Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Diffeomorphic Temporal Alignment Nets for Time-series Joint Alignment and Averaging

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single regularization-free inverse-consistency loss lets a neural network learn to jointly align and average misaligned time series, and that one fixed configuration outperforms tuned DTW-based and learned…

desk verdict Solid extension of the authors' own DTAN work, but the NCC 'superiority' claim overreaches because the triplet loss is a surrogate for the metric. read the letter →

arxiv 2502.06591 v1 pith:JCCNWZQO submitted 2025-02-10 cs.LG

Ron Shapira Weber , Oren Freifeld This is my paper · ORCID
classification cs.LG
keywords time-seriesjointalignmentdiffeomorphictemporalnetCPABdiffeomorphismsinverseconsistencyaveragingerrornearestcentroidclassificationUCRarchivevariable-lengthtimeseries
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that time-series joint alignment and averaging should be treated as a learning problem, and that the right learned objective removes the need for hand-tuned warp regularization. Its DTAN predicts a smooth, invertible time warp (a diffeomorphism) for each input signal, and a new loss called the Inverse Consistency Averaging Error (ICAE) checks that each class average, warped backward, still matches the original signals. The paper reports that this regularization-free scheme, with one fixed hyperparameter configuration, beats DTW-based barycenter averaging and earlier trained methods on nearest-centroid classification accuracy on the 84-dataset comparison set, and maintains top median accuracy across all 128 UCR datasets, while also aligning new signals in one forward pass. If it is right, a practical bottleneck in time-series analysis—computing meaningful averages and centroids from misaligned recordings—becomes much cheaper and does not require per-dataset tuning.

What carries the argument

The machinery is a temporal transformer network built on CPAB diffeomorphisms: warps obtained by integrating continuous piecewise-affine velocity fields, so they are smooth, invertible, and admit cheap closed-form gradients in 1D. The localization network predicts warp parameters from each input, and a differentiable resampler produces the aligned signal. The load-bearing new object is the ICAE loss (Equation 18), which computes each class centroid from the forward-warped signals and then warps that centroid back to each original signal with the inverse warp; the centroid-based triplet loss (Equation 20) adds inter-class separation. This backward-consistency check is what removes the need for a regularization term and, via masks over valid entries, lets the same loss handle variable-length signals.

What would settle it

Train the LICAE-only network on a synthetic ensemble with known ground-truth latent signals and warps; if it reaches near-zero ICAE while recovering warps far from the ground truth (or visibly pinched averages), the key assumption fails. On the empirical side, re-running the released code with the stated fixed configuration on the 84-dataset comparison set and finding median nearest-centroid accuracy below DBA's 0.657 would falsify the superiority claim.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that inverse consistency between the class average and the original signals is enough to prevent the degenerate 'pinching' solutions that plague unsupervised alignment. The network first warps every signal by an input-dependent CPAB diffeomorphism and averages the warped signals per class; ICAE then warps each class average back through the inverse warp and measures the squared error against the original signal. Minimizing that round-trip error forces the average to remain faithful to its class, and the paper argues this strongly discourages trivial or unrealistic warps without any smoothness regularizer. The experimental payload is that the LICAE-triplet variant—ICAE plus a centroid-margin term—reaches a median nearest-centroid accuracy of 0.707 on the 84-dataset comparison set and 0.670 over all 128 datasets with one fixed configuration, compared with 0.657 for DBA and 0.604 for the regularized single-configuration DTAN on the same comparison set. The same trained network aligns previously unseen test signals at inference time and handles variable-length inputs.

Load-bearing premise

The load-bearing premise is that satisfying the inverse-consistency loss cannot be done by a distorted, 'pinching' warp that maps the deformed average back onto the original signals; if some datasets admit such a solution, the regularization-free advantage could disappear.

Editorial extensions

If this is right

  • A single trained DTAN can align and average unseen test signals from any class, without knowing their labels, in one fast forward pass.
  • The ICAE formulation supports joint alignment of variable-length signals directly, averaging only over valid time points and avoiding specialized losses or boundary constraints.
  • No warp-smoothness regularization is needed, so practitioners can avoid dataset-specific tuning of the two prior regularization hyperparameters; the paper reports one fixed configuration across all 128 UCR datasets.
  • Adding a cross-entropy classification head to the same network improves both alignment-driven nearest-centroid accuracy and class separation in the learned embedding.
  • Applying PCA to the aligned signals requires fewer principal components to explain most variance, and reconstructing the original signals from few components acts as denoising.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The inverse-consistency principle is a generic mechanism: the same 'average forward, check backward' loss could replace smoothness priors in image congealing or functional registration, though the paper tests only 1D signals.
  • The UCR benchmark scores centroids by classification, so it rewards class separation more than alignment fidelity; a synthetic benchmark with known ground-truth warps would directly test whether ICAE recovers true latent alignments.
  • Because the alignment is amortized rather than recomputed, new data drawn from a distribution unlike the training set may not align well; a meta-learning or fine-tuning extension would address that gap, which the paper does not explore.
  • The fast forward-pass alignment could serve as an online preprocessing layer for streaming time series, but the paper only demonstrates batch inference and does not analyze distribution shift over time.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper extends the authors' earlier Diffeomorphic Temporal Alignment Net (DTAN) framework for joint alignment and averaging of time-series ensembles. The new content includes the inverse-consistency averaging error (ICAE) as a regularization-free objective, an inverse-consistent centroid triplet loss (LICAE-triplet), recurrent DTAN (RDTAN), a multi-task classification extension (MT-DTAN), an evaluation of several deep backbones for the localization network, and a PCA-based case study. The main empirical claim is that a single fixed hyperparameter configuration of DTAN with LICAE-triplet achieves state-of-the-art / superior NCC accuracy across the 128-dataset UCR archive.

Significance. If the claims are established, the framework has useful practical properties: it removes the need for warp regularization, supports variable-length signals, aligns new data by a single fast forward pass, and improves downstream PCA. The paper also ships code, reports controlled comparisons of three losses under a fixed configuration, and includes timing measurements that quantify the inference advantage over DTW-based methods. These are concrete strengths. However, the benchmark evidence for the central superiority claim is weaker than the abstract suggests, and the key theoretical claim that ICAE avoids degenerate warps is supported only by a single qualitative example.

major comments (4)
  1. [§5.2, Table 2 and Eq. (20)] The comparison is partially circular for LICAE-triplet. Equation (20) optimizes max(0, ∥u_a − μ_p∘T^{−θ_i}∥² − ∥u_a − μ_n∘T^{−θ_i}∥² + α), which is a margin version of the distance ∥u_test∘T^{θ} − μ_k∥ used by the NCC evaluation metric. The jump from LICAE (0.665) to LICAE-triplet (0.707) on the 84-dataset part of Table 2 is thus the expected result of optimizing a surrogate of the evaluation metric, and it does not by itself establish better joint alignment or averaging quality. I recommend adding an evaluation that is not aligned with the training objective, such as ground-truth warp recovery on synthetic data or a reconstruction-fidelity measure, or explicitly limiting the claim to NCC performance.
  2. [§5.2.1, Table 2] The abstract's claim of "superiority over contemporary averaging methods" is not supported by Table 2 as a whole. On the same 84 datasets, single-configuration LICAE-triplet reaches 0.707 median NCC, while SoftDBA-div reports 0.708, DTANlibcpab 0.705, ResNet-TW 0.711, and DTANDIFW 0.749. The comparison is not matched because the latter numbers use per-dataset test-set hyperparameter selection, but the blanket superiority claim should be replaced by a matched-protocol comparison or a carefully qualified statement. In addition, Part 4 of Table 2 reports only the two proposed losses on all 128 datasets and includes no DBA, SoftDTW, or WCSS baseline, so the full-archive superiority claim is not demonstrated.
  3. [§5.2.2, Table 2] The "single fixed HP configuration" is not configuration-free. The configuration was taken from [7] as the one with the highest number of wins on the 84 test sets, where [7] explicitly tuned hyperparameters on the test data. This makes the comparison among the three losses internally fair, but it does not support the claim that the method works without dataset-specific tuning, because the chosen configuration already encodes knowledge of these test sets. The authors should either select the configuration using only training/validation data or report results across several fixed configurations to show robustness.
  4. [§4.1, Eq. (18)] The key claim that Equation (18) "strongly discourages trivial solutions or unrealistic warps" is load-bearing for the regularization-free contribution, but it is supported only by one qualitative example (BeetleFly, Figure 8). No analysis of the loss landscape or conditions excluding degenerate, pinching solutions is given. I request either a formal statement with sufficient conditions under which ICAE excludes such solutions, or a systematic empirical check, e.g., measuring warp distortion against known ground-truth warps on synthetic data across many datasets.
minor comments (4)
  1. [Abstract and §1] There is a typo: "Tempiral Transform Net" should be "Temporal Transform Net."
  2. [Eq. (18) and Algorithm 1] Equation (18) writes LICAE with an unspecified ∥·∥ norm, while Algorithm 1 and Eq. (20) use the squared ℓ2 norm. For consistency, Eq. (18) should use ∥·∥²_{ℓ2}.
  3. [References] Reference [25] is missing an author name: "M. Cuturi and M. ," should be "M. Cuturi and M. Blondel."
  4. [Table 2] Since the paper reports 5 runs per condition, Table 2 would benefit from reporting an interquartile range or min-max spread in addition to median and best, so the reader can assess the stochastic stability of the three losses.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the new losses are defined from first principles and evaluated on external held-out data; remaining caveats are benchmark-interpretation or correctness concerns, not derivation circularity.

full rationale

The derivation chain is self-contained. The new objectives—LICAE (Eq. 18), LICAE-triplet (Eq. 20), and the multi-task loss (Eq. 23)—are defined directly from the warping model and class centroids, and the paper's central quantitative claims are evaluated on the external UCR archive with held-out test labels, so no reported number is forced by construction from a fitted parameter. The framework self-citations to [1] and [2] are disclosed as earlier partial versions and are not used as a uniqueness or correctness argument for the new losses. The authors also disclose that the single fixed hyperparameter configuration was inherited from [7], where it was selected using test data; that is a benchmark-fairness caveat, not circularity. The nearest-centroid/triplet overlap is the most substantive caveat: LICAE-triplet is a margin loss over centroid distances and therefore closely tracks the NCC evaluation metric, and this should temper claims that Table 2 independently proves superior alignment or averaging. However, the triplet term is a Jacobian-weighted inverse-warp analog rather than the identical NCC objective, and test labels are not used in training, so the improvement is not statistically forced by definition. The unsupported claim in Section 4.1 that 'Equation 18 strongly discourages trivial solutions' is a correctness risk (a zero-variance pinching solution can also drive LICAE to zero), but that is not a circular step. Overall, no load-bearing part of the derivation reduces to its own inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's core machinery (CPAB warps, DTAN, ICAE) is inherited from prior work; the paper contributes new losses and combinations. The free parameters are mostly fixed hyperparameters; none are fitted to the target benchmark in a per-dataset way, but the single configuration was selected using test performance in prior work [7]. The load-bearing assumption is that inverse consistency alone prevents trivial solutions.

free parameters (5)
  • N_p (number of CPAB subintervals) = 16
    Fixed for all experiments, selected based on [7] where this value was reported to yield the highest number of NCC wins (Section 5.2).
  • scaling-and-squaring parameter = 8
    Fixed for all experiments, from [7] (Section 5.2).
  • number of recurrences (RDTAN/stacked warps) = 4
    Fixed in the main experiments; the authors state 'we fixed the number of recurrences to 4' (Section 5.2).
  • triplet loss margin alpha = 1
    Set to 1 in all experiments and stated to be dataset-independent (Section 4.3).
  • multi-task weight lambda_ce = 1
    Set to 1 by default in the multi-task loss (Section 4.7).
assumptions (5)
  • domain assumption Generated signals follow u_i = v_i composed with w_i with latent warps w_i; the w_i are invertible (Eq. 1-2).
    This generative model frames all misalignment as temporal reparameterization; if real data contain amplitude or noise differences, the model absorbs them into the warp or residual.
  • standard math CPAB warps form a family of C1 diffeomorphisms closed under inversion, and support efficient closed-form gradients (from [14,15,7]).
    The method relies on these properties for differentiability and for the inverse-consistency computation in ICAE.
  • domain assumption The localization network learns input-dependent warps that generalize from training to test signals.
    The claimed benefit of fast generalization to new data assumes the network extrapolates to unseen inputs (Section 4.5).
  • ad hoc to paper ICAE has no spurious degenerate optima (e.g., pinching) in the datasets considered.
    The regularization-free claim depends on the inverse-consistency term preventing trivial solutions; the paper gives intuition and one illustrative example but no proof (Section 4.1, Figure 8).
  • domain assumption NCC accuracy is a valid proxy for joint-alignment quality.
    The benchmark evaluates NCC rather than direct alignment error since ground-truth warps are unavailable in UCR data (Section 5.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffeomorphic Temporal Alignment Nets for Time-series Joint Alignment and Averaging." pith.science (2026). https://pith.science/paper/JCCNWZQO

@misc{pith2026250206591,
  author       = {Pith},
  title        = {Pith review of: Diffeomorphic Temporal Alignment Nets for Time-series Joint Alignment and Averaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCCNWZQO}},
  note         = {Machine review of arXiv:2502.06591}
}
read the original abstract

In time-series analysis, nonlinear temporal misalignment remains a pivotal challenge that forestalls even simple averaging. Since its introduction, the Diffeomorphic Temporal Alignment Net (DTAN), which we first introduced (Weber et al., 2019) and further developed in (Weber & Freifeld, 2023), has proven itself as an effective solution for this problem (these conference papers are earlier partial versions of the current manuscript). DTAN predicts and applies diffeomorphic transformations in an input-dependent manner, thus facilitating the joint alignment (JA) and averaging of time-series ensembles in an unsupervised or a weakly-supervised manner. The inherent challenges of the weakly/unsupervised setting, particularly the risk of trivial solutions through excessive signal distortion, are mitigated using either one of two distinct strategies: 1) a regularization term for warps; 2) using the Inverse Consistency Averaging Error (ICAE). The latter is a novel, regularization-free approach which also facilitates the JA of variable-length signals. We also further extend our framework to incorporate multi-task learning (MT-DTAN), enabling simultaneous time-series alignment and classification. Additionally, we conduct a comprehensive evaluation of different backbone architectures, demonstrating their efficacy in time-series alignment tasks. Finally, we showcase the utility of our approach in enabling Principal Component Analysis (PCA) for misaligned time-series data. Extensive experiments across 128 UCR datasets validate the superiority of our approach over contemporary averaging methods, including both traditional and learning-based approaches, marking a significant advancement in the field of time-series analysis.

Figures

Figures reproduced from arXiv: 2502.06591 by the authors.

Figure 1
Figure 1. An illustration of the joint-alignment problem in ECG data. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The benefits of joint alignment for dimensionality reduction, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. DTAN joint alignment demonstrated on a class of the “Trace” dataset [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: The effect of the smoothness prior on the predicted warps in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: The effect of the regularization HP. The figures shows 10 samples (gray) from the ECGFiveDays dataset with their estimated [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The Inverse Consistency Averaging Error loss in a two-class [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Training procedure on the BeetleFly dataset. The first column depicts the input data (for better visualization, the top panel shows 3 random signals while the bottom 10 signals and their average are in blue). (Top) The Within-Class Sum of Squares (WCSS) loss reduces va…
Figure 9
Figure 9. Figure 9: JA of variable-length data (Dataset: ShakeGestureWiimoteZ) using the proposed LICAE. Shaded area is ± std. dev. any diffeomorphism that is representable by integrating a Lipshitz￾continuous stationary velocity field can be approximated by a CPAB diffeomorphism [14, 15]…
Figure 12
Figure 12. Figure 12: Joint alignment and averaging of the (top) [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: Recurrent DTAN (RDTAN) JA of synthetic data. (a) latent source and (b) 10 perturbed signals (gray) and their average (blue). [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 14
Figure 14. Figure 14: Nearest Centroid Classifier (NCC) performance as a [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]
Figure 15
Figure 15. Figure 15: DTAN vs. MT-DTAN comparison across different back [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TimePoint: Accelerated Time Series Alignment via Self-Supervised Keypoint and Descriptor Learning

    cs.CV 2025-05 conditional novelty 6.0 of 10

    TimePoint learns sparse keypoints and descriptors from synthetic 1D signals and applies DTW to these, yielding large speedups and modest accuracy gains over full-signal DTW on real-world benchmarks.

Reference graph

Works this paper leans on

75 extracted references · 70 canonical work pages · cited by 1 Pith paper

  1. [7]

    Closed-form diffeomorphic transformations for time series alignment,

    I. Martinez, E. Viles, and I. G. Olaizola, “Closed-form diffeomorphic transformations for time series alignment,” in International Conference on Machine Learning . PMLR, 2022, pp. 15 122–15 158

    work page 2022

  2. [1]

    Diffeomorphic temporal alignment nets,

    R. S. Weber, M. Eyal, N. Skafte Detlefsen, O. Shriki, and O. Freifeld, “Diffeomorphic temporal alignment nets,” in Advances in neural information processing systems, vol. 32, 2019. DIFFEOMORPHIC TEMPORAL ALIGNMENT NETWORKS 14

    work page 2019

  3. [2]

    Regularization-free diffeomor- phic temporal alignment nets,

    R. S. Weber and O. Freifeld, “Regularization-free diffeomor- phic temporal alignment nets,” in International Conference on Machine Learning. PMLR, 2023, pp. 30 794–30 826

    work page 2023

  4. [3]

    Dynamic-programming approach to continuous speech recognition,

    H. Sakoe, “Dynamic-programming approach to continuous speech recognition,” 1971 Proc. the International Congress of Acoustics, Budapest, 1971

    work page 1971

  5. [4]

    Mumford and A

    D. Mumford and A. Desolneux, Pattern theory: the stochastic analysis of real-world signals. AK Peters/CRC Press, 2010

    work page 2010

  6. [5]

    The ucr time series archive,

    H. A. Dau, A. Bagnall, K. Kamgar, C.-C. M. Yeh, Y. Zhu, S. Gharghabi, C. A. Ratanamahatana, and E. Keogh, “The ucr time series archive,” IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 6, pp. 1293–1305, 2019

    work page 2019

  7. [6]

    Residual Networks as Flows of Velocity Fields for Diffeomorphic Time Series Alignment

    H. Huang, B. B. Amor, X. Lin, F. Zhu, and Y. Fang, “Residual networks as flows of velocity fields for diffeomorphic time series alignment,” arXiv preprint arXiv:2106.11911, 2021

    work page Pith review arXiv 2021

  8. [8]

    Learning from one example through shared densities on transforms,

    E. G. Miller, N. E. Matsakis, and P. A. Viola, “Learning from one example through shared densities on transforms,” in CVPR, vol. 1. IEEE, 2000, pp. 464–471

    work page 2000

Show all 75 references
  1. [9]

    Data driven image models through continuous joint alignment,

    E. G. Learned-Miller, “Data driven image models through continuous joint alignment,” IEEE TPAMI, vol. 28, no. 2, pp. 236–250, 2006

    2006

  2. [10]

    Unsupervised joint alignment of complex images,

    G. B. Huang, V. Jain, and E. Learned-Miller, “Unsupervised joint alignment of complex images,” in ICCV. IEEE, 2007, pp. 1–8

    2007

  3. [11]

    Learning to align from scratch,

    G. Huang, M. Mattar, H. Lee, and E. G. Learned-Miller, “Learning to align from scratch,” in NIPS, 2012, pp. 764–772

    2012

  4. [12]

    Least squares congealing for unsupervised alignment of images,

    M. Cox, S. Sridharan, S. Lucey, and J. Cohn, “Least squares congealing for unsupervised alignment of images,” in CVPR, vol. 2008. NIH Public Access, 2008, p. 1

    2008

  5. [13]

    Least-squares congealing for large numbers of images,

    ——, “Least-squares congealing for large numbers of images,” in ICCV. IEEE, 2009, pp. 1949–1956

    2009

  6. [14]

    Highly-expressive spaces of well-behaved trans- formations: Keeping it simple,

    O. Freifeld, S. Hauberg, K. Batmanghelich, and J. W. Fisher III, “Highly-expressive spaces of well-behaved trans- formations: Keeping it simple,” in ICCV, 2015

    2015

  7. [15]

    Transformations based on continuous piecewise-affine velocity fields,

    ——, “Transformations based on continuous piecewise-affine velocity fields,” IEEE TPAMI, 2017

    2017

  8. [16]

    Finite-dimensional Lie algebras for fast diffeomorphic image registration,

    M. Zhang and P. T. Fletcher, “Finite-dimensional Lie algebras for fast diffeomorphic image registration,” in IPMI, 2015

    2015

  9. [17]

    Fast diffeomorphic image registration via fourier- approximated lie algebras,

    ——, “Fast diffeomorphic image registration via fourier- approximated lie algebras,” IJCV, 2018

    2018

  10. [18]

    Spatial transformer networks,

    M. Jaderberg, K. Simonyan, A. Zisserman et al., “Spatial transformer networks,” in Advances in neural information processing systems, 2015, pp. 2017–2025

    2015

  11. [19]

    Inverse compositional spatial trans- former networks,

    C.-H. Lin and S. Lucey, “Inverse compositional spatial trans- former networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 2568– 2576

    2017

  12. [20]

    Deep diffeomorphic transformer networks,

    N. Skafte Detlefsen, O. Freifeld, and S. Hauberg, “Deep diffeomorphic transformer networks,” in CVPR, 2018

    2018

  13. [21]

    The ucr time series classification archive,

    Y. Chen, E. Keogh, B. Hu, N. Begum, A. Bagnall, A. Mueen, and G. Batista, “The ucr time series classification archive,” July 2015

    2015

  14. [22]

    Dynamic programming algorithm optimization for spoken word recognition,

    H. Sakoe and S. Chiba, “Dynamic programming algorithm optimization for spoken word recognition,”IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 26, no. 1, pp. 43–49, 1978

    1978

  15. [23]

    A global averag- ing method for dynamic time warping, with applications to clustering,

    F. Petitjean, A. Ketterlin, and P. Ganc ¸arski, “A global averag- ing method for dynamic time warping, with applications to clustering,” Pattern Recognition, vol. 44, no. 3, pp. 678–693, 2011

    2011

  16. [24]

    Dynamic time warping averaging of time series allows faster and more accurate classification,

    F. Petitjean, G. Forestier, G. I. Webb, A. E. Nicholson, Y. Chen, and E. Keogh, “Dynamic time warping averaging of time series allows faster and more accurate classification,” in Data Mining (ICDM), 2014 IEEE International Conference on. IEEE, 2014, pp. 470–479

    2014

  17. [25]

    Soft-dtw: a differentiable loss function for time-series,

    M. Cuturi and M. , “Soft-dtw: a differentiable loss function for time-series,” arXiv preprint arXiv:1703.01541, 2017

    2017 arXiv

  18. [26]

    Differentiable divergences between time series,

    M. Blondel, A. Mensch, and J.-P. Vert, “Differentiable divergences between time series,” in International Conference on Artificial Intelligence and Statistics . PMLR, 2021, pp. 3853–3861

    2021

  19. [27]

    tslearn: a machine learning toolkit ded- icated to time-series data (2017),

    R. Tavenard, “tslearn: a machine learning toolkit ded- icated to time-series data (2017),” URL https://github. com/rtavenar/tslearn, 2017

    2017

  20. [28]

    Fast global alignment kernels,

    M. Cuturi, “Fast global alignment kernels,” in ICML, 2011

    2011

  21. [29]

    Time series alignment with global invariances,

    T. Vayer, L. Chapel, N. Courty, R. Flamary, Y. Soullard, and R. Tavenard, “Time series alignment with global invariances,” arXiv preprint arXiv:2002.03848, 2020

    2002 arXiv

  22. [30]

    Neural time warping for multiple sequence alignment,

    K. Kawano, T. Kutsuna, and S. Koide, “Neural time warping for multiple sequence alignment,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 3837–3841

    2020

  23. [31]

    Shape analysis of elastic curves in euclidean spaces,

    A. Srivastava, E. Klassen, S. H. Joshi, and I. H. Jermyn, “Shape analysis of elastic curves in euclidean spaces,” IEEE TPAMI, 2010

    2010

  24. [32]

    Registration of functional data using fisher-rao metric,

    A. Srivastava, W. Wu, S. Kurtek, E. Klassen, and J. S. Marron, “Registration of functional data using fisher-rao metric,” arXiv preprint arXiv:1103.3817, 2011

    2011 arXiv

  25. [33]

    An unsupervised learning model for deformable medical image registration,

    G. Balakrishnan, A. Zhao, M. R. Sabuncu, J. Guttag, and A. V. Dalca, “An unsupervised learning model for deformable medical image registration,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 9252–9260

    2018

  26. [34]

    Dreaming more data: Class-dependent distributions over diffeomorphisms for learned data augmentation,

    S. Hauberg, O. Freifeld, A. B. L. Larsen, J. W. F. III, and L. K. Hansen, “Dreaming more data: Class-dependent distributions over diffeomorphisms for learned data augmentation,” in AISTATS, 2016

    2016

  27. [35]

    Explicit disentangle- ment of appearance and perspective in generative models,

    N. Skafte Detlefsen and S. Hauberg, “Explicit disentangle- ment of appearance and perspective in generative models,” NeurIPS, 2019

    2019

  28. [36]

    Cyclic diffeomor- phic transformer nets for contour alignment,

    I. Kaufman, R. S. Weber, and O. Freifeld, “Cyclic diffeomor- phic transformer nets for contour alignment,” in 2021 IEEE International Conference on Image Processing (ICIP). IEEE, 2021, pp. 349–353

    2021

  29. [37]

    Single pair cross-modality super resolution,

    G. Shacht, D. Danon, S. Fogel, and D. Cohen-Or, “Single pair cross-modality super resolution,” in CVPR, 2021

    2021

  30. [38]

    Probabilistic spatial transformer networks,

    P. Schw¨obel, F. R. Warburg, M. Jørgensen, K. H. Madsen, and S. Hauberg, “Probabilistic spatial transformer networks,” in UAI, 2022

    2022

  31. [39]

    Leveraging statistical shape priors in gan- based ECG synthesis,

    N. Neifar, A. Ben-Hamadou, A. Mdhaffar, M. Jmaiel, and B. Freisleben, “Leveraging statistical shape priors in gan- based ECG synthesis,” arXiv preprint arXiv:2211.02626 , 2022

    2022 arXiv

  32. [40]

    Personalized monitoring in home healthcare: An assistive system for post hip replacement rehabilitation,

    A. Kryeem, S. Raz, D. Eluz, D. Itah, H. Hel-Or, and I. Shimshoni, “Personalized monitoring in home healthcare: An assistive system for post hip replacement rehabilitation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1868–1877. DIFFEOM...

    2023

  33. [41]

    Continuous piecewise-affine based motion model for image animation,

    H. Wang, F. Liu, Q. Zhou, R. Yi, X. Tan, and L. Ma, “Continuous piecewise-affine based motion model for image animation,” in AAAI, 2024

    2024

  34. [42]

    Action assessment in reha- bilitation: Leveraging machine learning and vision-based analysis,

    A. Kryeem, N. Boutboul, I. Bear, S. Raz, D. Eluz, D. Itah, H. Hel-Or, and I. Shimshoni, “Action assessment in reha- bilitation: Leveraging machine learning and vision-based analysis,” Computer Vision and Image Understanding , vol. 251, p. 104228, 2025

    2025

  35. [43]

    Trainable highly-expressive activation functions,

    I. Chelly, S. E. Finder, S. Ifergane, and O. Freifeld, “Trainable highly-expressive activation functions,” in European Confer- ence on Computer Vision. Springer, 2024, pp. 200–217

    2024

  36. [44]

    Digraf: Diffeomorphic graph-adaptive activation function,

    K. S. I. Mantri, X. Wang, C.-B. Sch ¨onlieb, B. Ribeiro, B. Bevilacqua, and M. Eliasof, “Digraf: Diffeomorphic graph-adaptive activation function,” in Advances in Neural Information Processing Systems , A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C...

    2024

  37. [45]

    Temporal transformer networks: Joint learning of invariant and discriminative time warping,

    S. Lohit, Q. Wang, and P. Turaga, “Temporal transformer networks: Joint learning of invariant and discriminative time warping,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 12 426– 12 435

    2019

  38. [46]

    Deep learning of warping functions for shape analysis,

    E. Nunez and S. H. Joshi, “Deep learning of warping functions for shape analysis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020, pp. 866–867

    2020

  39. [47]

    Srvfnet: A generative network for unsupervised multiple diffeomorphic functional alignment,

    E. Nunez, A. Lizarraga, and S. H. Joshi, “Srvfnet: A generative network for unsupervised multiple diffeomorphic functional alignment,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 4481–4489

    2021

  40. [48]

    Srvfregnet: Elastic function registration using deep neural networks,

    C. Chen and A. Srivastava, “Srvfregnet: Elastic function registration using deep neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4462–4471

    2021

  41. [49]

    libcpab,

    N. S. Detlefsen, “libcpab,” https://github.com/SkafteNicki/ libcpab, 2018

    2018

  42. [50]

    Comput- ing large deformation metric mappings via geodesic flows of diffeomorphisms,

    M. F. Beg, M. I. Miller, A. Trouv ´e, and L. Younes, “Comput- ing large deformation metric mappings via geodesic flows of diffeomorphisms,” IJCV, 2005

    2005

  43. [51]

    Recurrent neural networks for time series classification,

    M. H¨ usken and P. Stagge, “Recurrent neural networks for time series classification,” Neurocomputing, vol. 50, pp. 223–235, 2003

    2003

  44. [52]

    Recurrent neural network based language model

    T. Mikolov, M. Karafi´at, L. Burget, J. Cernock` y, and S. Khu- danpur, “Recurrent neural network based language model.” in Interspeech, vol. 2, no. 3. Makuhari, 2010, pp. 1045–1048

    2010

  45. [53]

    Deep learning for time series classification: a review,

    H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Deep learning for time series classification: a review,” arXiv preprint arXiv:1809.04356, 2018

    2018 arXiv

  46. [54]

    Time series classification from scratch with deep neural networks: A strong baseline,

    Z. Wang, W. Yan, and T. Oates, “Time series classification from scratch with deep neural networks: A strong baseline,” in Neural Networks (IJCNN), 2017 International Joint Con- ference on. IEEE, 2017, pp. 1578–1585

    2017

  47. [55]

    Identity mappings in deep residual networks,

    K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” in European conference on computer vision. Springer, 2016, pp. 630–645

    2016

  48. [56]

    Inceptiontime: Finding alexnet for time series classification,

    H. Ismail Fawaz, B. Lucas, G. Forestier, C. Pelletier, D. F. Schmidt, J. Weber, G. I. Webb, L. Idoumghar, P.-A. Muller, and F. Petitjean, “Inceptiontime: Finding alexnet for time series classification,” Data Mining and Knowledge Discovery, vol. 34, no. 6, pp. 1936–1962, 2020

    1936

  49. [57]

    Inception-v4, inception-resnet and the impact of residual connections on learning,

    C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi, “Inception-v4, inception-resnet and the impact of residual connections on learning,” in Thirty-first AAAI conference on artificial intelligence, 2017

    2017

  50. [58]

    Multilevel wavelet decomposition network for interpretable time series analysis,

    J. Wang, Z. Wang, J. Li, and J. Wu, “Multilevel wavelet decomposition network for interpretable time series analysis,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018, pp. 2437–2446

    2018

  51. [59]

    A log-euclidean polyaffine framework for locally rigid or affine registration,

    V. Arsigny, O. Commowick, X. Pennec, and N. Ayache, “A log-euclidean polyaffine framework for locally rigid or affine registration,” in BIR. Springer, 2006

    2006

  52. [60]

    Sparse adaptive parameterization of variability in image ensembles,

    S. Durrleman, S. Allassonni`ere, and S. Joshi, “Sparse adaptive parameterization of variability in image ensembles,” IJCV, 2013

    2013

  53. [61]

    Bayesian mixed effect atlas estimation with a diffeomorphic deforma- tion model,

    S. Allassonniere, S. Durrleman, and E. Kuhn, “Bayesian mixed effect atlas estimation with a diffeomorphic deforma- tion model,” SIAM Journal on Imaging Sciences , 2015

    2015

  54. [62]

    Deriving the CPAB derivative,

    O. Freifeld, “Deriving the CPAB derivative,” Ben-Gurion University, Tech. Rep., 2018

    2018

  55. [63]

    Gaussian processes in machine learning,

    C. E. Rasmussen, “Gaussian processes in machine learning,” in Advanced lectures on machine learning. Springer, 2004, pp. 63–71

    2004

  56. [64]

    A prototypical triplet loss for cover detection,

    G. Doras and G. Peeters, “A prototypical triplet loss for cover detection,” in ICASSP. IEEE, 2020

    2020

  57. [65]

    On the average value of correlated time series, with applications in dendroclimatology and hydrometeorology,

    T. M. Wigley, K. R. Briffa, and P. D. Jones, “On the average value of correlated time series, with applications in dendroclimatology and hydrometeorology,” Journal of climate and Applied Meteorology, vol. 23, no. 2, pp. 201–213, 1984

    1984

  58. [66]

    Gusfield, Algorithms on strings, trees, and sequences: computer science and computational biology

    D. Gusfield, Algorithms on strings, trees, and sequences: computer science and computational biology . Cambridge University Press, 1997

    1997

  59. [67]

    Fast computation of wasserstein barycenters,

    M. Cuturi and A. Doucet, “Fast computation of wasserstein barycenters,” in International Conference on Machine Learn- ing, 2014, pp. 685–693

    2014

  60. [68]

    Time series averaging using multi-tasking autoencoder,

    T. Terefe, M. Devanne, J. Weber, D. Hailemariam, and G. Forestier, “Time series averaging using multi-tasking autoencoder,” in 2020 IEEE 32nd International Conference on Tools with Artificial Intelligence (ICTAI). IEEE, 2020, pp. 1065–1072

    2020

  61. [69]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” CoRR, 2014. [Online]. Available: http://arxiv.org/abs/1412.6980

    2014 arXiv

  62. [70]

    tsai - a state-of-the-art deep learning library for time series and sequential data,

    I. Oguiza, “tsai - a state-of-the-art deep learning library for time series and sequential data,” Github, 2022. [Online]. Available: https://github.com/timeseriesAI/tsai

    2022

  63. [71]

    Asymptotic theory for the principal component analysis of a vector random function: some applications to statistical inference,

    J. Dauxois, A. Pousse, and Y. Romain, “Asymptotic theory for the principal component analysis of a vector random function: some applications to statistical inference,” Journal of multivariate analysis, vol. 12, no. 1, pp. 136–154, 1982

    1982

  64. [72]

    Some tools for functional data analysis,

    J. O. Ramsay and C. Dalzell, “Some tools for functional data analysis,” Journal of the Royal Statistical Society: Series B (Methodological), vol. 53, no. 3, pp. 539–561, 1991

    1991

  65. [73]

    Discovering precise DIFFEOMORPHIC TEMPORAL ALIGNMENT NETWORKS 16 temporal patterns in large-scale neural recordings through robust and interpretable time warping,

    A. H. Williams, B. Poole, N. Maheswaranathan, A. K. Dhawale, T. Fisher, C. D. Wilson, D. H. Brann, E. M. Trautmann, S. Ryu, R. Shustermanet al., “Discovering precise DIFFEOMORPHIC TEMPORAL ALIGNMENT NETWORKS 16 temporal patterns in large-scale neural recordings through robust ...

    2020

  66. [74]

    A deep moving-camera background model,

    G. Erez, R. S. Weber, and O. Freifeld, “A deep moving-camera background model,” in European Conference on Computer Vision. Springer, 2022, pp. 177–194

    2022

  67. [75]

    Spacejam: a lightweight and regularization- free method for fast joint alignment of images,

    N. Barel, R. S. Weber, N. Mualem, S. E. Finder, and O. Freifeld, “Spacejam: a lightweight and regularization- free method for fast joint alignment of images,” in European Conference on Computer Vision. Springer, 2024, pp. 180– 197

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.