Pith. sign in

REVIEW 2 major objections 4 minor 300 references

This thesis argues that normalization and classification can be built once on Lie groups and gyrogroups and then instantiated across the manifolds used in practice.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

One Lie-group/gyrogroup framework unifies batch normalization and logistic-regression classifiers across SPD, rotation, correlation, Grassmannian, and constant-curvature manifolds, with new hyperbolic and SPD geometries.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A serious, well-documented thesis that synthesizes nine peer-reviewed papers; the genuinely new value is the unified normalization/classification framework and the pseudo-reductive gyrogroup analysis, with a real but narrow caveat about the 'all known gyrogroups' claim. the 2 major comments →

arxiv 2607.19305 v2 pith:VRN2AWJF submitted 2026-07-21 cs.LG cs.AImath.DG

Riemannian Deep Learning: Modules, Networks, and Geometries

Chen Ziheng This is my paper
classification cs.LG cs.AImath.DG
keywords Riemannian deep learningLie groupsgyrogroupsbatch normalizationmultinomial logistic regressionSPD manifoldshyperbolic embeddingscorrelation matrices
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This doctoral thesis seeks to unify Riemannian deep learning by lifting two core neural modules—batch normalization and multinomial logistic regression—from Euclidean space and from individual manifolds to broad algebraic classes of manifolds. It proposes LieBN for Lie groups with invariant metrics and GyroBN for pseudo-reductive gyrogroups, and proves that both control the Riemannian sample mean and variance. It then extends MLR to general Riemannian manifolds via Riemannian trigonometry, requiring only an explicit logarithmic map. The thesis also builds networks for hyperbolic, correlation, and SPD representations and introduces learnable and fast SPD metrics. If these claims hold, practitioners can reuse one normalization and one classification layer across the SPD, rotation, Grassmannian, hyperbolic, spherical, and correlation manifolds.

Core claim

The central claim is that the algebraic structures already present on common manifolds are enough to make deep-learning modules universal. Algorithm 1 (LieBN) uses group translations for centering and biasing and tangent-space scaling at the identity, and the thesis proves that under any invariant metric this controls the Riemannian mean and variance (Props. 61–62, 65). GyroBN extends the same principle to pseudo-reductive gyrogroups—groupoids in which gyrations of left inverses are trivial—and recovers LieBN as a special case. For classification, Riemannian Multinomial Logistic Regression (RMLR) replaces the point-to-hyperplane infimum with a Riemannian-trigonometric formula, so a manifold

What carries the argument

The load-bearing object is the pseudo-reductive gyrogroup: an algebraic structure satisfying the first three gyrogroup axioms plus pseudo-reduction (gyr[a,x]=id for a a left inverse of x). This axiom restores the left cancellation law ⊖x⊕(x⊕y)=y, which makes left gyrotranslation a gyroisometry; that isometry is what lets GyroBN center a batch and provably move its mean while controlling the variance. The companion machinery for classification is Riemannian trigonometry: RMLR computes logits as Riemannian distances to geodesic hyperplanes, reducing classification to evaluating a logarithmic map and an inner product. Together the two mechanisms turn group-like structure and a logarithmic map i

Load-bearing premise

The normalization guarantee rests on the assumption that every manifold in the framework's coverage is pseudo-reductive and that its gyrations are gyroisometries; if any admitted gyrogroup violates this, the provable mean-and-variance control for that manifold fails.

What would settle it

Compute gyr[a,x]z and the left-cancellation residual ∥⊖x⊕(x⊕y)⊖y∥ on the Grassmannian's ONB representation for a=⊖x; a single pair with nonzero residual at machine precision would break the proof that left gyrotranslation is a gyroisometry, and with it GyroBN's mean-control theorem for that geometry.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Networks on SO(n), SPD, full-rank correlation, Grassmannian, and constant-curvature manifolds can all use the same LieBN/GyroBN layer with provable mean and variance control instead of manifold-specific recipes.
  • Any manifold with an explicit geodesic distance and logarithmic map gets an intrinsic classifier; RMLR subsumes existing LogEig and hyperbolic classifiers as special cases.
  • Correlation matrices become a first-class representation: CorNets provide MLR, FC, and convolutional layers directly on Cor+(n) with accurate gradients.
  • SPD geometry can be adapted to data via parameterized matrix logarithms, and Cholesky-product metrics give fast, closed-form, numerically stable operators for SPD networks.

Where Pith is reading between the lines

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

  • If the pseudo-reductivity criterion is as general as the thesis asserts, the same normalization module should transfer to future gyrogroup-based architectures, including attention or transformer layers defined on gyrovector spaces; a clean test is to instantiate GyroBN inside a gyro-attention block and compare convergence.
  • The algebraic checklist (pseudo-reductivity plus gyroisometric gyrations) gives future work a quick necessary condition: any newly proposed manifold gyro-structure must be checked for these two properties before mean-variance normalization can be claimed.
  • RMLR's reliance only on a logarithmic map suggests a practical recipe for manifolds without closed-form geodesic distances: approximate the logarithmic map and plug it into the same classifier, which would extend intrinsic classification beyond the manifolds listed.
  • The learnable SPD metrics blur the line between architecture search and geometry search; one could optimize the metric parameters jointly with network weights to interpolate among known SPD geometries inside a single model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The thesis develops a unified framework for Riemannian deep learning, with three strands: reusable modules (batch normalization and multinomial logistic regression), manifold-specific network architectures (Proper Velocity hyperbolic networks, Busemann-based hyperbolic layers, full-rank correlation networks), and learnable/fast SPD geometries (ALEM and product Cholesky metrics). The central theoretical claims are that LieBN controls Riemannian sample mean and variance on Lie groups with invariant metrics (Props. 61–62, 65), that GyroBN extends this control to pseudo-reductive gyrogroups with gyroisometric gyrations (Thms. 75, 77–99), and that RMLR subsumes existing Euclidean/SPD/Lie classifiers as special cases (Sec. 4.3, Tab. 4.1). These claims are supported by proofs in Appendix B and by experiments across vision, signal processing, graph learning, and genomics. The thesis also introduces several new geometric constructions—CRIM, PV coordinates, ALEM, PCM/BWCM—and releases PyTorch toolboxes for LieBN and GyroBN.

Significance. If the results hold, the thesis delivers a genuinely general algebraic foundation for normalization: a single construction (GyroBN) covers Lie groups, the Grassmannian, constant-curvature spaces, and full-rank correlation manifolds, with provable control of first- and second-order sample statistics rather than heuristic normalization. The RMLR framework similarly unifies a dispersed set of manifold classifiers under one logarithmic-map formulation. The thesis is also strong on the practical side: closed-form operators for several geometries, explicit numerical stability and efficiency comparisons (Tabs. 6.13, 6.16, 6.17), and reproducible public code for the two normalization toolboxes. The main risk is the breadth claim attached to pseudo-reductivity: the central theorem is conditional on an algebraic property, and the universal statement about 'all known gyrogroups' is not proven as a classification theorem. However, the listed instantiations each come with proofs in Appendix B, so the conditional theorem itself is defensible. This is a scope caveat rather than a demonstrated flaw.

major comments (2)
  1. [Sec. 3.3.1, Def. 73] The text asserts that the GyroBN conditions (pseudo-reductivity and gyroisometric gyrations) 'are satisfied by all known gyrogroups in machine learning,' but the support is a per-manifold list (Prop. 74 for the Grassmannian; seven instantiations in Sec. 3.3.4 with proofs in App. B), not a general theorem. If any of the newly derived closed-form gyro-structures (Lorentz, projected hypersphere, row-wise correlation) failed pseudo-reductivity, the guaranteed mean/variance control would not follow for that manifold. I recommend weakening the universal claim to 'all gyrogroups considered in this thesis' or adding a summary table that explicitly verifies pseudo-reductivity and gyroisometry for each instantiation. This is a correctness-risk concern about scope, not a counterexample to the conditional theorem.
  2. [Secs. 2.9.5 and 3.3.4] The K-stereographic gyration formula is stated to be valid only when D_st > 0 (Eq. 2.147), and positive-curvature logarithm/parallel-transport maps are restricted away from antipodal configurations. Since GyroBN on the sphere and projected hypersphere relies on gyroisometries at arbitrary batch points, the manuscript should clarify whether the normalization guarantee extends to points where D_st = 0 or to antipodal pairs, and if not, state the domain restriction in the GyroBN theorem. Similarly, the radius-model formulas exclude singular configurations; a short paragraph on domain validity would make the coverage claim precise.
minor comments (4)
  1. [Tab. 3.11 (and Tab. 6.8)] NTU60 results are reported as single accuracy values, whereas other tables report mean±std (e.g., Tab. 3.9). Please report mean±std over multiple runs, or state the protocol used for these entries, so the reader can assess variance.
  2. [Sec. 3.2.1] The claim that CRIM is 'the first non-trivial right-invariant SPD metric' is strong. Please provide a brief literature pointer or a sentence substantiating that no other non-commutative right-invariant SPD metric appears in the prior work.
  3. [Sec. 3.2.5.1, Tab. 3.6] The notation ⊕_LieAI and ⊙ for the power-deformed operations is introduced only in the table; a one-sentence definition of how P^θ pulls back the group operation would improve readability for a reader who has not followed the deformation discussion.
  4. [Sec. 1.3] The list of excluded papers is helpful for provenance, but the phrasing 'Since Chapter X follows ...' in the final paragraph could be misread as saying the excluded papers are prerequisites. Consider moving that explanatory paragraph to an appendix or editorial preface.

Circularity Check

0 steps flagged

No significant circularity; the conditional algebraic theorems are self-contained.

full rationale

I find no load-bearing circularity. The central claims are conditional on explicitly stated algebraic hypotheses—invariant metrics for LieBN and pseudo-reductive gyrogroups with gyroisometric gyrations for GyroBN—and the proofs are carried out from the definitions in the appendix, not obtained by fitting parameters or by invoking the author's own prior papers as the proof. The statements that GyroBN recovers LieBN (Tab. 3.13) and that RMLR subsumes LogEig/Lie classifiers (Tab. 4.1) are mathematical containments visible in the equations; they are specializations, not predictions forced by fitted inputs. The only caveat worth flagging is the breadth claim in Sec. 3.3.1 that 'These conditions are satisfied by all known gyrogroups in machine learning,' which is supported case-by-case (Prop. 74, Sec. 3.3.4, App. B) rather than by a general classification theorem. That is a scope/correctness risk for the universality claim, not a circularity in the conditional GyroBN theorem. The self-citations to the author's own publications are compilation artifacts of a thesis and are not load-bearing: none of the derivations reduces to an unverified self-citation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 4 invented entities

The central claims rest on standard Riemannian geometry (O'Neill, Bridson-Haefliger, Absil-Boumal), on the intrinsic Gaussian population model (Eq. 3.7), on per-manifold verification that the relevant gyro-structures are pseudo-reductive, and on standard matrix-calculus identities (Daleckii-Krein, Cholesky differentials). The principal non-standard premise is the assertion that all known machine-learning gyrogroups satisfy the pseudo-reductive law (Sec. 3.3.1), verified case by case rather than proven in general. Free parameters are dominated by the per-dataset deformation exponent θ, which the headline 'Best θ' numbers depend on.

free parameters (4)
  • θ (metric deformation exponent) = per-dataset values such as 1.5, 0.5, -0.5
    Deformation exponent for (θ,α,β)-AIM, θ-LCM, θ-CRIM, θ-PCM, θ-BWCM; tuned per data set from candidate grids and reported as 'Best θ' in Tables 3.9-3.10 and A.15. The headline accuracies depend on this selection.
  • (α, β) metric-family parameters = (1, 0)
    The thesis states '(α,β) = (1,0)' is simply set and only θ is tuned; these are fixed degrees of freedom of the metric family rather than derived values.
  • PHCM hemisphere weights α_i = 1 (canonical)
    PHCM is instantiated with all weights equal to 1 (Sec. 2.9.2); other weightings are part of the design space but unexplored.
  • Curvature K of constant-curvature models = -1 or +1
    Standard fixed curvature values for hyperbolic/spherical models; assumed rather than fitted in GyroBN and PV experiments.
axioms (6)
  • domain assumption Intrinsic Gaussian density on Lie groups (Eq. 3.7) with MLE equal to the sample Fréchet mean.
    Prop. 61 anchors LieBN's mean control; the MLE identification is inherited from prior studies (Rem. 63 cites [176, 35, 34]) and is not re-derived in full generality in the thesis.
  • domain assumption Every machine-learning gyrogroup is pseudo-reductive with gyroisometric gyrations.
    Sec. 3.3.1 asserts this for 'all known gyrogroups'; the document proves per-manifold (Prop. 74 for Grassmannian) and corrects the prior Nguyen-Yang isometry proof, so the general claim is asserted, not proven from axioms.
  • standard math Weighted Fréchet means exist and are unique on the relevant geodesic balls.
    Used for running-statistic updates (Alg. 1) and WFM computations; cited to [2, Thm. 2.1].
  • standard math Busemann functions and horospheres behave as in Hadamard spaces.
    Sec. 5.3 builds BMLR/BFC on Busemann functions of the Poincaré and Lorentz models, relying on Bridson-Haefliger Ch. II.8 facts.
  • standard math Daleckii-Krein formula and Cholesky differential identities hold as stated.
    All SPD backpropagation (Sec. 2.8) and the CRIM operators (Thm. 69) use these identities.
  • standard math Constant-curvature spaces are classified by dimension and curvature (O'Neill Cor. 8.25).
    Sec. 2.9.5 invokes this classification to justify the model zoo (Poincaré, Lorentz, Beltrami-Klein, sphere, projected hypersphere) used in GyroBN and PV experiments.
invented entities (4)
  • Pseudo-reductive gyrogroup independent evidence
    purpose: Algebraic foundation for GyroBN; relaxes left-reduction (G4) so left cancellation and gyroisometry of gyrotranslation hold without full gyrogroup axioms.
    Falsifiable through normalization experiments on the Grassmannian, five CCSs, and correlation manifolds (Secs. 3.3.4-3.3.5), and through the corrected proof of Grassmannian translation isometry.
  • CRIM (Cholesky Right-Invariant Metric) independent evidence
    purpose: First claimed non-trivial right-invariant SPD metric; enables right-translation LieBN on SPD manifolds.
    Closed-form geodesic/exponential/log operators stated in Thm. 69 and geodesic completeness in Cor. 70; evaluated on Radar, HDM05, FPHA (Tab. 3.9).
  • Proper Velocity (PV) coordinate model of hyperbolic space independent evidence
    purpose: Unconstrained hyperbolic representation that avoids FP32 failures of Möbius/Lorentz scalar multiplication.
    Falsifiable via failure-rate and round-trip error measurements (Tabs. 5.1-5.2) and via accuracy on image, graph, and TEB tasks.
  • ALEM (Adaptive Log-Euclidean Metric) independent evidence
    purpose: Learnable SPD geometry via parameterized matrix logarithms, adapting the metric during network training.
    Learned basis parameters are visualized (Fig. 6.2) and accuracy/efficiency measured on HDM05, FPHA, AFEW (Tabs. 6.2-6.4).

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Riemannian Deep Learning: Modules, Networks, and Geometries." pith.science (2026). https://pith.science/paper/VRN2AWJF

@misc{pith2026260719305,
  author       = {Pith},
  title        = {Pith review of: Riemannian Deep Learning: Modules, Networks, and Geometries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VRN2AWJF}},
  note         = {Machine review of arXiv:2607.19305}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep neural networks on manifold-valued representations have attracted growing interest, but many basic components remain tied to specific manifolds, rely on Euclidean approximations, or require costly and numerically fragile geometric operations. This thesis develops a unified framework for Riemannian deep learning from three complementary perspectives: reusable neural modules, manifold-specific network architectures, and the design of underlying geometries. It generalizes batch normalization from Euclidean spaces and individual manifolds to broad classes of Lie groups and gyrogroups, and extends multinomial logistic regression from Euclidean space to SPD manifolds and then to general Riemannian manifolds. It further develops neural networks for several important geometric representations, including an unconstrained model of hyperbolic space, Busemann-based hyperbolic learning, and full-rank correlation matrices. Finally, it introduces adaptive and computationally efficient Riemannian metrics on SPD manifolds, including learnable Log-Euclidean geometries and fast, stable Cholesky-based geometries. The proposed methods are supported by theoretical analysis and validated through numerical experiments and applications in vision, signal processing, graph learning, and genomics.

Figures

Figures reproduced from arXiv: 2607.19305 by Chen Ziheng.

Figure 2.1
Figure 2.1. Figure 2.1: The black stars denote 2 × 2 correlation matrices, while the red, green, and blue dots denote corresponding SPD matrices. The black dots denote the boundary of the SPD cone. size pairwise dependencies rather than raw variances. Only recently have Riemannian structures been developed for correlation matrices. The space of n×n full-rank correlation matrices, denoted by Cor+ (n), forms a Rieman￾nian manifol… view at source ↗
Figure 3.1
Figure 3.1. Figure 3.1: Illustration of LieBN on the SPD, rotation, and correlation Lie groups. The [PITH_FULL_IMAGE:figures/full_fig_p086_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: Minimal examples of applying LieBN. • Validation of the effectiveness of our LieBN framework by extensive experiments on different geometries.2 Outline. Sec. 3.2.2 recalls the invariant metrics and Lie structures used by LieBN. Sec. 3.2.3 revisits Euclidean BN and RBN. Sec. 3.2.4 develops LieBN on Lie groups under left- and right-invariant metrics and establishes its statistical control. Sec. 3.2.5 insta… view at source ↗
Figure 3.3
Figure 3.3. Figure 3.3: Visualization of input and output 30 × 30 SPD matrices in LieBN using 2 × 2 Riemannian t-SNE embeddings. The first row shows the input and output under different metrics. Due to the significant difference in magnitude between the t-SNE embeddings of LieBN’s input and output, the second row separately visualizes the LieBN output (at a smaller scale). the CRIM-based LieBN shows less efficiency, due to the … view at source ↗
Figure 3.4
Figure 3.4. Figure 3.4: Test accuracy curves corresponding to Tab. [PITH_FULL_IMAGE:figures/full_fig_p105_3_4.png] view at source ↗
Figure 3.5
Figure 3.5. Figure 3.5: Illustration of GyroBN on manifold-valued data. Blue points, green points, [PITH_FULL_IMAGE:figures/full_fig_p107_3_5.png] view at source ↗
Figure 3.6
Figure 3.6. Figure 3.6: Minimal examples of applying GyroBN. 3.3.2 Pseudo-Reductive Gyrogroups Given a gyrogroup (G, ⊕), the left gyrotranslation by x ∈ G is defined as Lx : G → G, Lx(y) = x ⊕ y, ∀y ∈ G. (3.32) If any gyrotranslation is a gyroisometry, we can use gyrotranslation to center manifold￾valued samples for the normalization layer. Nguyen and Yang [159] shows that any left gyrotranslation on the SPD and Grassmannian ma… view at source ↗
Figure 3.7
Figure 3.7. Figure 3.7: A conceptual comparison of the derivation logic in our work with that in [PITH_FULL_IMAGE:figures/full_fig_p110_3_7.png] view at source ↗
Figure 3.8
Figure 3.8. Figure 3.8: Visualization of GyroBN across different geometries. Blue and green points [PITH_FULL_IMAGE:figures/full_fig_p131_3_8.png] view at source ↗
Figure 3.9
Figure 3.9. Figure 3.9: Training and testing curves of 1-block GyroGr on two NTU data sets. [PITH_FULL_IMAGE:figures/full_fig_p134_3_9.png] view at source ↗
Figure 4.1
Figure 4.1. Figure 4.1: Conceptual illustration of SPD hyperplanes induced by [PITH_FULL_IMAGE:figures/full_fig_p147_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: Illustration of the deformation (left) and Venn diagram (right) of metrics on SPD manifolds, where IEM, SREM, and 1 4 PAM denote Inverse Euclidean Metric, Square Root Euclidean Metric, and Polar Affine Metric scaled by 1/4, respectively. 4.3.3 SPD Multinomial Logistic Regressions This section showcases our RMLR framework on the SPD manifold. We first systemati￾cally discuss the power-deformed geometries … view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: Conceptual illustration of SPD hyperplanes induced by five families of [PITH_FULL_IMAGE:figures/full_fig_p154_4_3.png] view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: Conceptual illustration of a Lie hyperplane. Each pair of antipodal black [PITH_FULL_IMAGE:figures/full_fig_p157_4_4.png] view at source ↗
Figure 5.1
Figure 5.1. Figure 5.1: Illustration: red curves are different horospheres of [PITH_FULL_IMAGE:figures/full_fig_p182_5_1.png] view at source ↗
Figure 5.2
Figure 5.2. Figure 5.2: Validation accuracy curves on ImageNet-1k. [PITH_FULL_IMAGE:figures/full_fig_p189_5_2.png] view at source ↗
Figure 5.3
Figure 5.3. Figure 5.3: Illustration of the Log-Euclidean 1D convolution with two kernels. The [PITH_FULL_IMAGE:figures/full_fig_p198_5_3.png] view at source ↗
Figure 5.4
Figure 5.4. Figure 5.4: Illustration of the PHCM convolution and MLR. The multi-channel input [PITH_FULL_IMAGE:figures/full_fig_p200_5_4.png] view at source ↗
Figure 5.5
Figure 5.5. Figure 5.5: Illustration of the decision hyperplanes in the correlation MLRs under five [PITH_FULL_IMAGE:figures/full_fig_p204_5_5.png] view at source ↗
Figure 5.6
Figure 5.6. Figure 5.6: Distribution of per-sample coefficients of variation of diagonal variances on [PITH_FULL_IMAGE:figures/full_fig_p206_5_6.png] view at source ↗
Figure 5.7
Figure 5.7. Figure 5.7: Distribution of per-sample coefficients of variation of diagonal variances [PITH_FULL_IMAGE:figures/full_fig_p207_5_7.png] view at source ↗
Figure 5.8
Figure 5.8. Figure 5.8: Distribution of ratios of diagonal to off-diagonal entries on FPHA. [PITH_FULL_IMAGE:figures/full_fig_p208_5_8.png] view at source ↗
Figure 5.9
Figure 5.9. Figure 5.9: Distribution of ratios of diagonal to off-diagonal entries on HDM05. [PITH_FULL_IMAGE:figures/full_fig_p208_5_9.png] view at source ↗
Figure 6.1
Figure 6.1. Figure 6.1: Accuracy curves on the FPHA data set. SPDNet SPDNetBN ALog MUL DIV RELU 85.73±0.80 86.83±0.74 87.8±0.71 88.07±1.13 86.65±0.68 [PITH_FULL_IMAGE:figures/full_fig_p229_6_1.png] view at source ↗
Figure 6.2
Figure 6.2. Figure 6.2: Visualization of parameters in the ALog layer on the HDM05 and FPHA [PITH_FULL_IMAGE:figures/full_fig_p231_6_2.png] view at source ↗
Figure 6.3
Figure 6.3. Figure 6.3: Geodesic interpolation of SPD matrices under different Riemannian metrics. [PITH_FULL_IMAGE:figures/full_fig_p247_6_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

300 extracted references · 19 linked inside Pith

  1. [1]

    ICCV , year =

    Is second-order information helpful for large-scale visual recognition? , author =. ICCV , year =

  2. [2]

    CVPR , year =

    Towards faster training of global covariance pooling networks by iterative matrix square root normalization , author =. CVPR , year =

  3. [3]

    ICCV , year =

    Matrix backpropagation for deep networks with structured layers , author =. ICCV , year =

  4. [4]

    arXiv preprint arXiv:1509.07838 , year=

    Training deep networks with structured layers by matrix backpropagation , author=. arXiv preprint arXiv:1509.07838 , year=

    Pith/arXiv arXiv

  5. [5]

    2008 , publisher=

    Functions of matrices: theory and computation , author=. 2008 , publisher=

    2008

  6. [6]

    2007 , publisher=

    Positive Definite Matrices , author=. 2007 , publisher=

    2007

  7. [7]

    2013 , series=

    Matrix analysis , author=. 2013 , series=

    2013

  8. [8]

    BMVC , year=

    Improved bilinear pooling with cnns , author=. BMVC , year=

  9. [9]

    NeurIPS , year =

    Backpropagation-friendly eigendecomposition , author =. NeurIPS , year =

  10. [10]

    Essentials of Pad

    George Jr, A and others , year=. Essentials of Pad

  11. [11]

    The theory and application of the Pad

    Baker Jr, George A , year=. The theory and application of the Pad

  12. [12]

    Baker, George Allen and Gammel, John L , year=. The Pad

  13. [13]

    Van Assche, Walter , journal=. Pad

  14. [14]

    General theorems on the convergence of sequences of Pad

    Wall, HS , journal=. General theorems on the convergence of sequences of Pad

  15. [15]

    An Overview , author=

    On Rational Function Techniques and Pade Approximants. An Overview , author=. 2002 , publisher=

    2002

  16. [16]

    ICCV , year =

    Bilinear cnn models for fine-grained visual recognition , author =. ICCV , year =

  17. [17]

    CVPR , year =

    Imagenet: A large-scale hierarchical image database , author =. CVPR , year =

  18. [18]

    CVPR , year =

    G2DeNet: Global Gaussian distribution embedding network and its application to visual recognition , author =. CVPR , year =

  19. [19]

    CVPR , year =

    Deep global generalized Gaussian networks , author =. CVPR , year =

  20. [20]

    IEEE TPAMI , year =

    Robust Differentiable SVD , author =. IEEE TPAMI , year =

  21. [21]

    Dang and K

    Z. Dang and K. M. Yi and Y. Hu and F. Wang and P. Fua and M. Salzmann , booktitle =

  22. [22]

    Dang and K.M

    Z. Dang and K.M. Yi and F. Wang and Y. Hu and P. Fua and M. Salzmann , journal =

  23. [23]

    IEEE TPAMI , year =

    Deep cnns meet global covariance pooling: Better representation and generalization , author =. IEEE TPAMI , year =

  24. [24]

    CVPR , year =

    An investigation into the stochasticity of batch whitening , author =. CVPR , year =

  25. [25]

    ICCV , year =

    Understanding generalized whitening and coloring transform for universal style transfer , author =. ICCV , year =

  26. [26]

    CVPR , year =

    Image-to-image translation via group-wise deep whitening-and-coloring transformation , author =. CVPR , year =

  27. [27]

    ECCV , year =

    Semantic segmentation with second-order pooling , author =. ECCV , year =

  28. [28]

    CVPR , year =

    Decorrelated batch normalization , author =. CVPR , year =

  29. [29]

    1981 , publisher=

    Approximation theory and methods , author=. 1981 , publisher=

    1981

  30. [30]

    CVPR , year =

    Deep residual learning for image recognition , author =. CVPR , year =

  31. [31]

    Welinder and S

    P. Welinder and S. Branson and T. Mita and C. Wah and F. Schroff and S. Belongie and P. Perona , Institution =

  32. [32]

    arXiv preprint arXiv:1306.5151 , year=

    Fine-grained visual classification of aircraft , author=. arXiv preprint arXiv:1306.5151 , year=

    Pith/arXiv arXiv

  33. [33]

    2013 , address =

    3D Object Representations for Fine-Grained Categorization , booktitle =. 2013 , address =

    2013

  34. [34]

    arXiv preprint arXiv:1409.1556 , year=

    Very deep convolutional networks for large-scale image recognition , author=. arXiv preprint arXiv:1409.1556 , year=

    Pith/arXiv arXiv

  35. [35]

    Novel Dataset for Fine-Grained Image Categorization

    Aditya Khosla and Nityananda Jayadevaprakash and Bangpeng Yao and Li Fei-Fei. Novel Dataset for Fine-Grained Image Categorization. First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition. 2011

    2011

  36. [36]

    2003 , publisher=

    Introductory lectures on convex optimization: A basic course , author=. 2003 , publisher=

    2003

  37. [37]

    ICCV , year =

    Dynamic label graph matching for unsupervised video re-identification , author =. ICCV , year =

  38. [38]

    CVPR , year =

    Compact bilinear pooling , author =. CVPR , year =

  39. [39]

    CVPR , year =

    Low-Rank Bilinear Pooling for Fine-Grained Classification , author =. CVPR , year =

  40. [40]

    CVPR , year =

    Monet: Moments embedding network , author =. CVPR , year =

  41. [41]

    CVPR , year =

    Kernel Pooling for Convolutional Neural Networks , author =. CVPR , year =

  42. [42]

    ECCV , year =

    Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition , author =. ECCV , year =

  43. [43]

    ECCV , year =

    Grassmann Pooling as Compact Homogeneous Bilinear Pooling for Fine-Grained Visual Classification , author =. ECCV , year =

  44. [44]

    NeurIPS , year =

    Learning deep bilinear transformation for fine-grained image representation , author =. NeurIPS , year =

  45. [45]

    ECCV , year =

    Toward Faster and Simpler Matrix Normalization via Rank-1 Update , author =. ECCV , year =

  46. [46]

    arXiv preprint arXiv:1211.7102 , year=

    SVD based image processing applications: state of the art, contributions and research challenges , author=. arXiv preprint arXiv:1211.7102 , year=

    Pith/arXiv arXiv

  47. [47]

    2007 , publisher=

    Understanding complex datasets: data mining with matrix decompositions , author=. 2007 , publisher=

    2007

  48. [48]

    arXiv preprint arXiv:1312.6034 , year=

    Deep inside convolutional networks: Visualising image classification models and saliency maps , author=. arXiv preprint arXiv:1312.6034 , year=

    Pith/arXiv arXiv

  49. [49]

    ECCV , year =

    Visualizing and understanding convolutional networks , author =. ECCV , year =

  50. [50]

    ICCV , year =

    Interpretable explanations of black boxes by meaningful perturbation , author =. ICCV , year =

  51. [51]

    ICCV , year =

    Understanding deep networks via extremal perturbations and smooth masks , author =. ICCV , year =

  52. [52]

    CVPR , year =

    There and back again: Revisiting backpropagation saliency methods , author =. CVPR , year =

  53. [53]

    arXiv preprint arXiv:1412.6806 , year=

    Striving for simplicity: The all convolutional net , author=. arXiv preprint arXiv:1412.6806 , year=

    Pith/arXiv arXiv

  54. [54]

    ICCV , year =

    Grad-cam: Visual explanations from deep networks via gradient-based localization , author =. ICCV , year =

  55. [55]

    BMVC , year=

    Rise: Randomized input sampling for explanation of black-box models , author=. BMVC , year=

  56. [56]

    NeurIPS , year =

    Real time image saliency for black box classifiers , author =. NeurIPS , year =

  57. [57]

    CVIU , volume=

    Classifier-agnostic saliency map extraction , author=. CVIU , volume=. 2020 , publisher=

    2020

  58. [58]

    CVPR , year =

    Densely connected convolutional networks , author =. CVPR , year =

  59. [59]

    CVPR , year =

    Rethinking the inception architecture for computer vision , author =. CVPR , year =

  60. [60]

    Psychometrika , volume=

    The approximation of one matrix by another of lower rank , author=. Psychometrika , volume=. 1936 , publisher=

    1936

  61. [61]

    Delving Deeper into the Whorl of Flower Segmentation

    Maria-Elena Nilsback and Andrew Zisserman. Delving Deeper into the Whorl of Flower Segmentation. Image and Vision Computing. 2009

    2009

  62. [62]

    Mathematische Nachrichten , volume=

    A note on von Neumann's trace inequalitv , author=. Mathematische Nachrichten , volume=. 1991 , publisher=

    1991

  63. [63]

    Monatshefte f

    A trace inequality of John von Neumann , author=. Monatshefte f. 1975 , publisher=

    1975

  64. [64]

    ICCV , year =

    Why Approximate Matrix Square Root Outperforms Accurate SVD in Global Covariance Pooling? , author =. ICCV , year =

  65. [65]

    2015 , URL =

    Inceptionism: Going Deeper into Neural Networks , author =. 2015 , URL =

    2015

  66. [66]

    The Annals of Applied Statistics , volume=

    Non-Euclidean statistics for covariance matrices, with applications to diffusion tensor imaging , author=. The Annals of Applied Statistics , volume=. 2009 , publisher=

    2009

  67. [67]

    SIAM journal on matrix analysis and applications , volume=

    Geometric means in a novel vector space structure on symmetric positive-definite matrices , author=. SIAM journal on matrix analysis and applications , volume=. 2007 , publisher=

    2007

  68. [68]

    IJCV , year =

    A Riemannian framework for tensor computing , author =. IJCV , year =

  69. [69]

    Afsari, Bijan , booktitle=

  70. [70]

    NeurIPS , year =

    A new metric on the manifold of kernel matrices with application to matrix geometric means , author =. NeurIPS , year =

  71. [71]

    arXiv preprint arXiv:1708.02862 , year=

    Webvision database: Visual learning and understanding from web data , author=. arXiv preprint arXiv:1708.02862 , year=

    Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2006.07159 , year=

    Are we done with ImageNet? , author=. arXiv preprint arXiv:2006.07159 , year=

    Pith/arXiv arXiv 2006

  73. [73]

    Communications of the ACM , volume=

    Solution of the matrix equation AX+ XB= C [F4] , author=. Communications of the ACM , volume=. 1972 , publisher=

    1972

  74. [74]

    CVPR , year =

    Iterative normalization: Beyond standardization towards efficient whitening , author =. CVPR , year =

  75. [75]

    NeurIPS , year =

    Universal Style Transfer via Feature Transforms , author =. NeurIPS , year =

  76. [76]

    CVPR , year =

    RobustNet: Improving Domain Generalization in Urban-Scene Segmentation via Instance Selective Whitening , author =. CVPR , year =

  77. [77]

    ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift f

    Iterative berechung der reziproken matrix , author=. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift f. 1933 , publisher=

    1933

  78. [78]

    arXiv preprint arXiv:2104.10935 , year=

    So-vit: Mind visual tokens for vision transformer , author=. arXiv preprint arXiv:2104.10935 , year=

    Pith/arXiv arXiv

  79. [79]

    ICML , year =

    Batch normalization: Accelerating deep network training by reducing internal covariate shift , author =. ICML , year =

  80. [80]

    ICCV , year =

    Switchable whitening for deep representation learning , author =. ICCV , year =

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.