Pith. sign in

REVIEW 4 major objections 6 minor 23 references

SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read SIDDA is a domain-adaptation training algorithm that uses a dynamically regularized Sinkhorn divergence and trainable loss weights; paired with D4-equivariant networks, it lifts target-domain accuracy by up to about 40 percentage points…

desk verdict A useful empirical DA paper whose 'out-of-the-box' claim needs an ablation of the dynamic sigma schedule to hold. read the letter →

arxiv 2501.14048 v2 pith:HF7R7TPA submitted 2025-01-23 cs.LG astro-ph.GAcs.AIcs.CV

Sneh Pandya , Purvik Patel , Brian D. Nord , Mike Walmsley , Aleksandra Ćiprijanović This is my paper · ORCID
classification cs.LGastro-ph.GAcs.AIcs.CV
keywords domainadaptationSinkhorndivergenceequivariantneuralnetworksoptimaltransportmodelcalibrationcovariateshiftgalaxymorphologyclassificationdihedralgroup
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

Domain adaptation usually requires expensive per-dataset tuning of the distance metric and the balance between classification and alignment losses. SIDDA replaces both knobs with dynamic mechanisms: the entropy penalty of a Sinkhorn divergence is recomputed every epoch from the current source-target latent distances, and the classification-versus-alignment balance is learned through two trainable weights. The paper claims this makes effective domain alignment nearly turnkey, and it demonstrates the claim on simulated shapes, simulated astronomical objects, noisy and blurred handwritten digits, real galaxy images from two surveys, and optical-to-radar remote sensing images. Reported gains reach about 40 percentage points of target-domain accuracy, are larger with D4-equivariant networks than with plain CNNs, and come with order-of-magnitude improvements in expected calibration error and Brier score on some target domains. If true, scientists could cheaply take a classifier trained on one telescope's labeled data and use it on another telescope's unlabeled data.

What carries the argument

The load-bearing object is the Sinkhorn divergence $S_\sigma(\mu,\nu) = OT_\sigma(\mu,\nu) - \frac{1}{2}OT_\sigma(\mu,\mu) - \frac{1}{2}OT_\sigma(\nu,\nu)$, a symmetrized, entropy-regularized optimal transport distance that interpolates between the Wasserstein distance as $\sigma\to0$ and MMD as $\sigma\to\infty$. SIDDA uses it as the domain-alignment loss, but instead of fixing $\sigma$ it recomputes $\sigma_\ell$ each epoch from the current maximum pairwise squared distance between layer-normalized source and target latents, with a floor to keep the Sinkhorn iterations stable; a combined batch of source and target images ensures shared batch statistics. The second mechanism is a trainable loss balance inspired by uncertainty-weighted multi-task learning, in which two scalars $\eta_1,\eta_2$ control the relative weight of cross-entropy and alignment loss and are clipped so alignment never dominates classification. The argument for pairing this with equivariant networks is that a constrained latent space (for example, one built from $D_4$-equivariant convolutions) makes source and target distributions easier to align; the paper supports this with lower Jensen-Shannon distances for higher dihedral group orders.

What would settle it

Apply SIDDA to a fresh domain-shift pair using its default schedule and a warm-up length that is not tuned to that dataset; if target-domain accuracy gains over the no-DA baseline vanish, or if good results require changing the constants 0.05 and 0.01 in the schedule, the out-of-the-box claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that domain adaptation can be made largely self-tuning by choosing the Sinkhorn divergence as the alignment loss and actively scheduling the two quantities that normally need tuning: the entropic regularization $\sigma$ of the transport plan and the weights of the classification and alignment losses. Concretely, SIDDA sets $\sigma_\ell = \max\bigl(0.05 \max_{i,j}\|z_i - z^*_j\|^2, 0.01\bigr)$ each epoch from a combined batch of source and target latents, and writes the total loss as $\frac{1}{2\eta_1^2}\mathcal{L}_{\mathrm{CE}} + \frac{1}{2\eta_2^2}\mathcal{L}_{\mathrm{DA}} + \log(|\eta_1\eta_2|)$ with trainable $\eta_1,\eta_2$ clipped so the alignment term never dominates. On five datasets the paper reports that SIDDA improves unlabeled target-domain accuracy in every setup, with the largest gain about $+40$ percentage points on simulated astronomical objects, and that $D_4$-equivariant networks consistently beat CNNs and produce lower Jensen-Shannon distance between source and target latent distributions. It also reports order-of-magnitude calibration improvements in the best cases, and a clear boundary: on the extreme optical-to-SAR shift, target gains are small and source accuracy drops slightly.

Load-bearing premise

The out-of-the-box promise rests on the automatic rule $\sigma_\ell = \max\bigl(0.05 \max_{i,j}\|z_i - z_j^*\|^2, 0.01\bigr)$ being equally good across datasets, but the paper tunes the warm-up phase separately for every dataset, so that premise is not yet demonstrated.

Editorial extensions

If this is right

  • A model trained on labeled images from one survey can be applied to unlabeled images from a second survey, with the largest reported gain of about 40 percentage points in target-domain accuracy and no manual loss-weight tuning.
  • Equivariant networks and SIDDA reinforce each other: in the reported experiments the D4 model beats the CNN both with and without adaptation, and target accuracy rises with dihedral group order up to D8.
  • SIDDA can serve as a calibration regularizer, not just an accuracy booster: in the successful alignments the ECE and Brier score drop on the target domain, sometimes by more than an order of magnitude.
  • The method's computational cost stays close to MMD (minutes on one GPU) while approaching the accuracy of a near-Wasserstein formulation that takes two to three times longer.
  • On cross-wavelength shifts (optical to SAR) the gains are real but small and come with a source-domain accuracy drop, so the method's reach stops where the latent distributions are too far apart to align near the classifier.

Reading between the lines

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

  • A direct test of the turnkey premise would be to apply SIDDA to a new survey pair with the default schedule and a fixed warm-up; the paper leaves warm-up length as a per-dataset tuned value, so the zero-tuning reading of the method is stronger than the evidence.
  • Because the paper ties alignment quality to the Jensen-Shannon distance between latent distributions, one can use that distance as a cheap, label-free early stopping or dataset-selectivity criterion for when SIDDA is likely to help.
  • The D8 overfitting suggests the group order is another dial: a model that anneals equivariance strength during training might capture the accuracy benefit of high-order symmetry without the late-training overfitting the paper reports.
  • The calibration improvement raises the possibility that DA can be used as a principled uncertainty regularizer in safety-critical image classification, a connection the paper introduces but does not develop.
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 / 6 minor

Summary. The paper introduces SIDDA, a domain adaptation method that aligns source and target latent distributions using the Sinkhorn divergence with a per-epoch adaptive entropy regularization and trainable loss weights. The method is evaluated on simulated datasets, MNIST-M, Galaxy Zoo Evo, and MRSSC2, combined with CNNs and D4-equivariant networks, and compared with MMD, Wasserstein distance, and fixed loss coefficients. The central claim is that SIDDA provides 'out-of-the-box' domain adaptation requiring minimal hyperparameter tuning while improving target accuracy by up to 40 percentage points and calibration by an order of magnitude.

Significance. If the central claim holds, SIDDA is a practically valuable contribution: it offers a nearly turnkey distance-based DA method that works with standard CNN architectures and improves further with equivariant networks. The study is carefully executed in its empirical parts: three seeds per configuration, uncertainties reported, a precise algorithm description, and public code and data availability. The calibration results, especially the order-of-magnitude ECE/Brier improvements on the shapes and astronomical datasets, are striking. However, the significance is qualified by the fact that the two mechanisms that are supposed to remove hyperparameter dependence, the sigma schedule in Eq. (7) and the loss-weighting clipping rule, are not ablated, and the warm-up phase is explicitly tuned per experiment.

major comments (4)
  1. [Section 4.2, Table A1] The warm-up duration is explicitly tuned for each dataset and model (5 to 30 epochs), as stated in Section 4.2 and Table A1, which directly contradicts the abstract's 'out-of-the-box' and 'minimal hyperparameter tuning' claim. Since the warm-up determines when DA is activated, the reported gains could depend on this per-dataset choice rather than on a robust automatic schedule. Please either justify the warm-up as a mild and non-critical hyperparameter (for example, by showing performance is insensitive to it) or ablate it, and adjust the claims accordingly.
  2. [Section 2.3, Eq. (7)] The dynamic Sinkhorn regularization in Eq. (7) introduces constants (0.05, 0.01) and the max operator without any ablation. The paper notes that 'we found that this threshold worked best' for the eta2/eta1 >= 0.25 clip, which is another hand-picked constant. Section 5.6 only varies the loss coefficients, not the sigma schedule. This is the central new ingredient that is supposed to make the method hyperparameter-free; without an ablation of the prefactor, the floor, and the choice of max versus a quantile or mean, the claim of a robust automatic schedule is not established.
  3. [Section 5.5, Table 5] The MMD baseline uses a single fixed Gaussian kernel width (epsilon = 0.05) with no tuning or multi-kernel approach, despite the paper's own Section 2.1 noting MMD's sensitivity to kernel choice. As a result, the comparison in Table 5 may understate MMD's performance, and the conclusion that SIDDA 'combines the performance of the Wasserstein distance with the efficiency of MMD' is not fully supported. Please include a tuned or multi-kernel MMD baseline, or clearly frame the comparison as 'against an untuned MMD' rather than against MMD in general.
  4. [Section 5.6, Table 6] The comparison with fixed loss coefficients uses only three ad-hoc formulations (LC,D, LC,10D, L10C,D). The paper itself states that 'additional model initializations and experiments are necessary to draw more definitive conclusions regarding the stability of dynamic loss weighting.' This is an explicit limitation that weakens the claim that trainable coefficients avoid manual loss-weight tuning; a small grid over the fixed coefficient would strengthen the evidence.
minor comments (6)
  1. [Section 2.3] The text refers to 'Figure 2.3' when describing the pipeline, but the figure is numbered 'Figure 1'; fix the cross-reference.
  2. [Section 1] The phrase 'representation learningingraphautoencodersandpredictingmolecular properties' appears to be missing spaces and possibly a conjunction; correct the typo.
  3. [Section 5.1] In the sentence 'It is also seen that the latent distribution of theD4 is more clustered...' there is a missing space before 'D4'; similar spacing issues occur throughout the paper (e.g., 'theD4-DA model').
  4. [Section 2.3] The discussion of Kendall et al. [2018] admits that no uncertainty is output by the model and that 'the exact written form of loss weights is not important.' This is a reasonable heuristic, but it should be stated more plainly as an adaptive weighting scheme rather than inheriting the uncertainty-based justification from Kendall et al.
  5. [Section 4.2, Algorithm 1] Algorithm 1 clips eta2 as max(eta2, 0.25*eta1) after the optimizer step, but the text in Section 2.3 states the constraint eta2/eta1 >= 0.25; make the relationship explicit (for example, note that clipping is applied to the updated values).
  6. [Tables 1, 4, and 5] The table captions contain 'T able' with a space, likely a LaTeX artifact; correct the formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SIDDA's target-accuracy gains are measured empirically and are not implied by its loss definition; admitted per-dataset tuning is a robustness limitation, not a circular reduction.

full rationale

Walking the claimed derivation chain: the paper defines LDA as the Sinkhorn divergence (Eq. 5, Sec. 2.3) and the total loss as Eq. 6 with trainable weights, plus the dynamic sigma schedule of Eq. 7. Target-domain accuracy, ECE, and Brier score are then measured on held-out target test sets (Tables 1 and 4). No equation in the paper defines these metrics as functions of LDA or of sigma_l, so the improvement is not true by construction. The dynamic-sigma schedule is a heuristic (0.05 prefactor and 0.01 floor) taken from the geomloss/Feydy scaling convention, and the paper explicitly admits empirical choices: 'For our implementation, we found that this threshold worked best' (Sec. 2.3, eta2/eta1 clip) and 'The duration of the warm-up phase was tuned for each experiment' (Sec. 4.2, Table A1). Those admissions weaken the 'out-of-the-box/minimal tuning' claim, but they are correctness and robustness limitations rather than circular reductions: the reported accuracies do not reduce to those tuned values by definition. The self-citations (Pandya et al. 2023 for choosing D4 and for the group-order robustness trend; Ciprijanovic et al. for DeepAstroUDA) are contextual and are not used to prove SIDDA's central results; the SIDDA-vs-baseline comparisons in Tables 5 and 6 are independent experiments against MMD, Wasserstein, and fixed-loss-weight variants. No uniqueness theorem or ansatz is imported from the authors' prior work to force the method's form. Hence no step in the derivation chain is circular; score 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The ledger shows that SIDDA's 'out-of-the-box' claim rests on several hand-chosen constants (0.05, 0.01, the η clip) and on per-dataset warm-up tuning, plus standard DA and Sinkhorn theory borrowed from prior work. There are no invented physical entities. The empirical accuracy results do not depend on fitted constants in a circular way, but the automation claim is fragile.

free parameters (4)
  • sigma scale prefactor = 0.05
    Eq. 7; chosen by hand with a unit-cube scaling justification from Feydy et al.; no ablation reported.
  • sigma lower bound = 0.01
    Eq. 7; introduced to avoid numerical instability near the unregularized Wasserstein limit; no ablation.
  • eta ratio clip = eta2/eta1 >= 0.25
    Section 2.3 and Algorithm 1 line 14: authors state they found this threshold 'worked best and stabilized training'; a hand-chosen constant.
  • warm-up epochs = 5, 10, 20, or 30 depending on dataset/model
    Table A1; the warm-up duration is explicitly tuned per experiment, which undermines the 'minimal hyperparameter tuning' claim.
assumptions (6)
  • domain assumption Covariate shift: P(X) differs between source and target but P(Y|X) is unchanged.
    Standard DA setting stated in the Abstract and Section 1; all experiments assume label-conditional distributions are preserved.
  • standard math Sinkhorn divergence S_sigma is a suitable unbiased distance between empirical latent distributions.
    Uses Feydy et al. 2018 results (Eq. 5) that S_sigma corrects the bias of regularized OT and interpolates between OT and MMD.
  • standard math JS-distance lower bound on target loss (Shui et al., Eq. 10) holds and motivates minimizing Sinkhorn loss.
    Section 2.4; the bound is cited but not re-derived; the paper's interpretation that reducing JS distance improves target performance relies on it.
  • domain assumption Equivariant features (D4) reduce the cost of aligning latent distributions compared to generic CNN features.
    Section 2.4 argues symmetry-constrained features align more easily; supported only by empirical results, not proven.
  • domain assumption Layer normalization of latent vectors stabilizes the dynamic sigma computation.
    Section 4.1: the authors state they found layer norm useful for stable sigma computation; no theoretical guarantee.
  • ad hoc to paper Kendall-style uncertainty weighting remains valid when no uncertainty is output by the model.
    Section 2.3 adapts Eq. 6 from Kendall et al. 2018 by absorbing the factor of two, which the paper acknowledges may not be exactly motivated in the classification setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks." pith.science (2026). https://pith.science/paper/HF7R7TPA

@misc{pith2026250114048,
  author       = {Pith},
  title        = {Pith review of: SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HF7R7TPA}},
  note         = {Machine review of arXiv:2501.14048}
}
abstract

Modern neural networks (NNs) often do not generalize well in the presence of a "covariate shift"; that is, in situations where the training and test data distributions differ, but the conditional distribution of classification labels remains unchanged. In such cases, NN generalization can be reduced to a problem of learning more domain-invariant features. Domain adaptation (DA) methods include a range of techniques aimed at achieving this; however, these methods have struggled with the need for extensive hyperparameter tuning, which then incurs significant computational costs. In this work, we introduce SIDDA, an out-of-the-box DA training algorithm built upon the Sinkhorn divergence, that can achieve effective domain alignment with minimal hyperparameter tuning and computational overhead. We demonstrate the efficacy of our method on multiple simulated and real datasets of varying complexity, including simple shapes, handwritten digits, and real astronomical observations. SIDDA is compatible with a variety of NN architectures, and it works particularly well in improving classification accuracy and model calibration when paired with equivariant neural networks (ENNs). We find that SIDDA enhances the generalization capabilities of NNs, achieving up to a $\approx40\%$ improvement in classification accuracy on unlabeled target data. We also study the efficacy of DA on ENNs with respect to the varying group orders of the dihedral group $D_N$, and find that the model performance improves as the degree of equivariance increases. Finally, we find that SIDDA enhances model calibration on both source and target data--achieving over an order of magnitude improvement in the ECE and Brier score. SIDDA's versatility, combined with its automated approach to domain alignment, has the potential to advance multi-dataset studies by enabling the development of highly generalizable models.

Figures

Figures reproduced from arXiv: 2501.14048 by the authors.

Figure 1
Figure 1. SIDDA pipeline. The source and target domain batches of size n, xn and x ∗ n, are first concatenated into a single batch X before being passed into the model. After passing through the convolutional layers, the neural network produces a combined batch of latent vectors, Z, extracted from the final linear layer. This layer is positioned just before the output layer, which generates the class probabilities, Y. Both Z … view at source ↗
Figure 2
Figure 2. Example images for simulated datasets in the source domain (top row) and the target domain (bottom row) with corresponding labels. Left Panels: Shapes dataset, featuring lines, rectangles, and circles, simulated with DeepBench. This dataset includes variations in object positions and orientations, with Poisson noise added and normalized relative to the image signal in the target domain. Middle Panels: Astronomical o… view at source ↗
Figure 3
Figure 3. Top Panel: Example source domain images from the GZ Evo dataset with corresponding labels. Images are from GZ2 data observed by SDSS. Bottom Panel: Example target domain images from the GZ Evo dataset with the same labels. Images are from GZ DESI (combined observations from the DESI Imaging Surveys). science project that labels galaxy images through online participation. GZ Evo combines labeled image datasets across… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Top Panel: Example source domain optical images from the MRSSC2 dataset with corresponding labels. Bottom Panel: Example target domain SAR images from the MRSSC2 dataset with the same labels. to the desired symmetries. Therefore, they learn transformation-invariant fea…
Figure 5
Figure 5. Figure 5: MNIST-M (Noise) latent distributions visualized with isomaps. Source (solid) and target (hollow) latent distributions are plotted atop each other to visualize latent distribution misalignment. The inclusion of DA clearly improves the alignment of source and target late…
Figure 6
Figure 6. Figure 6: Evolution of the trainable coefficients, η −2 1 and η −2 2 , and the Sinkhorn plan regularization strength σℓ for CNN￾DA trained on MNIST-M (Noise) after the LCE-only warm￾up period. The shaded regions correspond to 1σ uncertainties from three training runs initialized…
Figure 7
Figure 7. Figure 7: Jensen-Shannon (JS) distances for CNN-DA and DN -DA (N ∈ {1, 2, 4, 8}) models trained on MNIST-M (Noise). Shaded regions correspond to 1σ uncertainties from three training runs initialized with varying random seeds. All models underwent a 30-epoch warm-up phase without…
Figure 8
Figure 8. Figure 8: MRSSC2 latent distributions, visualized using isomaps, with the source domain shown as solid markers and the target domain as hollow markers. Both the CNN and D4 models exhibit substantially less clustering in the latent space—compared to the MNIST-M (Noise) dataset sh…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 5 canonical work pages

  1. [5]

    URL https: //doi.org/10.1007/978-3-031-16749-2_9

    doi: 10.1007/978-3-031-16749-2_9. URL https: //doi.org/10.1007/978-3-031-16749-2_9. Gabriele Cesa, Leon Lang, and Maurice Weiler. A program to build E(N)-equivariant steerable CNNs. In International Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=WE4qe9xlnQw. A. Ćiprijanović, A. Lewis, K. Pedro, S. Madireddy, B. Nord, G....

    doi:10.1007/978-3-031-16749-2_9

  2. [8]

    Jean Feydy, Thibault Séjourné, François-Xavier Vialard, Shun ichi Amari, Alain Trouvé, and Gabriel Peyré

    URL https://api.semanticscholar.org/ CorpusID:222209143. Jean Feydy, Thibault Séjourné, François-Xavier Vialard, Shun ichi Amari, Alain Trouvé, and Gabriel Peyré. Interpolating between optimal transport and mmd using sinkhorn divergences, 2018. URL https:// arxiv.org/abs/1810.08278. Jean Feydy, Thibault Séjourné, François-Xavier Vialard, Shun-ichi Amari, ...

    arXiv 2018

  3. [11]

    URL https://arxiv.org/abs/2403.02314. John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, An- drew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, ...

    arXiv 2021

  4. [12]

    Guoliang Kang, Lu Jiang, Yi Yang, and Alexander Hauptmann

    URL https://arxiv.org/abs/2406.10685. Guoliang Kang, Lu Jiang, Yi Yang, and Alexander Hauptmann. Contrastive adaptation network for unsupervised domain adaptation.2019 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 4888–4897, 2019. URLhttps:// api.semanticscholar.org/CorpusID:57572938. Alex Kendall, Yarin Gal, and Roberto C...

    arXiv 2019

  5. [14]

    URL https://api.semanticscholar.org/ CorpusID:271403939. 24 J. Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory, 37(1):145–151, 1991. doi: 10.1109/18.61115. Kang Liu, Jian Yang, and Shengyang Li. Remote- Sensing Cross-Domain Scene Classification: A Dataset and Benchmark. Remote Sensing, 14(18): 4635, September...

    arXiv doi:10.48550/arxiv.2208.07422 1991

  6. [16]

    On the Decreasing Power of Kernel and Distance based Nonparametric Hypothesis Tests in High Dimensions

    URL https://arxiv.org/abs/1406.2083. Andrea Roncoli, Aleksandra Ćiprijanović, Maggie Voet- berg, Francisco Villaescusa-Navarro, and Brian Nord. Domain Adaptive Graph Neural Networks for Con- straining Cosmological Parameters Across Multiple Data Sets. arXiv e-prints, art. arXiv:2311.01588, November 2023. doi: 10.48550/arXiv.2311.01588. Y. Rubner, C. Tomas...

    work page Pith review arXiv 2023

  7. [20]

    URL https://www

    doi: 10.2172/1989920. URL https://www. osti.gov/biblio/1989920. Mike Walmsley, Chris Lintott, Tobias Géron, Sandor Kruk, Coleman Krawczyk, Kyle W Willett, Steven Bamford, Lee S Kelvin, Lucy Fortson, Yarin Gal, William Keel, Karen L Masters, Vihang Mehta, Brooke D Simmons, Rebecca Smethurst, Lewis Smith, Elisabeth M Baeten, and Christine Macmillan. Galaxy ...

    work page arXiv

  8. [21]

    doi: 10.1093/mnras/stad2919

    ISSN 0035-8711. doi: 10.1093/mnras/stad2919. Mike Walmsley, Micah Bowles, Anna M. M. Scaife, Jason Shingirai Makechemu, Alexander J. Gordon, Annette M. N. Ferguson, Robert G. Mann, James Pearson, Jürgen J. Popp, Jo Bovy, Josh Speagle, Hugh Dickinson, Lucy Fortson, Tobias Géron, Sandor Kruk, Chris J. Lintott, Kameswara Mantha, Devina Mohan, David O’Ryan, a...

    arXiv

Show all 23 references
  1. [22]

    URL https://arxiv.org/abs/1911.08251. Kyle W. Willett, Chris J. Lintott, Steven P. Bamford, Karen L. Masters, Brooke D. Simmons, Kevin R. V. Casteels, Edward M. Edmondson, Lucy F. Fortson, Sugata Kaviraj, William C. Keel, Thomas Melvin, Robert C. Nichol, M. Jordan Raddick, Kev...

    1911 arXiv

  2. [26]

    URL http: //dx.doi.org/10.1109/TEVC.2019.2890858

    doi: 10.1109/tevc.2019.2890858. URL http: //dx.doi.org/10.1109/TEVC.2019.2890858. Baochen Sun, Jiashi Feng, and Kate Saenko. Correla- tion alignment for unsupervised domain adaptation. ArXiv, abs/1612.01939, 2016. URL https://api. semanticscholar.org/CorpusID:10084602. Paxson ...

    2019

  3. [746]

    URLhttp: //dx.doi.org/10.1051/0004-6361/202141938

    doi: 10.1051/0004-6361/202141938. URLhttp: //dx.doi.org/10.1051/0004-6361/202141938. J. L. Sersic. Photometry of southern galaxies: NGC

    doi

  4. [2014]

    Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas Guibas

    URL https://api.semanticscholar.org/ CorpusID:13347901. Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas Guibas. Vector neurons: A general framework for so(3)-equivariant networks, 2021. URL https: //arxiv.org/abs/2104.12229. Li Deng. Th...

    2021 arXiv

  5. [2015]

    Alex Hernández-García and Peter König

    URL https://arxiv.org/abs/1512.03385. Alex Hernández-García and Peter König. Data augmentation instead of explicit regularization, 2020. URL https://arxiv.org/abs/1806.03852. Sergey Ioffe and Christian Szegedy. Batch normal- ization: Accelerating deep network training by re- d...

    2020 arXiv

  6. [2016]

    Weitao Du, He Zhang, Yuanqi Du, Qi Meng, Wei Chen, Bin Shao, and Tie-Yan Liu

    URL https://api.semanticscholar.org/ CorpusID:12047850. Weitao Du, He Zhang, Yuanqi Du, Qi Meng, Wei Chen, Bin Shao, and Tie-Yan Liu. Se(3) equivariant graph neural networks with complete local frames, 2022. URL https://arxiv.org/abs/2110.14811. Gamaleldin F. Elsayed, Prajit R...

    2022 arXiv

  7. [2018]

    URL https://arxiv.org/abs/1802.03690. S. Kullback and R. A. Leibler. On information and sufficiency. The Annals of Mathematical Statistics, 22(1):79–86, March 1951. doi: 10. 1214/aoms/1177729694. URL https://doi.org/10. 1214%2Faoms%2F1177729694. Jinlong Li, Runsheng Xu, Jin Ma...

    1951 arXiv

  8. [2019]

    URL https://arxiv.org/abs/1912.02292. Dylan Nelson, Volker Springel, Annalisa Pillepich, Vicente Rodriguez-Gomez, Paul Torrey, Shy Genel, Mark Vogelsberger, Ruediger Pakmor, Federico Marinacci, Rainer Weinberger, Luke Kelley, Mark Lovell, Benedikt Diemer, and Lars Hernquist. T...

    1912 arXiv

  9. [2020]

    Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis

    URL https://arxiv.org/abs/2002.02959. Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so(3) equivariant representations with spherical cnns, 2018. URL https://arxiv.org/abs/1711.06721. Abolfazl Farahani, Sahar Voghoei, Khaled M. Rash...

    2002 arXiv

  10. [2022]

    doi: 10.1088/1475-7516/ 2022/09/004

    ISSN 1475-7516. doi: 10.1088/1475-7516/ 2022/09/004. URL http://dx.doi.org/10.1088/ 1475-7516/2022/09/004. Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence...

    2022 doi

  11. [2023]

    AleksandraĆiprijanović, DianaKafkes, GregorySnyder, F

    doi: 10.1088/2632-2153/acca5f. AleksandraĆiprijanović, DianaKafkes, GregorySnyder, F. Javier Sánchez, Gabriel Nathan Perdue, Kevin Pedro, Brian Nord, Sandeep Madireddy, and Stefan M. Wild. DeepAdversaries: examining the robustness of deep learning models for galaxy morphology ...

    arXiv

  12. [2024]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun

    URL https://arxiv.org/abs/2401.12588. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition,

    arXiv

  13. [5128]

    Changjian Shui, Qi Chen, Jun Wen, Fan Zhou, Christian Gagné, and Boyu Wang

    The Observatory, 78:24–29, February 1958. Changjian Shui, Qi Chen, Jun Wen, Fan Zhou, Christian Gagné, and Boyu Wang. A novel domain adaptation theory with jensen–shannon divergence. Knowledge-Based Systems, 257:109808, 2022. ISSN 0950-7051. doi: https://doi.org/10.1016/j.knos...

    1958

  14. [6904]

    URL https://doi

    doi: 10.1145/3400066. URL https://doi. org/10.1145/3400066. 27 Appendix A. Training and Model Configurations T able A1. Model training configurations for each dataset. The learning rate followed a multiplicative decay schedule from10−2 to 10−4, applied at epochs⌊Total Epochs/3...

    doi

  15. [8828]

    URL http: //dx.doi.org/10.1109/TPAMI.2010.161

    doi: 10.1109/TPAMI.2010.161. URL http: //dx.doi.org/10.1109/TPAMI.2010.161. Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization, 2016. URL https: //arxiv.org/abs/1607.06450. Alexander Bogatskiy, Brandon Anderson, Jan T. Offermann, Marwah Roussi, David W...

    2010 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.