Pith. sign in

REVIEW 4 major objections 4 minor 36 references

FAAGC: Feature Augmentation on Adaptive Geodesic Curve Based on the shape space theory

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

Pith's one-line read FAAGC claims that sampling synthetic features along a per-class geodesic arc in pre-shape space improves few-shot classification, reporting gains on five benchmarks and a roughly 9x speedup over the earlier FAGC.

desk verdict Useful incremental follow-up on FAGC with a real speedup, but the headline claim of consistent superiority is contradicted by the paper's own tables; referee it, expecting major revision. read the letter →

arxiv 2501.18619 v2 pith:OYSTNMS4 submitted 2025-01-25 cs.CV cs.LG

classification cs.CVcs.LG
keywords FeatureaugmentationShapespacetheoryPre-shapeGeodesiccurveDatascarcityFew-shotlearningImageclassificationGradientdescent
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

FAAGC is a feature-level data augmentation method for deep classifiers trained with very few examples per class. The paper claims that projecting each class's features into the pre-shape space, the unit sphere obtained after centering and normalizing, and fitting an adaptive geodesic arc to them yields synthetic samples that stay closer to the true feature distribution than linear interpolation or noise-based augmentation. Across five image benchmarks reduced to 3–20 samples per class, the method reports consistent accuracy gains, with the largest improvement on CIFAR-100@5 (67.87% vs 66.41% without augmentation). It also claims a large speedup over the earlier FAGC approach, cutting geodesic-curve fitting time from about 352 seconds to 39 seconds on CIFAR-10@5. If correct, the contribution is a generic, classifier-agnostic plug-in that makes scarce-data training more reliable without hand-designed image transforms.

What carries the argument

The machinery is the pre-shape space construction and the adaptive geodesic arc. A feature vector $v\in\mathbb{R}^d$ is expanded to paired coordinates $(v[1],v[1]),\dots,(v[d],v[d])$, centered and normalized to a unit-sphere vector $\tau\in S^{2d-3}_*$; by shape space theory, all objects of one shape lie on a great circle, so a class is represented by an arc $\Gamma(\tau_{\text{start}},\tau_{\text{end}})(z)=\frac{\sin[(1-z)\theta]}{\sin\theta}\tau_{\text{start}}+\frac{\sin(z\theta)}{\sin\theta}\tau_{\text{end}}$, with $\theta$ the geodesic distance between endpoints and $z\in[0,1]$ the normalized arc position. The endpoints are learned per class: the loss $L_{\text{train}}=L_{\text{Sim}}+\beta\,L_{\text{Diverg}}$ aligns sampled points with real features by cosine similarity and forces the learnable positions $t$ to match $U(0,1)$ by sorted Wasserstein distance, while the endpoints are kept on the sphere by optimizing the unnormalized Euclidean vectors and projecting. This is what carries the argument: the fitted arc is the generative model, and sampling along it is the augmentation.

What would settle it

Fit the FAAGC geodesic for each class on a few-shot training set and measure the mean geodesic distance from every training feature to the fitted arc, or train a binary discriminator to separate real features from FAAGC-sampled features. If the residuals are large or the discriminator easily separates the two, for instance with a simple linear probe achieving above-chance accuracy, the arc does not represent the class and the augmented points are off-manifold, which would invalidate the claimed mechanism.

Watch

Extended reading notes

Core claim

The central claim is that, for each class, the deep features of a small training set lie close to a single geodesic arc in the pre-shape space, and that sampling new features along that arc creates useful training data. The paper implements this by duplicating each feature coordinate to form 2D landmarks, subtracting the mean, and normalizing onto the unit sphere $S^{2d-3}_*$; it then learns two endpoint vectors per class by gradient descent on a loss that combines a cosine-similarity term between sampled and real features with a Wasserstein term that keeps learnable sampling parameters close to the uniform distribution used at generation time. After fitting, augmented features are sampled from $\Gamma(\tau_{\text{start}},\tau_{\text{end}})(z)$ with $z\sim U(0,1)$. On five benchmarks at 5 samples per class, FAAGC reports the best accuracy among all compared feature-level augmentation methods, including a 1.46-point gain over the no-augmentation baseline on CIFAR-100@5, and it outperforms FAGC while reducing fitting time from 351.83s to 39.39s on CIFAR-10@5.

Load-bearing premise

The load-bearing premise is that each class's deep features, after duplicating coordinates, centering, and normalizing onto the unit sphere, lie close to a single geodesic arc, and that any point sampled on that arc still looks like a valid member of the class; the paper offers no geometric or statistical test of this assumption.

Editorial extensions

If this is right

  • At 5 samples per class, FAAGC reports the highest accuracy among compared methods on all five datasets; the largest gain over no augmentation is +1.46 points on CIFAR-100@5.
  • Gains are largest when data is scarcest: at 3 samples per class CIFAR-100 improves from 57.83% to 59.39%, while at 20 samples per class the gain shrinks to 0.26 points.
  • FAAGC improves accuracy for k-NN, SVM, and MLP classifiers, and across ViT, Swin, ResNet, and EfficientNet backbones, indicating it is not tied to one architecture.
  • The method stacks with image-level augmentation: combining FAAGC with RandAugment or AugMix on CIFAR-100@5 gives additional gains over either alone.
  • Compared with FAGC, FAAGC reduces fitting time from 351.83s to 39.39s on CIFAR-10@5 while improving accuracy from 85.20% to 85.84%.

Reading between the lines

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

  • (Editorial inference) The single-arc-per-class model should be testable directly: fitting FAAGC on a class with two well-separated visual modes, such as a digit written with open or closed top, would likely produce a high residual loss, predicting that a mixture of two geodesic arcs would outperform one.
  • (Editorial inference) The projection duplicates every coordinate to synthesize 2D landmarks; if the up-dimension step distorts the geometry, alternative embeddings, such as pairing consecutive coordinates or using a spherical map of the original vector, may change augmentation quality, and a controlled comparison would isolate the projection's role.
  • (Editorial inference) The reported speedup makes online augmentation during classifier training feasible, a setting the paper leaves unexplored, since the current pipeline augments features offline before classifier training.
  • (Editorial inference) The method's success suggests that pre-trained deep features of a class, in the low-data regime, concentrate near low-dimensional great-circle arcs; if so, principal geodesic analysis or multi-geodesic mixtures could yield further gains without new data.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 proposes FAAGC, a feature-level data augmentation method that projects deep image features into a pre-shape space by duplicating each feature coordinate, centering, and normalizing onto a unit sphere. For each class, it fits an adaptive geodesic arc by optimizing two endpoint vectors and a set of sampling parameters with a loss that combines a sample-similarity term and a distribution-divergence term, then augments by sampling uniformly along the fitted arc. Experiments on CIFAR-10, CIFAR-100, CUB-200, Fashion-MNIST, and Caltech101 under 5-samples-per-class conditions compare FAAGC with several feature augmentation baselines and with the prior FAGC method, reporting accuracy gains in several settings and a large runtime reduction.

Significance. If the results are taken at face value, the main contributions are a gradient-based geodesic fitting procedure that is roughly an order of magnitude faster than FAGC (Table V) and consistent gains over no augmentation and over most baselines in low-data settings (Tables I and IV). The method is simple, modular, and the appendix proof of equivalence between the geodesic and interpolation forms is mathematically sound. However, the paper's headline claim that FAAGC 'consistently outperforms all baseline methods' is not supported by its own tables, and the algorithm description contains an internal inconsistency that must be resolved before the results can be reproduced. No code or per-seed results are provided, so the small reported differences cannot be independently checked.

major comments (4)
  1. [Section IV-B, Tables I-III] The claim that 'FAAGC consistently outperforms all baseline methods under data-limited conditions across all benchmark datasets' is directly contradicted by the reported numbers. In Table I, on Caltech101@5, FAGC achieves 88.72±0.03 while FAAGC achieves 88.12±0.01, a deficit of 0.60 points that exceeds the reported standard deviations. In Table II, on CIFAR-100@5 with an MLP classifier, FAGC obtains 66.28 while FAAGC obtains 66.23. In Table III, on CIFAR-100@5 with Swin-Transformer, Manifold-Mixup obtains 73.34 while FAAGC obtains 73.26. These are not incidental comparisons: FAGC is the prior geodesic method that FAAGC claims to improve, and Manifold-Mixup is a central baseline. The central claim must be revised to state precisely in which settings FAAGC wins, and the text should discuss the cases where it does not.
  2. [Section III-B and Algorithm 1] There is an internal inconsistency about whether the sampling parameter t is resampled or learned. The text states that 'the sampling parameters are treated as a set of learnable parameters of size m' and that optimization updates t, but Algorithm 1, line 2, says 'Sample t ∼ N(0, I)' inside the while loop, and line 4 applies a sigmoid. If t is re-sampled at every iteration, there is no persistent parameter for Adam to update at line 10, and the divergence loss L_Diverg cannot align the learned t with the final augmentation distribution z ∼ U(0,1). If t is initialized once and then updated, line 2 should say 'Initialize t'. The authors must specify the exact training procedure, since the method's reproducibility hinges on this distinction.
  3. [Section III-A and III-B] The geometric premise of the method is that each class's pre-shape features are well approximated by a single geodesic arc, and that sampling points on that arc produces semantically valid class members. This premise is load-bearing but is never tested. The fitting loss L_Sim is minimized, but no diagnostic is reported for the residual geodesic distance after fitting, nor is there any comparison with alternative low-dimensional models such as principal geodesic analysis or multiple arcs. Without such evidence, the uniform sampling along a single fitted arc may generate off-manifold features. I recommend adding a quantitative test that reports, for each dataset and class, the fit quality of the learned geodesic relative to the spread of the training features.
  4. [Section IV-A and IV-B] The hyperparameters β, η_p, and η_t are selected by grid search on the reduced CIFAR-100 training set and then applied uniformly to all datasets and backbones. Since CIFAR-100@5 is also the dataset on which the largest accuracy improvement is reported (1.46 points in Table I), the evaluation does not fully separate hyperparameter choice from method effectiveness. Please add an analysis showing sensitivity across datasets when the hyperparameters are tuned per dataset, or justify that the chosen values are not the source of the reported gains.
minor comments (4)
  1. [Table I] The no-augmentation baselines are reported with ±0.00 standard deviation, despite the statement that 'all later experiments are conducted 6 times with different random seeds.' Please clarify whether these baselines are deterministic or were run once; if they were run multiple times, the zero standard deviation is suspicious.
  2. [Table V] The training times for FAAGC are 39.39 s for 5 samples per class and 39.34 s for 10 samples per class, which are almost identical. This is surprising because the fitting workload should scale with m. Please explain whether the timing includes a fixed overhead or whether the measurements are noisy, and report standard deviations for the runtime.
  3. [Section III-A] The up-dimension procedure duplicates each feature value to form landmark pairs, i.e., v'_i = [(v_i[1], v_i[1]), ..., (v_i[d], v_i[d])]. After centering by the mean of each coordinate, the x and y coordinates of every landmark are identical, so all landmarks lie on the line y=x. This makes the configuration degenerate as a 2D landmark set. Please clarify whether this degeneracy is intentional and how it affects the interpretation of the pre-shape space projection.
  4. [Throughout] There are several typos and grammatical errors, such as 'a object' in Section II-B and 'featires' in Section III-C. A careful proofread is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: FAAGC's geodesic fitting is optimized on training features and evaluated on held-out test data; inherited self-citations are not load-bearing.

full rationale

The derivation chain is self-contained with respect to circularity. FAAGC fits per-class geodesic endpoints (tau_start, tau_end) and sampling parameters t to the training pre-shape features by minimizing L_Sim and L_Diverg (Eqs. 12-15), then generates augmented samples by Eq. (11), and evaluates classifiers on held-out test sets. The fitted quantities are not the evaluated accuracies, so the improvement claim is empirically testable rather than forced by construction. The paper inherits the up-dimension procedure, the z~U(0,1) sampling choice, and the loss-weighting strategy from [10] (same research group), and it cites [9] for the geodesic interpolation formula, but these are preprocessing or inherited training details rather than the target result; no uniqueness theorem is imported and no equation equates the augmentation output to its own evaluation. The appendix's equivalence of Eqs. (7) and (11) is a direct trigonometric identity, not a renaming of the claimed result. Note: Section IV-B's statement that FAAGC 'consistently outperforms all baseline methods' conflicts with Table I (Caltech101@5: FAGC 88.72 vs FAAGC 88.12; CIFAR-100@5 MLP: FAGC 66.28 vs FAAGC 66.23); this is a correctness and evidence-consistency concern, not circularity, and does not change the circularity score.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the standard geometry of Kendall's shape space, plus two unvalidated modeling assumptions: that duplicated CNN features behave like landmarks and that each class is concentrated on a geodesic arc. The free parameters are the loss weight, two learning rates, the classifier loss weights, and the sampling distribution; several were tuned on CIFAR-100.

free parameters (5)
  • beta (geodesic loss weight) = 0.3
    Tuned by grid search on reduced CIFAR-100 (Section IV-A) and applied to all datasets; balances L_Sim and L_Diverg.
  • eta_p (endpoint learning rate) = 0.0003
    Tuned by grid search on reduced CIFAR-100; controls updates to v_start and v_end.
  • eta_t (sampling parameter learning rate) = 0.003
    Tuned by grid search on reduced CIFAR-100; controls updates to t.
  • p_g and lambda (classifier loss weights) = 0.3 and 0.5
    Taken from FAGC [10] and used in Eq. (17) to weight original versus augmented features.
  • sampling distribution z = U(0,1)
    Chosen for consistency with prior work [10] and used both in divergence loss and final augmentation.
assumptions (4)
  • standard math Kendall's pre-shape space geometry: after centering and normalization, features lie on a unit sphere and geodesic distance is arccos of inner product (Eq. 5).
    Used throughout Section II-B to define the geodesic curve and Eq. (11) for interpolation.
  • domain assumption Deep network features, after duplicating each dimension, centering, and normalizing, are valid landmark configurations whose semantic class structure is preserved.
    Invoked in Section III-A, Eq. (9) and Eqs. (2)-(3); no analysis shows CNN features satisfy landmark geometry.
  • ad hoc to paper Each class's feature distribution can be well approximated by a single geodesic arc on the sphere.
    Core modeling premise of Section III-B; the loss fits only two endpoints and one sampling parameter per sample, forcing a one-dimensional curve.
  • domain assumption Sampling parameters z uniformly in [0,1] yields a sensible augmentation scale.
    Adopted in Section III-C for consistency with FAGC [10]; no data-dependent justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FAAGC: Feature Augmentation on Adaptive Geodesic Curve Based on the shape space theory." pith.science (2026). https://pith.science/paper/OYSTNMS4

@misc{pith2026250118619,
  author       = {Pith},
  title        = {Pith review of: FAAGC: Feature Augmentation on Adaptive Geodesic Curve Based on the shape space theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OYSTNMS4}},
  note         = {Machine review of arXiv:2501.18619}
}
read the original abstract

Deep learning models have been widely applied across various domains and industries. However, many fields still face challenges due to limited and insufficient data. This paper proposes a Feature Augmentation on Adaptive Geodesic Curve (FAAGC) method in the pre-shape space to increase data. In the pre-shape space, objects with identical shapes lie on a great circle. Thus, we project deep model representations into the pre-shape space and construct a geodesic curve, i.e., an arc of a great circle, for each class. Feature augmentation is then performed by sampling along these geodesic paths. Extensive experiments demonstrate that FAAGC improves classification accuracy under data-scarce conditions and generalizes well across various feature types.

Figures

Figures reproduced from arXiv: 2501.18619 by the authors.

Figure 1
Figure 1. Workflow for Feature Augmentation on Adaptive Geodesic Curves (FAAGC) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Learnable sampling parameters t allow adaptive alignment between sampled points and original samples, avoiding mismatched pairings and reducing loss. sampled vectors obtained during training are denoted as τ ∗ = [τ ∗ 1 , . . . , τ ∗ i , . . . , τ ∗ m], where τ ∗ i = finterp(τstart, τend, ti). Here, τ ∗ represents a matrix formed by arranging the m sampled vectors as column vectors. To ensure that the newly generated… view at source ↗
Figure 3
Figure 3. Accuracy Gains from the FAAGC module under different image [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Effect of the loss-function weight β on classification accuracy [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Effect of learning rates ηp and ηt on classification accuracy To further simplify the coefficient of τstart, we recall the trigonometric identity sin(θ − s) = sin θ cos s − cos θ sin s. This allows us to rewrite sin θ cos s − cos θ sin s sin θ = sin(θ − s) sin θ . (22)…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 17 canonical work pages

  1. [1]

    A survey on deep learning in medical image analysis,

    G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I. S´anchez, “A survey on deep learning in medical image analysis,”Medical image analysis, vol. 42, pp. 60–88, 2017

  2. [2]

    Machine learning for molecular and materials science,

    K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, “Machine learning for molecular and materials science,”Nature, vol. 559, no. 7715, pp. 547–555, 2018

  3. [3]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,”Advances in neural informa- tion processing systems, vol. 25, 2012

  4. [4]

    A survey on image data augmen- tation for deep learning,

    C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmen- tation for deep learning,”Journal of big data, vol. 6, no. 1, pp. 1–48, 2019

  5. [5]

    Dataset augmentation in feature space,

    T. DeVries and G. W. Taylor, “Dataset augmentation in feature space,” arXiv preprint arXiv:1702.05538, 2017

  6. [6]

    Manifold mixup: Better representations by interpolating hidden states,

    V . Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, D. Lopez-Paz, and Y . Bengio, “Manifold mixup: Better representations by interpolating hidden states,” inInternational conference on machine learning. PMLR, 2019, pp. 6438–6447

  7. [7]

    Feature-level smote: Augmenting fault samples in learnable feature space for imbal- anced fault diagnosis of gas turbines,

    D. Liu, S. Zhong, L. Lin, M. Zhao, X. Fu, and X. Liu, “Feature-level smote: Augmenting fault samples in learnable feature space for imbal- anced fault diagnosis of gas turbines,”Expert Systems with Applications, vol. 238, p. 122023, 2024

  8. [8]

    Shape manifolds, procrustean metrics, and complex projective spaces,

    D. G. Kendall, “Shape manifolds, procrustean metrics, and complex projective spaces,”Bulletin of the London mathematical society, vol. 16, no. 2, pp. 81–121, 1984

Show all 36 references
  1. [9]

    Recognition of multiple configurations of objects with limited data,

    Y . Han, B. Wang, M. Idesawa, and H. Shimai, “Recognition of multiple configurations of objects with limited data,”Pattern Recognition, vol. 43, no. 4, pp. 1467–1475, 2010

  2. [10]

    Gcfa: Geodesic curve feature augmen- tation via shape space theory,

    Y . Han, G. Wan, and B. Wang, “Gcfa: Geodesic curve feature augmen- tation via shape space theory,”arXiv preprint arXiv:2312.03325, 2023

  3. [11]

    Autoaug- ment: Learning augmentation strategies from data,

    E. D. Cubuk, B. Zoph, D. Mane, V . Vasudevan, and Q. V . Le, “Autoaug- ment: Learning augmentation strategies from data,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 113–123

  4. [12]

    Randaugment: Practical automated data augmentation with a reduced search space,

    E. D. Cubuk, B. Zoph, J. Shlens, and Q. V . Le, “Randaugment: Practical automated data augmentation with a reduced search space,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 702–703

  5. [13]

    Augmix: A simple data processing method to improve robustness and uncertainty,

    D. Hendrycks, N. Mu, E. D. Cubuk, B. Zoph, J. Gilmer, and B. Lak- shminarayanan, “Augmix: A simple data processing method to improve robustness and uncertainty,”arXiv preprint arXiv:1912.02781, 2019

  6. [14]

    Synthetic data augmentation using gan for improved liver lesion clas- sification,

    M. Frid-Adar, E. Klang, M. Amitai, J. Goldberger, and H. Greenspan, “Synthetic data augmentation using gan for improved liver lesion clas- sification,” in2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). IEEE, 2018, pp. 289–293

  7. [15]

    Training generative adversarial networks with limited data,

    T. Karras, M. Aittala, J. Hellsten, S. Laine, J. Lehtinen, and T. Aila, “Training generative adversarial networks with limited data,”Advances in neural information processing systems, vol. 33, pp. 12 104–12 114, 2020

  8. [16]

    Domain-guided data augmentation for deep learning on medical imaging,

    C. Athalye and R. Arnaout, “Domain-guided data augmentation for deep learning on medical imaging,”PloS one, vol. 18, no. 3, p. e0282532, 2023

  9. [17]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,”arXiv preprint arXiv:1412.6572, 2014

  10. [18]

    Auto-encoding variational bayes,

    D. P. Kingma, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

  11. [19]

    A simple feature augmentation for domain generalization,

    P. Li, D. Li, W. Li, S. Gong, Y . Fu, and T. M. Hospedales, “A simple feature augmentation for domain generalization,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 8886–8895

  12. [20]

    On feature normalization and data augmentation,

    B. Li, F. Wu, S.-N. Lim, S. Belongie, and K. Q. Weinberger, “On feature normalization and data augmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 383–12 392

  13. [21]

    Data augmentation with variational autoencoders and manifold sampling,

    C. Chadebec and S. Allassonni `ere, “Data augmentation with variational autoencoders and manifold sampling,” inDeep Generative Models, 12 and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Con- junction ...

  14. [22]

    Feature space augmentation for long-tailed data,

    P. Chu, X. Bian, S. Liu, and H. Ling, “Feature space augmentation for long-tailed data,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX

  15. [23]

    Springer, 2020, pp. 694–710

  16. [24]

    D. G. Kendall, D. Barden, T. K. Carne, and H. Le,Shape and shape theory. John Wiley & Sons, 2009

  17. [25]

    Kendall shape-vae: Learning shapes in a generative framework,

    S. Vadgama, J. M. Tomczak, and E. J. Bekkers, “Kendall shape-vae: Learning shapes in a generative framework,” inNeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations, 2022

  18. [26]

    Glass: Geometric latent augmentation for shape spaces,

    S. Muralikrishnan, S. Chaudhuri, N. Aigerman, V . G. Kim, M. Fisher, and N. J. Mitra, “Glass: Geometric latent augmentation for shape spaces,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 18 552–18 561

  19. [27]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,”ICLR, 2021

  20. [28]

    Learning with a wasserstein loss,

    C. Frogner, C. Zhang, H. Mobahi, M. Araya, and T. A. Poggio, “Learning with a wasserstein loss,”Advances in neural information processing systems, vol. 28, 2015

  21. [29]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

  22. [30]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

  23. [31]

    The caltech-ucsd birds-200-2011 dataset,

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The caltech-ucsd birds-200-2011 dataset,” 2011

  24. [32]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”arXiv preprint arXiv:1708.07747, 2017

  25. [33]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,

    L. Fei-Fei, R. Fergus, and P. Perona, “Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,” in2004 conference on computer vision and pattern recognition workshop. IEEE, 2004, pp. 178–178

  26. [34]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

  27. [35]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

  28. [36]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

Pith tools

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