Pith. sign in

REVIEW 4 major objections 7 minor 138 references

Quasi-SVD claims that a fully parallel, differentiable factorisation with one exactly-orthogonal Lie-parameterised factor can replace sequential SVD, achieving SSIM 0.89–0.94, 3–20× speedups, and real-time throughput above 25 FPS in clinica

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 ↗

Quasi-SVD learns a Lie-constrained approximate SVD whose one-sided orthogonal factor enables GPU-parallel medical imaging decompositions above 25 FPS with SSIM 0.89–0.94.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Genuinely novel idea and real medical benchmarks, but the 'exact orthogonality' claim doesn't hold under the default Taylor implementation, and the ULM decomposition numbers are worse than the abstract implies — worth refereeing with revisions. the 4 major comments →

arxiv 2607.25967 v1 pith:I2PFX233 submitted 2026-07-28 cs.CV cs.LGcs.NAmath.NA

Quasi-SVD: Learning a Lie-constrained matrix factorisation for real-time imaging

Christopher Hahne This is my paper
classification cs.CV cs.LGcs.NAmath.NA MSC 65F15
keywords Quasi-SVDsingular value decompositionLie group exponential mapGPU parallel factorisationreal-time medical imagingultrasound localisation microscopyMueller matrix polarimetrylearned matrix factorisation
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

Quasi-SVD is the claim that the singular value decomposition — the inherently sequential, rate-limiting step in ultrasound localisation microscopy and Mueller matrix polarimetry — can be replaced by a learned, fully parallel factorisation fast enough for live clinical workflows. The method makes one factor, X, exactly orthogonal by construction: X is the matrix exponential of a skew-symmetric matrix, so it lives in the special orthogonal group and never drifts. The singular values D and the second factor Y are recovered through soft-constrained training losses, and a lemma proves this asymmetric design is sufficient: any classical SVD can be rewritten with a determinant-one left factor by flipping a column-sign pair, so one hard orthogonality constraint is provably enough for a valid factorisation A = X D Y^T. Across two clinical regimes — massive batches of 3×3 Mueller matrices and single 500×500 ultrasound clutter-filtering matrices — the Lie-constrained network reports SSIM 0.89–0.94, 3–20× speedups over classical and randomised SVD baselines, and throughput above 25 FPS, including transfer to a second polarimetric instrument. If right, the paper turns previously offline or sub-real-time imaging steps into live ones and reframes SVD approximation as a deployment-oriented design choice rather than a purely numerical one.

Core claim

Quasi-SVD factorises A as A = X D Y^T with X ∈ SO(m) exactly orthogonal by construction: a network or closed-form rule predicts a skew-symmetric matrix S, and X = exp(S). The singular values are extracted via spectral invariance, d = diag^{-1}(√|X^T A A^T X|), and Y = A^T X D^{-1} completes the factorisation. Lemma 1 justifies the one-sided design: flipping one column sign in U and the matching column in V puts any SVD's left factor in SO(m), so a Lie-parameterised left factor always exists and the right factor needs only a soft orthogonality penalty. The paper claims this design is provably sufficient, fully parallel (span O(log m) versus O(n log n) for classical sequential solvers), and em

What carries the argument

The central mechanism is the exponential map from skew-symmetric matrices to the special orthogonal group, exp: so(m) → SO(m). Because exp(S)^T exp(S) = I, the learned left factor X is exactly orthogonal by construction — no soft penalty, no drift. Only X carries this hard constraint: the singular values come from the orthogonal similarity transform d = diag^{-1}(√|X^T A A^T X|), and Y = A^T X D^{-1} is guided toward orthogonality by a soft loss. Lemma 1 ('asymmetric orthogonality suffices') is the load-bearing result: sign-flipping one column pair of a classical SVD yields a left factor in SO(m), so one hard constraint is a theorem, not a heuristic. The LieNN variant predicts the skew-symme

Load-bearing premise

The load-bearing premise, acknowledged in the paper itself, is transfer: per-pixel polarimetric models drop in absolute accuracy on a second instrument (Table 3), the ULM model is trained on 20 of 25 sequences and tested on only 5 (Section 4.1), and the synthetic evaluation assumes full-rank inputs (Supp. D.1) — if any of these transfer conditions fails in a new clinical setting, the real-time deployment claim gives way.

What would settle it

Take the published NPP-trained per-pixel LieNN+RNN model and apply it to polarimetric data from a third, unseen optical instrument (different magnification, illumination, and spectral response), comparing per-pixel azimuth to the reference solver: if SSIM falls well below the 0.83 observed on the MUC transfer test (Table 3), the cross-instrument robustness claim fails. In parallel, feed the ULM model a rank-deficient Casorati matrix (e.g., from saturated or empty frames), which the paper's Supp. D.1 synthetic setup explicitly excludes, and check whether reconstruction SSIM and the 3.28 ms runt

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

If this is right

  • Ultrasound localisation microscopy clutter filtering, previously offline or throttled to roughly 4 FPS, becomes a live step: the LieNN+RNN variant reconstructs filtered frames at SSIM 0.89 in 3.28 ms per 500×500 matrix, more than 20× faster than the reference solver.
  • Per-pixel Mueller matrix polarimetry for neurosurgical tissue characterisation runs above 25 FPS at 3×3 scale, with SSIM up to 0.94 on the brain-tissue test set, enabling intraoperative polarimetric feedback.
  • Enforcing exact orthogonality on one factor via the Lie exponential eliminates the numerical drift that soft-constrained learned decompositions accumulate, while remaining fully differentiable and trainable in a few GPU-hours per modality.
  • The per-pixel formulation transfers across polarimetric instruments without retraining: models trained on one optical setup applied to a second setup keep SSIM 0.83 with the RNN variant, supporting instrument-agnostic deployment.
  • The work–span analysis indicates the runtime gap over classical solvers widens as matrix size and batch size grow, so the advantage is largest exactly in the high-throughput regimes clinical pipelines need.

Where Pith is reading between the lines

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

  • Editorial inference: the same recipe — one hard Lie-constrained factor plus soft recovery of the rest — is portable to any SVD-heavy, latency-sensitive pipeline (e.g., MRI coil compression, video background subtraction, or large-scale batched low-rank factorisation) whose downstream task consumes reconstruction quality rather than exact spectral components; the paper demonstrates the trade but doe
  • Editorial inference: because only the left factor is exactly orthogonal, downstream tasks that consume the right factor directly would inherit its error. The paper's own ULM table shows right-factor orthogonality error near 20.8 (versus 0 for the reference) despite SSIM 0.89 — so in that regime the output is a good reconstruction but a poor SVD, a distinction the headline fidelity metric can mask.
  • Editorial inference: the reported cross-instrument transfer is plausibly a consequence of the strictly per-pixel, context-free formulation. Adding spatial or temporal context — which the paper names as future work — would likely raise in-distribution accuracy but could erode exactly the instrument-agnostic generalisation it highlights; this tension is directly testable.
  • Editorial inference: the real-time claim is benchmarked at two matrix extremes, so a natural stress test is a third regime — for instance, rank-deficient inputs, which the paper's own synthetic evaluation (Supp. D.1) explicitly excludes. If a clinically encountered rank-deficient Casorati or Mueller matrix breaks the learned mapping, the deployment envelope is narrower than the headline throughput
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

4 major / 7 minor

Summary. The paper proposes Quasi-SVD, a learned, GPU-parallel approximation of the singular value decomposition. The central idea is to enforce exact orthogonality on one factor, X, by parameterizing it as the matrix exponential of a skew-symmetric matrix over the Lie algebra so(m), while recovering the singular values D and the right factor Y through closed-form estimates and soft losses. The framework is instantiated in several variants (analytic, unconstrained neural net, Lie-constrained neural net, end-to-end) with optional RNN refinement, and evaluated on two medical imaging tasks: Mueller matrix polarimetry (3x3 per-pixel matrices) and ultrasound localization microscopy (500x500 Casorati matrices). The reported results show SSIM 0.89–0.94 for the best LieNN+RNN variants, speedups relative to cuSOLVER and randomised SVD, and cross-instrument transfer to the MUC dataset. The authors argue that the asymmetric hard-orthogonality constraint is sufficient for a valid factorization and is key to eliminating numerical drift found in soft-constrained methods.

Significance. If the advertised property — exact orthogonality of one learned factor via the Lie exponential — were actually delivered at the tested scales, this would be a useful contribution: it combines a principled geometric constraint with a fully parallel, differentiable computation graph, and it is validated on realistic clinical workloads with a comparison against strong baselines. The paper also contains useful elements: a clear work–span analysis, an ablation over model variants, a domain-shift experiment, and a planned code release. However, the central theoretical claim is not supported by the implementation described in the manuscript. The default 9-term Taylor approximation of exp(S) is not exactly orthogonal for finite k, yet the paper claims 'exact orthogonality' and reports ||X||_ort = 0.00 at m=500. This inconsistency affects the core contribution, and the factorization accuracy numbers in the ULM table are very weak. The empirical reconstruction SSIM and speed results may still stand, but the paper's framing as a hard-guarantee method needs major revision.

major comments (4)
  1. [Section 3.6 / Eq. (3) / Table 4] The paper's central claim is that the Lie parameterization 'guarantees exact orthogonality' for the left factor. However, Section 3.6 states that for general matrices the default implementation uses a 9-term Taylor expansion of exp(S). For a skew-symmetric S and finite truncation order k, exp_k(S)^T exp_k(S) ≠ I. This is not a minor numerical detail: Eqs. (5)–(6) reconstruct A as X D Y^T = X X^T A, so the reconstruction identity depends on X being exactly orthogonal. Table 4 nevertheless reports ||X||_ort = 0.00 for LieNN+RNN at m=n=500. This is inconsistent with the stated 9-term Taylor implementation unless the values are rounded below the reported precision or a different exponentiation routine (e.g., exact exp, Padé, or Cayley) was used. The manuscript must state the actual routine. If the 0.00 is a rounding artifact, the 'exact orthogonality' guarantee is vacuous at the precision re
  2. [Table 4, ULM row (LieNN+RNN)] The ULM results for the best Lie-constrained model show ||D||_rel = 0.97 and ||Y||_ort = 20.80. Since n=500, ||I||_F = sqrt(500) ≈ 22.36, so ||Y Y^T − I||_F ≈ 465; this means the predicted right factor is far from orthogonal. A relative singular-value error of 0.97 means the recovered spectrum is essentially uncorrelated with the ground truth. These numbers are not consistent with the statement in Section 4.6 that LieNN+RNN 'closely approximates the GT while preserving orthogonality and low relative errors across factors.' The high SSIM (0.89) may reflect reconstruction-level fidelity, but the factorization itself is not accurate. The paper's contribution includes 'decomposition accuracy' as a stated goal; this table undermines that claim and should be addressed explicitly, either by reporting the actual factor errors, by discussing why downstream reconstruction tolerates them, or by tem
  3. [Lemma 1 / Eqs. (5)–(6) / Supp. A.4] Lemma 1 shows that for any A there exists an SO(m) left factor in some SVD-like decomposition. This is a parameterization fact, not a guarantee that the learned X is the true left singular-vector matrix. Eq. (5) computes diag(X^T A A^T X); this equals the squared singular values only when X equals the true left singular matrix. For an approximate X, this quantity is not the spectrum of A, and the Gershgorin bound in Supp. A.4.1 does not bound the error introduced by a learned X. The paper's phrase 'provably sufficient for valid factorisation' is therefore overstated: what is proven is the existence of a parameterization, not that the network's output yields a valid SVD. The authors should clarify this distinction and either provide a bound on the actual approximation error or remove the implication that the learned decomposition inherits the exactness of the Lie parameterization.
  4. [Abstract / Section 4.6] The abstract claims 'accelerates computation by 3–20× relative to cuSOLVER and randomised SVD.' The ULM table shows a ~23× speedup over cuSOLVER, which supports the upper end, but the MMP table (Table 2) shows LieNN+RNN at 14.19 ms vs. GT at 36.98 ms, a speedup of about 2.6×, below the stated lower bound of 3×. The text in Section 4.6 also says the method 'halves SVD runtime for MMP,' which is inconsistent with the reported numbers (the ratio is closer to one-third). Please reconcile the speedup claims with the tables, or state that the 3–20× range refers only to selected configurations.
minor comments (7)
  1. [Section 6] Typos: 'clincial' should be 'clinical', 'proposd' should be 'proposed'.
  2. [Section 4.5] The source code link is given as '*placeholder*'. For reproducibility, a real URL or a statement about availability upon acceptance is needed.
  3. [Table 2 / Figure 2] The caption mentions 'shaded second-best' but no shading is visible in the provided table; please clarify or remove.
  4. [Section 3.1, Eq. (4)] The Rodrigues formula is written as expr(S) but should be exp_r(S) or a similarly defined notation; the current expression is easy to confuse with a typo.
  5. [Table 4] The rSVD row reports NaN for ||D||_rel and ||Y||_ort. The caption says 'NaN for numerical instability,' but this should be explained in the text so readers know whether these are missing values or actual infinities.
  6. [Section 4.6 / Figure 4 caption] The caption of Figure 4 repeats the 'plots use logarithmic axes' note already present in the main text. This duplication is unnecessary.
  7. [Supp. D.1] The supplementary states that the error growth 'likely stems from the approximation of the matrix exponential.' This is an important limitation and should be mentioned in the main paper's discussion, as it directly qualifies the 'guaranteed' orthogonality claim.

Circularity Check

0 steps flagged

No significant circularity; the central derivation is self-contained and empirically anchored to external baselines.

full rationale

Quasi-SVD's derivation chain is not circular. The analytic construction X=exp(S), D=diag(sqrt(diag(X^T A A^T X))), Y=A^T X D^{-1} is an algebraic identity: for an exactly orthogonal X, X D Y^T = X X^T A = A, so D and Y are residuals of the chosen X rather than independently predicted quantities; their fidelity is then honestly evaluated against cuSOLVER ground truth. The learned variants are trained with supervised losses (L_sigma, L_end) against cuSOLVER outputs on training splits and tested on held-out sequences and a separate MUC instrument, so the reported SSIM, domain-transfer, and speed results are not fit-to-the-same-data predictions. Lemma 1 is a genuine existence proof (a sign flip yields det=1, hence SO(m)), not a uniqueness theorem imported from the authors. Self-citations (Hahne et al.) supply datasets, augmentation, and clinical context but are not load-bearing for the factorization mathematics. The one substantive weakness—Section 3.6 states that a 9-term Taylor expansion is the default for general matrices, and a truncated Taylor series of a skew-symmetric matrix is not exactly orthogonal, conflicting with the abstract's 'exact orthogonality' guarantee—is an internal-consistency/correctness issue, not a circularity: it does not make the claimed result equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The framework introduces no new physical entities. Its free parameters are the learned network weights, the unreported loss weighting constants, and the hand-chosen Taylor order. The most consequential assumptions are that cuSOLVER outputs are trustworthy ground truth and that models trained on one instrument or acquisition split transfer to the deployment setting.

free parameters (3)
  • Neural network weights θ (MLP and 3-step RNN) = per-dataset checkpoints, not shipped
    All learned variants (UNN, LieNN, E2E) depend on weights fitted to cuSOLVER outputs; no checkpoints or code are provided.
  • Loss weighting constants (λ1, λ2, λ3) = unspecified
    Eq. (14) defines the total training loss with three regularization constants, but their values are never reported, leaving the training objective underspecified.
  • Taylor expansion order k=9 (default) = 9
    Section 3.6 defaults to a 9-term Taylor expansion for general matrices; this hand-chosen truncation makes the 'exact orthogonality' guarantee approximate for the 500×500 ULM case.
axioms (6)
  • standard math Matrix exponential of a skew-symmetric matrix is orthogonal (exp(S)^T exp(S) = I).
    Proved in Supp. A.1 and invoked in Section 3.1 to justify X = exp(S) ∈ SO(m).
  • standard math Any SVD can be rewritten with its left factor in SO(m) by flipping one column sign in U and V.
    Lemma 1 / Supp. A.3; this justifies the asymmetric one-sided orthogonality design.
  • standard math Gershgorin circle theorem and Cauchy–Schwarz bounds on matrix coherence.
    Supp. A.4 uses these to justify the singular-value estimate, though the bounds do not control the learned network's approximation error.
  • domain assumption cuSOLVER SVD is treated as ground truth for training labels and evaluation.
    Section 4.4 designates cuSOLVER as GT; all learned losses and error metrics compare against it.
  • domain assumption Models trained on NPP and on the first 20 ULM sequences generalise to the MUC instrument and the last 5 ULM sequences.
    Sections 4.6 and 5 claim robust domain transfer; this generalisation across instruments and acquisition splits is load-bearing for the clinical claim.
  • ad hoc to paper The 9-term Taylor truncation behaves as an orthogonal matrix for practical purposes.
    Section 3.6 uses a 9-term truncation by default, but a truncated matrix exponential is not exactly orthogonal; reported X orthogonality errors of 0.00 are rounded, not exact.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Quasi-SVD: Learning a Lie-constrained matrix factorisation for real-time imaging." pith.science (2026). https://pith.science/paper/I2PFX233

@misc{pith2026260725967,
  author       = {Pith},
  title        = {Pith review of: Quasi-SVD: Learning a Lie-constrained matrix factorisation for real-time imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I2PFX233}},
  note         = {Machine review of arXiv:2607.25967}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Singular Value Decomposition (SVD) underlies matrix factorisation tasks across computational imaging, with medical applications increasingly demanding real-time processing. Yet SVD algorithms are inherently sequential, constraining real-time GPU throughput and limit online deployment in clinical pipelines. This study introduces Quasi-SVD, a differentiable, fully parallelized matrix factorization framework for GPUs. Rather than enforcing orthogonality on both factors, it guarantees exact orthogonality for a single Lie-parameterized factor while recovering the remaining components through soft constraints, enabling efficient parallel decomposition without iterative singular-vector optimization. This asymmetric design, provably sufficient for valid factorisation, achieves reconstruction fidelity of SSIM = 0.89-0.94 and accelerates computation by 3-20x relative to cuSOLVER and randomised SVD, enabling throughput above 25 FPS. Performance is evaluated on two medical imaging tasks spanning complementary computational regimes: (1) spatio-temporal background subtraction for ultrasound localisation microscopy, requiring high-dimensional matrix separation, and (2) Mueller matrix polarimetry for neurosurgical tissue characterisation, requiring massive batch processing of small matrices. Across both regimes and multiple imaging instruments, the proposed framework demonstrates robust domain transfer and throughput exceeding 25 FPS at clinical matrix scales, a rate sufficient for live image-guided workflows that classical solvers cannot currently support in these settings. By prioritising downstream reconstruction fidelity over exact spectral recovery, Quasi-SVD makes structured matrix factorisation practical for real-time imaging.

Figures

Figures reproduced from arXiv: 2607.25967 by Christopher Hahne.

Figure 1
Figure 1. Figure 1: Quasi-SVD model variants. The analytic model in (a) provides a closed-form solution using the matrix exponential for orthonormal projection without learnable components. The unconstrained neural network (UNN) in (b) introduces trainable weights and algebraic operations to factorise the input matrix. The Lie-based neural network (LieNN) in (c) predicts skew-symmetric parameters mapped to an exact orthogonal… view at source ↗
Figure 2
Figure 2. Figure 2: MMP benchmark comparison. The images show the per-pixel azimuth φi decomposed by the various models. (a) contains azimuth frames (388 × 516 pixels) from the NPP test set (remaining test samples are found in Supp. 6). (b) shows the azimuth results for the mouse uterine cervix image (600 × 700 pixels) Novikova et al. [2024], Pogudin [2024] for which the same models were used to investigate a model’s ability … view at source ↗
Figure 3
Figure 3. Figure 3: ULM clutter removal benchmark analysis. Each frame shows the B-mode intensity after spatio-temporal filtering. Rows depict results at different points in acquisition time with columns representing methods for comparison. Numbers at the bottom represent the average computation time for a m = n = 500 input matrix A. Images are colour-encoded for better visibility. performance. Taken together, the LieNN offer… view at source ↗
Figure 4
Figure 4. Figure 4: Computation time plots across matrix and batch sizes. Timings were measured on an Nvidia H100 GPU to accommodate large memory demands and averaged over 1000 iterations after 100 preceding warm-ups each. Runtime increases sharply with data size, amplifying the gap between the classical SVD and the proposed methods. The plots use logarithmic axes (base-10 on y, base-2 on x), visually compressing exponential … view at source ↗
Figure 5
Figure 5. Figure 5: Error growth for varying dimension m in square matrices. The curves depict the metrics from Section 4.2 of the main paper, representing the orthogonality errors ∥X∥ ort F and ∥Y∥ ort F as well as deviations for singular values in ∥D∥ rel F and a matrix reconstruction error ∥XDY⊤∥ rel F after decomposition. with diagonal vectors dX and dY from Eq. (5) in the main paper for X and Y, respectively. The intuiti… view at source ↗
Figure 6
Figure 6. Figure 6: MMP benchmark comparison (pt. 2). The images show the remaining NPP test samples of the per-pixel azimuth φi decomposed by the various models [PITH_FULL_IMAGE:figures/full_fig_p022_6.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

138 extracted references · 6 canonical work pages

  1. [1]

    Handbook for Automatic Computation: Linear Algebra , pages=

    Singular value decomposition and least squares solutions , author=. Handbook for Automatic Computation: Linear Algebra , pages=. 1971 , publisher=

    1971

  2. [3]

    2013 , publisher=

    Matrix computations , author=. 2013 , publisher=

    2013

  3. [4]

    SIAM journal on matrix analysis and applications , volume=

    Jacobi’s method is more accurate than QR , author=. SIAM journal on matrix analysis and applications , volume=. 1992 , publisher=

    1992

  4. [5]

    SIAM journal on scientific computing , volume=

    A hierarchically blocked Jacobi SVD algorithm for single and multiple graphics processing units , author=. SIAM journal on scientific computing , volume=. 2015 , publisher=

    2015

  5. [6]

    2026 , url =

    cuSOLVER: Dense Linear Algebra on GPUs , version =. 2026 , url =

    2026

  6. [7]

    ACM Transactions on Knowledge Discovery from Data (TKDD) , volume=

    Fast algorithms for approximating the singular value decomposition , author=. ACM Transactions on Knowledge Discovery from Data (TKDD) , volume=. 2011 , publisher=

    2011

  7. [8]

    Proceedings of the 2003 SIAM international conference on data mining , pages=

    Fast online svd revisions for lightweight recommender systems , author=. Proceedings of the 2003 SIAM international conference on data mining , pages=. 2003 , organization=

    2003

  8. [10]

    BIT Numerical Mathematics , volume=

    The truncated SVD as a method for regularization , author=. BIT Numerical Mathematics , volume=. 1987 , publisher=

    1987

  9. [11]

    SIAM Journal on Scientific and Statistical Computing , volume=

    Truncated singular value decomposition solutions to discrete ill-posed problems with ill-determined numerical rank , author=. SIAM Journal on Scientific and Statistical Computing , volume=. 1990 , publisher=

    1990

  10. [13]

    SIAM review , volume=

    Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions , author=. SIAM review , volume=. 2011 , publisher=

    2011

  11. [14]

    Linear Algebra and its Applications , volume=

    Rank revealing QR factorizations , author=. Linear Algebra and its Applications , volume=. 1987 , publisher=

    1987

  12. [15]

    SIAM Journal on Scientific Computing , volume=

    Efficient algorithms for computing a strong rank-revealing QR factorization , author=. SIAM Journal on Scientific Computing , volume=. 1996 , publisher=

    1996

  13. [16]

    SIAM Journal on Computing , volume=

    Fast Monte Carlo Algorithms for Matrices II: Computing a Low-Rank Approximation to a Matrix , author=. SIAM Journal on Computing , volume=. 2006 , publisher=

    2006

  14. [17]

    The random

    Musco, Cameron and Musco, Christopher , journal=. The random. 2017 , publisher=

    2017

  15. [18]

    JOSA A , volume=

    Interpretation of Mueller matrices based on polar decomposition , author=. JOSA A , volume=. 1996 , publisher=

    1996

  16. [19]

    Introduction and application , author=

    Principal component analysis: a beginner's guide—I. Introduction and application , author=. Weather , volume=. 1990 , publisher=

    1990

  17. [20]

    2009 IEEE international symposium on parallel & distributed processing , pages=

    Singular value decomposition on GPU using CUDA , author=. 2009 IEEE international symposium on parallel & distributed processing , pages=. 2009 , organization=

    2009

  18. [21]

    IEEE Transactions on Information Theory , volume=

    Tensor SVD: Statistical and computational limits , author=. IEEE Transactions on Information Theory , volume=. 2018 , publisher=

    2018

  19. [22]

    The flexible tensor singular value decomposition and its applications in multisensor signal fusion processing , journal =

    Jinfeng Huang and Feibin Zhang and Babak Safaei and Zhaoye Qin and Fulei Chu , keywords =. The flexible tensor singular value decomposition and its applications in multisensor signal fusion processing , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.ymssp.2024.111662 , url =

    arXiv 2024

  20. [23]

    Tensor SVD: Statistical and Computational Limits , year=

    Zhang, Anru and Xia, Dong , journal=. Tensor SVD: Statistical and Computational Limits , year=

  21. [24]

    International Conference on Machine Learning , pages=

    A stochastic PCA and SVD algorithm with an exponential convergence rate , author=. International Conference on Machine Learning , pages=. 2015 , organization=

    2015

  22. [25]

    International Conference on Machine Learning , pages=

    Fast stochastic algorithms for SVD and PCA: Convergence properties and convexity , author=. International Conference on Machine Learning , pages=. 2016 , organization=

    2016

  23. [26]

    Advances in neural information processing systems , volume=

    Randomized block krylov methods for stronger and faster approximate singular value decomposition , author=. Advances in neural information processing systems , volume=

  24. [28]

    LazySVD: Even Faster SVD Decomposition Yet Without Agonizing Pain , url =

    Allen-Zhu, Zeyuan and Li, Yuanzhi , booktitle =. LazySVD: Even Faster SVD Decomposition Yet Without Agonizing Pain , url =

  25. [29]

    IEEE Transactions on signal processing , volume=

    K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation , author=. IEEE Transactions on signal processing , volume=. 2006 , publisher=

    2006

  26. [30]

    2022 IEEE International Conference on Data Mining Workshops (ICDMW) , pages=

    SV-Learn: Learning Matrix Singular Values with Neural Networks , author=. 2022 IEEE International Conference on Data Mining Workshops (ICDMW) , pages=. 2022 , organization=

    2022

  27. [32]

    7th International Conference on Learning Representations , year=

    Spectral Inference Networks: Unifying Deep and Spectral Learning , author=. 7th International Conference on Learning Representations , year=

  28. [33]

    2022 , editor =

    Deng, Zhijie and Shi, Jiaxin and Zhu, Jun , booktitle =. 2022 , editor =

    2022

  29. [34]

    Proceedings of the 31st ACM international conference on information & knowledge management , pages=

    SVD-GCN: A simplified graph convolution paradigm for recommendation , author=. Proceedings of the 31st ACM international conference on information & knowledge management , pages=

  30. [35]

    ACM Transactions on Intelligent Systems and Technology , volume=

    Privacy-preserving point-of-interest recommendation based on simplified graph convolutional network for geological traveling , author=. ACM Transactions on Intelligent Systems and Technology , volume=. 2024 , publisher=

    2024

  31. [36]

    PRICAI 2014: Trends in Artificial Intelligence: 13th Pacific Rim International Conference on Artificial Intelligence, Gold Coast, QLD, Australia, December 1-5, 2014

    Fast learning of deep neural networks via singular value decomposition , author=. PRICAI 2014: Trends in Artificial Intelligence: 13th Pacific Rim International Conference on Artificial Intelligence, Gold Coast, QLD, Australia, December 1-5, 2014. Proceedings 13 , pages=. 2014 , organization=

    2014

  32. [37]

    Singular Value Decomposition and Neural Networks

    Bermeitinger, Bernhard and Hrycej, Tomas and Handschuh, Siegfried. Singular Value Decomposition and Neural Networks. Artificial Neural Networks and Machine Learning -- ICANN 2019: Deep Learning. 2019

    2019

  33. [38]

    2022 , isbn =

    Wu, Zhebin and Shu, Lin and Xu, Ziyue and Chang, Yaomin and Chen, Chuan and Zheng, Zibin , title =. 2022 , isbn =. doi:10.1145/3534678.3539436 , booktitle =

    arXiv 2022

  34. [39]

    Spatiotemporal Clutter Filtering of Ultrafast Ultrasound Data Highly Increases Doppler and fUltrasound Sensitivity , year=

    Demené, Charlie and Deffieux, Thomas and Pernot, Mathieu and Osmanski, Bruno-Félix and Biran, Valérie and Gennisson, Jean-Luc and Sieu, Lim-Anna and Bergel, Antoine and Franqui, Stéphanie and Correas, Jean-Michel and Cohen, Ivan and Baud, Olivier and Tanter, Mickael , journal=. Spatiotemporal Clutter Filtering of Ultrafast Ultrasound Data Highly Increases...

  35. [40]

    IEEE transactions on medical imaging , volume=

    Adaptive spatiotemporal SVD clutter filtering for ultrafast Doppler imaging using similarity of spatial singular vectors , author=. IEEE transactions on medical imaging , volume=. 2018 , publisher=

    2018

  36. [41]

    Nature Biomedical Engineering , volume=

    Performance benchmarking of microbubble-localization algorithms for ultrasound localization microscopy , author=. Nature Biomedical Engineering , volume=. 2022 , publisher=

    2022

  37. [42]

    doi:10.5281/zenodo.4343435 , url =

    Chavignon Arthur and Baptiste Heiles and Hingot Vincent and Lopez Pauline and Eliott Teston and Couture Olivier , title =. doi:10.5281/zenodo.4343435 , url =

    work page doi:10.5281/zenodo.4343435

  38. [43]

    2020 , note =

    Chavignon, Arthur and Heiles, Baptiste and Hingot, Vincent and Lopez, Pauline and Teston, Eliott and Couture, Olivier , title =. 2020 , note =

    2020

  39. [44]

    Nature , volume=

    Ultrafast ultrasound localization microscopy for deep super-resolution vascular imaging , author=. Nature , volume=. 2015 , publisher=

    2015

  40. [45]

    RF-ULM: Ultrasound Localization Microscopy Learned From Radio-Frequency Wavefronts , year=

    Hahne, Christopher and Chabouh, Georges and Chavignon, Arthur and Couture, Olivier and Sznitman, Raphael , journal=. RF-ULM: Ultrasound Localization Microscopy Learned From Radio-Frequency Wavefronts , year=

  41. [46]

    Magnetic resonance in medicine , volume=

    Low-rank plus sparse matrix decomposition for accelerated dynamic MRI with separation of background and dynamic components , author=. Magnetic resonance in medicine , volume=. 2015 , publisher=

    2015

  42. [47]

    Sampaio, Paulo and Lopez-Antu. M. Sci. Rep. , volume=. 2023 , publisher=

    2023

  43. [48]

    International journal of computer assisted radiology and surgery , pages=

    Near-real-time Mueller polarimetric image processing for neurosurgical intervention , author=. International journal of computer assisted radiology and surgery , pages=. 2024 , publisher=

    2024

  44. [49]

    Polarimetric feature analysis of Mueller matrices for brain tumor image segmentation , volume =

    Christopher Hahne and Ivan Diaz and Omar Rodriguez-Nunez and Elea Gros and Muriel Blatter and Theotim Lucas and David Hasler and Tatiana Novikova and Theoni Maragkou and Philippe Schucht and Richard McKinley , journal =. Polarimetric feature analysis of Mueller matrices for brain tumor image segmentation , volume =. 2025 , url =

    2025

  45. [50]

    Physically Consistent Image Augmentation for Deep Learning in Mueller Matrix Polarimetry , year=

    Hahne, Christopher and Rodriguez-Nunez, Omar and Gros, Elea and Lucas, Théotim and Hewer, Ekkehard and Novikova, Tatiana and Maragkou, Theoni and Schucht, Philippe and McKinley, Richard , journal=. Physically Consistent Image Augmentation for Deep Learning in Mueller Matrix Polarimetry , year=

  46. [51]

    IEEE Photonics Journal , year=

    Efficient Polarization Demosaicking Via Low-Cost Edge-Aware and Inter-Channel Correlation , author=. IEEE Photonics Journal , year=

  47. [52]

    Artificial Intelligence Review , volume=

    Incremental methods in face recognition: a survey , author=. Artificial Intelligence Review , volume=. 2021 , publisher=

    2021

  48. [53]

    IEEE Transactions on Image Processing , volume=

    Deep k-svd denoising , author=. IEEE Transactions on Image Processing , volume=. 2021 , publisher=

    2021

  49. [54]

    Evaluation of the robustness of learned MR image reconstruction to systematic deviations between training and test data for the models from the fastMRI challenge , author=. Machine Learning for Medical Image Reconstruction: 4th International Workshop, MLMIR 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 4 , pa...

    2021

  50. [55]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Learned compression of high dimensional image datasets , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  51. [56]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    FLAR-SVD: Fast and Latency-Aware Singular Value Decomposition for Model Compression , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  52. [57]

    Robotics: science and systems , volume=

    ElasticFusion: Dense SLAM without a pose graph , author=. Robotics: science and systems , volume=

  53. [58]

    2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Nerf-slam: Real-time dense monocular slam with neural radiance fields , author=. 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2023 , organization=

    2023

  54. [59]

    International journal of computer assisted radiology and surgery , volume=

    Learning how to robustly estimate camera pose in endoscopic videos , author=. International journal of computer assisted radiology and surgery , volume=. 2023 , publisher=

    2023

  55. [60]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Gaussian splatting slam , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  56. [61]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Tangent Space Backpropagation for 3D Transformation Groups , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  57. [62]

    ONNX: Open Neural Network Exchange Format for Machine Learning Models , year =

  58. [63]

    2023 , url =

    NVIDIA TensorRT: Programmable Inference Accelerator , author =. 2023 , url =

    2023

  59. [64]

    SIAM review , volume=

    Nineteen dubious ways to compute the exponential of a matrix, twenty-five years later , author=. SIAM review , volume=. 2003 , publisher=

    2003

  60. [65]

    SIAM Review , volume=

    Orthogonal matrices and the matrix exponential , author=. SIAM Review , volume=. 1990 , publisher=

    1990

  61. [66]

    Dai , keywords =

    Jian S. Dai , keywords =. Euler–Rodrigues formula variations, quaternion conjugation and intrinsic connections , journal =. 2015 , issn =. doi:https://doi.org/10.1016/j.mechmachtheory.2015.03.004 , url =

    work page doi:10.1016/j.mechmachtheory.2015.03.004 2015

  62. [68]

    Scientific Data , volume=

    FastMRI Prostate: A public, biparametric MRI dataset to advance machine learning for prostate cancer imaging , author=. Scientific Data , volume=. 2024 , publisher=

    2024

  63. [69]

    Radiology: Artificial Intelligence , volume=

    fastMRI Breast: A publicly available radial k-space dataset of breast dynamic contrast-enhanced MRI , author=. Radiology: Artificial Intelligence , volume=. 2025 , publisher=

    2025

  64. [70]

    Acm Sigkdd Explorations Newsletter , volume=

    Lessons from the Netflix prize challenge , author=. Acm Sigkdd Explorations Newsletter , volume=. 2007 , publisher=

    2007

  65. [71]

    Proceedings of KDD Cup 2011 , pages=

    The yahoo! music dataset and kdd-cup’11 , author=. Proceedings of KDD Cup 2011 , pages=. 2012 , organization=

    2011

  66. [72]

    http://yann

    The MNIST database of handwritten digits , author=. http://yann. lecun. com/exdb/mnist/ , year=

  67. [73]

    Nucleic acids research , volume=

    The international Genome sample resource (IGSR): A worldwide collection of genome variation incorporating the 1000 Genomes Project data , author=. Nucleic acids research , volume=. 2017 , publisher=

    2017

  68. [74]

    2010 IEEE International conference on data mining , pages=

    Factorization machines , author=. 2010 IEEE International conference on data mining , pages=. 2010 , organization=

    2010

  69. [75]

    Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval , pages=

    Neural factorization machines for sparse predictive analytics , author=. Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval , pages=

  70. [76]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  71. [77]

    Advances in Applied Clifford Algebras , volume=

    On SVD and Polar Decomposition in Real and Complexified Clifford Algebras , author=. Advances in Applied Clifford Algebras , volume=. 2024 , publisher=

    2024

  72. [78]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Cheap Orthogonal Constraints in Neural Networks: A Simple Parametrization of the Orthogonal and Unitary Group , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , editor =

    2019

  73. [79]

    International Conference on Machine Learning , pages=

    Cheap orthogonal constraints in neural networks: A simple parametrization of the orthogonal and unitary group , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  74. [81]

    2023 , urldate =

    NVIDIA Developer Forum Discussion , title =. 2023 , urldate =

    2023

  75. [82]

    2021 , urldate =

    Issue: Fix the non-converging issue of SVD on GPU for large matrices , howpublished =. 2021 , urldate =

    2021

  76. [84]

    2024 , urldate =

    Pogudin, Gleb , title =. 2024 , urldate =

    2024

  77. [87]

    Communications of the ACM , volume=

    Programming parallel algorithms , author=. Communications of the ACM , volume=. 1996 , publisher=

    1996

  78. [88]

    Journal of the ACM (JACM) , volume=

    The parallel evaluation of general arithmetic expressions , author=. Journal of the ACM (JACM) , volume=. 1974 , publisher=

    1974

  79. [89]

    Mathematics , volume=

    Computing the matrix exponential with an optimized Taylor polynomial approximation , author=. Mathematics , volume=. 2019 , publisher=

    2019

  80. [90]

    SIAM Journal on Scientific Computing , volume =

    Low-Rank Matrix Approximation Using the Lanczos Bidiagonalization Process , author =. SIAM Journal on Scientific Computing , volume =. 2000 , publisher =

    2000

Showing first 80 references.

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