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, et al. From detection of individual metastases to classification of lymph node status at the patient level: the camelyon17 challenge. IEEE Transactions on Medical Imaging, 2018

    work page 2018

  34. [34]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    work page 2009

  35. [35]

    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

  36. [36]

    A fine-grained analysis on distribution shift, 2021

    Olivia Wiles, Sven Gowal, Florian Stimberg, Sylvestre Alvise-Rebuffi, Ira Ktena, Krishnamurthy Dvijotham, and Taylan Cemgil. A fine-grained analysis on distribution shift, 2021

    work page 2021