Pith. sign in

REVIEW 3 major objections 6 minor 87 references

Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A hierarchy-balanced hyperbolic embedding layer outperforms Euclidean classifiers for out-of-distribution detection.

desk verdict A solid empirical case for hierarchy-aware hyperbolic prototypes in OOD detection, but the paper overclaims 'all 13' and lacks the matched Euclidean control needed to pin the gains on geometry. read the letter →

arxiv 2506.10146 v1 pith:VHJY3TYS submitted 2025-06-11 cs.LG cs.CV

Tejaswi Kasarla , Max van Spengler , Pascal Mettes This is my paper
classification cs.LGcs.CV
keywords HyperboliclearningOut-of-distributiondetectionPoincaréballHierarchicalembeddingsPrototype-basedclassificationNormbalancingOpenOODbenchmarkDistortionloss
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 asks whether the geometry of the last layer, rather than a new scoring rule, can make out-of-distribution (OOD) detection work better. Its answer is yes: when in-distribution classes become prototypes in the Poincaré ball, placed so that hyperbolic distances mirror a known class hierarchy and every node at the same hierarchy level sits at the same distance from the origin, unknown samples naturally land in low-norm regions near the origin while confident in-distribution samples sit near the boundary. The authors call this Balanced Hyperbolic Learning, and they show that standard OOD scoring functions—MSP, Energy, ODIN, KNN, Gram, and others—carry over by replacing logits with negative geodesic distances to the hyperbolic prototypes. In their experiments across 13 scoring functions and two in-distribution datasets with OpenOOD splits, the hyperbolic final layer consistently improves FPR@95 and AUROC over the Euclidean counterpart trained with the same backbone, and it outperforms prior hyperbolic and hyperspherical prototype methods. The practical payoff, if the claim holds, is that upgrading a detector can be as simple as swapping the classifier head and supplying a class hierarchy, with no extra outlier data and no new scoring function.

What carries the argument

The load-bearing object is the Balanced Hyperbolic Embedding: class prototypes in a 64-dimensional Poincaré ball learned by Riemannian SGD under a distortion loss $L_d = (d_B(p_i,p_j) - d_G(v_i,v_j))/d_G(v_i,v_j)$, which makes hyperbolic distances track Dijkstra graph distances in the class hierarchy, plus a norm-balancing loss $L_n = \frac{1}{n}\sum_{\ell}\sum_{i \in \ell}(\|p_i\| - m_\ell)^2$ that equalizes distances from the origin across nodes at the same hierarchy level. The second component is a hyperbolic distance-based cross-entropy loss, $\mathcal{L} = -\frac{1}{N}\sum_n \sum_k \log \frac{\exp(-d_B(z_{n,k},p_k))}{\sum_i \exp(-d_B(z_{n,i},p_i))}$, which trains the image encoder to align with the fixed prototypes. Together these turn the final layer into a set of geodesic distances, so any OOD scoring function that consumes logits, features, or probabilities can be applied directly to hyperbolic distances.

What would settle it

On a fixed OOD benchmark with unchanged training data, corrupt the hierarchy systematically—randomly permuting leaf labels, pruning levels, or substituting a hierarchy whose graph distances are known to disagree with human similarity judgments—and test whether the reported FPR@95 and AUROC gains over the Euclidean baseline shrink or reverse. A second decisive experiment is to use a class set with no meaningful hierarchy and compare hyperbolic prototypes against Euclidean ones; the paper's own assumption predicts the advantage should disappear.

Watch

Extended reading notes

Core claim

The paper's central claim is that a good hierarchical hyperbolic embedding is intrinsically better than a Euclidean classifier for telling in-distribution from out-of-distribution samples. The mechanism is the Poincaré norm: training the final layer so that image embeddings sit near hierarchy-defined prototypes pushes confident in-distribution samples toward the boundary of the ball and leaves unrelated OOD samples near the origin, so distances and softmax probabilities become informative OOD scores. Balanced Hyperbolic Learning makes this work in practice by fixing the prototype bias of existing hyperbolic embeddings: a distortion loss matches pairwise hyperbolic distances to undirected graph distances computed by Dijkstra's algorithm, and a norm-balancing loss keeps nodes at the same hierarchy level equidistant from the origin, preventing shallow or narrow subtrees from collapsing inward. The paper reports that all 13 tested scoring functions improve over their Euclidean baselines, that the approach beats prior hyperbolic and hyperspherical prototype methods, and that it natively supports hierarchical OOD generalization by identifying the most related in-distribution class for unknown samples.

Load-bearing premise

The method assumes that a correct and complete class hierarchy for the in-distribution classes is available and that undirected graph distance in that hierarchy faithfully reflects semantic similarity between classes, so that prototype placement encodes real relationships.

Editorial extensions

If this is right

  • Any existing OOD scoring function that consumes logits, features, or probabilities can be applied on top of hyperbolic prototypes, and the paper demonstrates consistent gains for 13 such functions without new outlier data.
  • A correct class hierarchy becomes a usable input signal for OOD detection, so datasets with existing taxonomies can be exploited without extra training data or scoring rules.
  • OOD samples can be assigned to their closest in-distribution class in the hierarchy more accurately than with Euclidean baselines, enabling hierarchical OOD generalization rather than a bare binary flag.
  • Because only the final layer and loss change, the approach transfers across backbones, embedding dimensions, and future scoring functions.
  • Swapping the final Euclidean layer for balanced hyperbolic prototypes is enough to push methods like maximum softmax probability from weak to competitive detectors, lowering FPR@95 by roughly 9 points on CIFAR-100.

Reading between the lines

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

  • The norm-separation plots suggest a norm-only detector may be nearly sufficient—one could threshold on the Poincaré norm alone—although the paper does not itself propose or evaluate such a detector.
  • The same prototype placement could be reused in few-shot, open-set, and zero-shot settings, where hierarchy-aware spacing is also load-bearing; that would be a direct application of the balancing loss beyond OOD detection.
  • Because the paper allows LLM-generated hierarchies as a source, the dependence on hierarchy quality is testable: a sweep that adds noise or edge-removals to the hierarchy would quantify how much semantic fidelity the OOD gains require.
  • The hierarchy need only describe the in-distribution classes, so the method could be combined with outlier-exposure training strategies; the paper deliberately excludes those, so the combination is an open question.
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

3 major / 6 minor

Summary. The paper introduces Balanced Hyperbolic Learning, a method that learns class prototypes in the Poincaré ball from a given class hierarchy using a distortion loss with norm balancing, then trains a backbone so that image embeddings align with these prototypes via a hyperbolic distance-based cross-entropy loss. The authors show how 13 established OOD scoring functions can be generalized to operate with hyperbolic prototypes, and report experiments on CIFAR-100 and ImageNet-100 under the OpenOOD protocol, claiming consistent improvements over Euclidean baselines, superiority over other hyperbolic methods and over prototype-based methods such as CIDER and PALM, and hierarchical OOD generalization.

Significance. If the claims hold, the work would provide a simple, generally applicable recipe: swap the final linear layer for hierarchy-derived hyperbolic prototypes and thereby improve a wide range of OOD detectors without altering the backbone or the scoring function. The paper's strengths include the breadth of the empirical study (13 scoring functions, multiple datasets, backbones, ablations, and visualizations), the explicit statement of the hierarchy assumption as a limitation, and the falsifiable nature of the central claim. However, the significance is currently tempered by the calibration confound and by counterexamples in the reported tables.

major comments (3)
  1. [§6.1.1, Table 2] The claim that all 13 scoring functions benefit from hyperbolic embeddings is contradicted by Table 2 on ImageNet-100: KNN FPR@95 worsens from 37.13 to 45.74, SCALE AUROC drops from 88.14 to 86.21, and GEN n-AUROC decreases from 81.04 to 80.24. The text states that 'our hyperbolic embeddings have a positive effect on all 13 scoring functions' and that 'the same outcome' holds for Table 2; this is inaccurate. Please revise the claim to reflect the actual pattern and discuss the exceptions.
  2. [§4.3 and Fig. 2] The comparison between the Euclidean baseline and the hyperbolic method is confounded by calibration hyperparameters: the hyperbolic method uses energy temperature T=10 and logit temperature γ=10, with prototypes scaled by 0.95, while the baseline uses T=1 and no analogous scaling. The ablation in Fig. 2 includes a non-hierarchical hyperbolic condition with one-hot prototypes, but it lacks a Euclidean fixed-prototype head with matched scale and temperature. As a result, the observed improvements could be due to better calibration of a fixed-prototype logit scale rather than to hyperbolic geometry, which is load-bearing for the abstract's claim that 'a good hierarchical hyperbolic embedding is preferred.' Please add a Euclidean prototype control with matched hyperparameters (or a temperature/scale sweep for both) to isolate the geometric effect.
  3. [Abstract and §7] The abstract claims 'consistent, significant improvement' over Euclidean counterparts, but no statistical significance test is reported and the averages in Tables 1 and 2 lack error bars, while Table 2 contains negative results. With only three training runs per method, some reported differences may be within noise. Please either add significance testing (e.g., paired tests across datasets or confidence intervals) or soften the language to 'consistent' only for the metrics and datasets where the direction is uniform.
minor comments (6)
  1. [§4.2, Eq. (9) and Algorithm 1] Equation (9) defines the norm loss as a sum of (p_l_i - m_l) without squaring, whereas Algorithm 1 line 6 uses a squared term; please align the equation with the algorithm.
  2. [§3.2 and Algorithm 1] The curvature parameter is set to c = -1 in Algorithm 1 and Section 4.2, but Table 7 reports c = 1 as the best setting and the Poincaré ball definition requires 1/c > 0; please correct the sign convention to avoid ambiguity.
  3. [Throughout] There are several typos: 'curcature' in §6.4.3, 'maixmum' in §6.1.2, 'distrbution' in §6.5, 'hierachical' in §6.3.3, and 'out-of-distrbution' in the Fig. 6 caption; please proofread.
  4. [Fig. 1] The caption contains incomplete sentences, e.g., 'The represents classifiers in Euclidean space' and 'The represents image embeddings'; please fix the wording.
  5. [§6.1.1] The statement that the method uses 'the exact same backbone and training procedure' as the baseline is not literally accurate, since the hyperbolic method introduces additional hyperparameters (logit temperature γ, prototype scaling, and energy temperature T); please rephrase to 'same backbone and training schedule'.
  6. [§4.1 and Limitations] The text says hierarchical information 'typically comes for free' but the Limitations section states that a correct and known hierarchy is assumed; please reconcile these statements.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity: the OOD claim is empirical; only the reported distortion metric is the training objective itself.

  1. self definitional [Section 4.2 (Eq. 7, Algorithm 1) and Section 6.1.2 (Table 3, left)]
    "We do so by directly minimizing the distortion (Sala et al. 2018) between the hyperbolic and graph distances. ... These values reveal that the baseline embeddings yield a much higher hierarchical distortion than our approach and are actually not well suited for standard classification."

    The 'Dist.' column in Table 3 is exactly the quantity minimized by Ld in Eq. 7: the average relative deviation between the hyperbolic pairwise distance dB(pi,pj) and the graph distance dG(vi,vj). Reporting that the proposed method has the lowest distortion is therefore a restatement of the optimization objective, true by construction, rather than independent evidence of embedding quality. This step is non-central: the paper's load-bearing OOD conclusion rests on held-out FPR/AUROC/AUPR numbers in the same tables, not on the distortion column.

full rationale

The central claim that balanced hyperbolic prototypes improve OOD detection is an empirical result, not a derivation: the paper fixes hierarchy-derived prototypes, trains a backbone with a distance-based cross-entropy loss, and then applies standard scoring functions to held-out OOD data. No predicted quantity is defined in terms of the fitted prototypes or temperatures in a way that forces the reported improvements. The energy-temperature choice (T=10 for the proposed method, T=1 for the baseline) is a per-method calibration choice that could affect the energy comparison, but it is a hyperparameter, not a built-in equivalence. Self-citations, such as Poincaré ResNet, appear as baselines or motivating references, not as load-bearing uniqueness or ansatz-importing arguments. The only true by-construction equivalence is the distortion metric reported in Table 3, which is the training loss of Eq. 7; this is a minor, non-central tautology. The acknowledged limitation of assuming a correct and known hierarchy is an assumption, not circularity.

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

The method relies on a supplied hierarchy, the assumption that graph distance matches semantic distance, and the hypothesis that OOD samples cluster near the origin or between ID clusters. All three are domain assumptions rather than derived facts. The main free parameters are the curvature, embedding dimension, prototype scaling, norm-loss weight, and the energy temperature, several of which were tuned on the evaluation sets.

free parameters (6)
  • curvature c = c = 1 (best in ablation)
    Selected by ablating c in {0.5, 0.75, 1, 1.5, 2.0} on CIFAR-100 with MSP; c=1 gives best OOD results (Table 7).
  • energy temperature T = T = 10 for hyperbolic, T = 1 for Euclidean baseline
    Chosen because 'these are the best performing settings for both' (Section 4.3), i.e., selected on the evaluation data.
  • logit temperature gamma = gamma = 10
    Multiplier on hyperbolic distances before softmax (Section 5, implementation details); fixed but arbitrary.
  • prototype scaling factor = 0.95
    Scales hyperbolic prototypes to avoid boundary issues (Section 5).
  • norm-loss weight tau = 0.01 for 2-level, 0.1 for deeper hierarchies
    Empirically set based on hierarchy depth (Section 4.2).
  • embedding dimension d = 64
    Set following literature; ablation shows stability across dimensions (Table 6).
assumptions (4)
  • domain assumption A correct, known hierarchy for the in-distribution classes is available.
    Stated in Section 4.1 and Limitations; if false, prototypes are meaningless.
  • domain assumption Undirected graph distance in the hierarchy is a faithful proxy for semantic class distance.
    Used in Section 4.2 as supervision for the distortion loss; no verification on OOD semantics.
  • domain assumption OOD samples tend to lie near the origin or between ID clusters in hyperbolic space.
    Central hypothesis illustrated in Figure 1b and used to justify the method; not proven.
  • standard math Poincare ball model formulas and exponential map are valid.
    Standard definitions in Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors." pith.science (2026). https://pith.science/paper/VHJY3TYS

@misc{pith2026250610146,
  author       = {Pith},
  title        = {Pith review of: Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VHJY3TYS}},
  note         = {Machine review of arXiv:2506.10146}
}
read the original abstract

Out-of-distribution recognition forms an important and well-studied problem in deep learning, with the goal to filter out samples that do not belong to the distribution on which a network has been trained. The conclusion of this paper is simple: a good hierarchical hyperbolic embedding is preferred for discriminating in- and out-of-distribution samples. We introduce Balanced Hyperbolic Learning. We outline a hyperbolic class embedding algorithm that jointly optimizes for hierarchical distortion and balancing between shallow and wide subhierarchies. We then use the class embeddings as hyperbolic prototypes for classification on in-distribution data. We outline how to generalize existing out-of-distribution scoring functions to operate with hyperbolic prototypes. Empirical evaluations across 13 datasets and 13 scoring functions show that our hyperbolic embeddings outperform existing out-of-distribution approaches when trained on the same data with the same backbones. We also show that our hyperbolic embeddings outperform other hyperbolic approaches, beat state-of-the-art contrastive methods, and natively enable hierarchical out-of-distribution generalization.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

87 extracted references · 71 canonical work pages

  1. [1]

    In: International Conference on Learning Representations (2018)

    Becigneul, G., Ganea, O.-E.: Riemannian adap- tive optimization methods. In: International Conference on Learning Representations (2018)

    2018

  2. [2]

    In: ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models (2023)

    Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing imagenet out-of-distribution detec- tion evaluation. In: ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models (2023)

    2023

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Bertinetto, L., Mueller, R., Tertikas, K., Saman- gooei, S., Lord, N.A.: Making better mistakes: Leveraging class hierarchies with deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12506–12515 (2020)

    2020

  4. [4]

    In: International Conference on Machine Learning (2020)

    Hamilton, W.: Latent variable modelling with hyperbolic normalizing flows. In: International Conference on Machine Learning (2020)

    2020

  5. [5]

    Flavors of geometry 31(59-115), 2 (1997)

    Cannon, J.W., Floyd, W.J., Kenyon, R., Parry, W.R., et al.: Hyperbolic geometry. Flavors of geometry 31(59-115), 2 (1997)

    1997

  6. [6]

    Journal of the American Statisti- cal Association 112(520), 1684–1696 (2017)

    Chen, Y.-C., Genovese, C.R., Wasserman, L.: Density level sets: Asymptotics, inference, and visualization. Journal of the American Statisti- cal Association 112(520), 1684–1696 (2017)

    2017

  7. [7]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3606–3613 (2014)

    2014

  8. [8]

    IEEE Transactions on Multimedia (2023)

    Cui, Y., Yu, Z., Peng, W., Tian, Q., Liu, L.: Rethinking few-shot class-incremental learning with open-set hypothesis in hyperbolic geome- try. IEEE Transactions on Multimedia (2023)

    work page 2023

Show all 87 references
  1. [9]

    In: The Eleventh Inter- national Conference on Learning Representa- tions (2022)

    Djurisic, A., Bozanic, N., Ashok, A., Liu, R.: Extremely simple activation shaping for out-of- distribution detection. In: The Eleventh Inter- national Conference on Learning Representa- tions (2022)

    2022

  2. [10]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp

    Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255 (2009). Ieee

    2009

  3. [11]

    IEEE signal processing magazine 29(6), 141–142 (2012)

    Deng, L.: The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine 29(6), 141–142 (2012)

    2012

  4. [12]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Dengxiong, X., Kong, Y.: Ancestor search: Gener- alized open set recognition via hyperbolic side information learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 4003–4012 (2023)

    2023

  5. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp

    Greeff, M., Krause, A.: Hierarchical image clas- sification using entailment cone embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 836–837 (2020)

    2020

  6. [14]

    In: International Conference on Machine Learning, pp

    Desai, K., Nickel, M., Rajpurohit, T., John- son, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning, pp. 7694–7731 (2023). PMLR

    2023

  7. [15]

    In: Computer Vision and Pattern Recognition (2022)

    Sebe, N., Oseledets, I.: Hyperbolic vision trans- formers: Combining improvements in metric learning. In: Computer Vision and Pattern Recognition (2022)

    2022

  8. [16]

    In: International Conference on Learning Representations (2023) Ghadimi Atigh, M., Keller-Ressel, M., Mettes, P.: Hyperbolic busemann learning with ideal pro- totypes

    Franco, L., Mandica, P., Munjal, B., Galasso, F.: Hyperbolic self-paced learning for self- supervised skeleton-based action representa- tions. In: International Conference on Learning Representations (2023) Ghadimi Atigh, M., Keller-Ressel, M., Mettes, P.: Hyperbolic busemann ...

    2023

  9. [17]

    In: International Conference on Machine Learning, pp

    Ganea, O., B´ ecigneul, G., Hofmann, T.: Hyper- bolic entailment cones for learning hierarchi- cal embeddings. In: International Conference on Machine Learning, pp. 1646–1655 (2018). PMLR

    2018

  10. [18]

    IEEE trans- actions on pattern analysis and machine intelli- gence 43(10), 3614–3631 (2020)

    Geng, C., Huang, S.-j., Chen, S.: Recent advances in open set recognition: A survey. IEEE trans- actions on pattern analysis and machine intelli- gence 43(10), 3614–3631 (2020)

    2020

  11. [19]

    In: International Conference on Machine Learning, pp

    Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning, pp. 1321–1330 (2017). PMLR

    2017

  12. [20]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recog- nition, pp

    Guo, Y., Wang, X., Chen, Y., Yu, S.X.: Clipped hyperbolic classifiers are super-hyperbolic clas- sifiers. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recog- nition, pp. 11–20 (2022)

    2022

  13. [21]

    In: International Conference on Computer Vision (2021)

    Gao, Z., Wu, Y., Jia, Y., Harandi, M.: Curvature generation in curved spaces for few-shot learn- ing. In: International Conference on Computer Vision (2021)

    2021

  14. [22]

    In: International Confer- ence on Machine Learning (2022)

    Song, D.: Scaling out-of-distribution detection for real-world settings. In: International Confer- ence on Machine Learning (2022). PMLR

    2022

  15. [23]

    IEEE Transac- tions on Neural Networks and Learning Systems (2023)

    Hong, J., Fang, P., Li, W., Han, J., Petersson, L., Harandi, M.: Curved geometric networks for visual anomaly recognition. IEEE Transac- tions on Neural Networks and Learning Systems (2023)

    2023

  16. [24]

    In: International Con- ference on Learning Representations (2016)

    Hendrycks, D., Gimpel, K.: A baseline for detect- ing misclassified and out-of-distribution exam- ples in neural networks. In: International Con- ference on Learning Representations (2016)

    2016

  17. [25]

    In: Advances in Neural Information Processing Systems (2021)

    Hsu, J., Gu, J., Wu, G., Chiu, W., Yeung, S.: Capturing implicit hierarchical structure in 3d biomedical images with self-supervised hyper- bolic representations. In: Advances in Neural Information Processing Systems (2021)

    2021

  18. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Hsu, Y.-C., Shen, Y., Jin, H., Kira, Z.: General- ized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10951–10960 (2020)

    2020

  19. [27]

    In: Proceed- ings of the IEEE International Conference on Computer Vision, pp

    He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level perfor- mance on imagenet classification. In: Proceed- ings of the IEEE International Conference on Computer Vision, pp. 1026–1034 (2015)

    2015

  20. [28]

    Krizhevsky, A., Hinton, G., et al.: Learning mul- tiple layers of features from tiny images (2009)

    2009

  21. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Kim, S., Jeong, B., Kwak, S.: Hier: Metric learn- ing beyond class labels via hierarchical regu- larization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19903–19912 (2023)

    2023

  22. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Oseledets, I., Lempitsky, V.: Hyperbolic image embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6418–6428 (2020)

    2020

  23. [31]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Kong, S., Ramanan, D.: Opengan: Open-set recog- nition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 813–822 (2021)

    2021

  24. [32]

    In: Computer Vision and Pattern Recognition (2020)

    Liu, S., Chen, J., Pan, L., Ngo, C.-W., Chua, T.- S., Jiang, Y.-G.: Hyperbolic visual embedding learning for zero-shot recognition. In: Computer Vision and Pattern Recognition (2020)

    2020

  25. [33]

    In: The Twelfth International Conference on Learning Representations (2024)

    Moore, K.: Learning with mixture of proto- types for out-of-distribution detection. In: The Twelfth International Conference on Learning Representations (2024). https://openreview. net/forum?id=uNkKaD3MCs

    2024

  26. [34]

    In: Interna- tional Conference on Learning Representations (2016)

    Loshchilov, I., Hutter, F.: Sgdr: Stochastic gra- dient descent with warm restarts. In: Interna- tional Conference on Learning Representations (2016)

    2016

  27. [35]

    21 Advances in neural information processing sys- tems 31 (2018)

    Lee, K., Lee, K., Lee, H., Shin, J.: A sim- ple unified framework for detecting out-of- distribution samples and adversarial attacks. 21 Advances in neural information processing sys- tems 31 (2018)

    2018

  28. [36]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Lee, H.: Hierarchical novelty detection for visual object recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1034–1042 (2018)

    2018

  29. [37]

    In: International Conference on Learning Representations (2018)

    Liang, S., Li, Y., Srikant, R.: Enhancing the reli- ability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations (2018)

    2018

  30. [38]

    In: International Conference on Machine Learn- ing, pp

    Law, M., Liao, R., Snell, J., Zemel, R.: Lorentzian distance learning for hyperbolic representations. In: International Conference on Machine Learn- ing, pp. 3672–3681 (2019). PMLR

    2019

  31. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 23946–23955 (2023)

    2023

  32. [40]

    In: Com- puter Vision and Pattern Recognition (2020)

    Long, T., Mettes, P., Shen, H.T., Snoek, C.G.M.: Searching for actions on the hyperbole. In: Com- puter Vision and Pattern Recognition (2020)

    2020

  33. [41]

    arXiv preprint arXiv:2108.13624 (2021)

    Cui, P.: Towards out-of-distribution generaliza- tion: A survey. arXiv preprint arXiv:2108.13624 (2021)

    2021 arXiv

  34. [42]

    Advances in Neural Information Processing Systems (2020)

    Liu, W., Wang, X., Owens, J., Li, Y.: Energy- based out-of-distribution detection. Advances in Neural Information Processing Systems (2020)

    2020

  35. [43]

    Stanford University (2015)

    Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. Stanford University (2015)

    2015

  36. [44]

    arXiv preprint arXiv:2402.15754 (2024)

    Liu, Y., Yang, T., Huang, S., Zhang, Z., Huang, H., Wei, F., Deng, W., Sun, F., Zhang, Q.: Hd-eval: Aligning large language model evalu- ators through hierarchical criteria decomposi- tion. arXiv preprint arXiv:2402.15754 (2024)

    2024 arXiv

  37. [45]

    In: Conference on Lifelong Learning Agents, pp

    Linderman, R., Zhang, J., Inkawhich, N., Li, H., Chen, Y.: Fine-grain inference on out- of-distribution data with hierarchical classifi- cation. In: Conference on Lifelong Learning Agents, pp. 162–183 (2023). PMLR

    2023

  38. [46]

    arXiv (2022)

    Li, L., Zhang, Y., Wang, S.: The euclidean space is evil: Hyperbolic attribute editing for few-shot image generation. arXiv (2022)

    2022

  39. [47]

    arXiv preprint arXiv:2305.06611 (2023)

    Mettes, P., Atigh, M.G., Keller-Ressel, M., Gu, J., Yeung, S.: Hyperbolic deep learning in computer vision: A survey. arXiv preprint arXiv:2305.06611 (2023)

    2023 arXiv

  40. [48]

    In: AAAI Conference on Artificial Intelligence (2022)

    Ma, R., Fang, P., Drummond, T., Harandi, M.: Adaptive poincar´ e point to set distance for few-shot classification. In: AAAI Conference on Artificial Intelligence (2022)

    2022

  41. [49]

    Communications of the ACM 38(11), 39–41 (1995)

    Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39–41 (1995)

    1995

  42. [50]

    In: Advances in Neural Information Processing Systems, vol

    Mathieu, E., Le Lan, C., Maddison, C.J., Tomioka, R., Teh, Y.W.: Continuous hierarchical rep- resentations with poincar´ e variational auto- encoders. In: Advances in Neural Information Processing Systems, vol. 32 (2019)

    2019

  43. [51]

    Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450 (2022)

    2022 arXiv

  44. [52]

    https://openreview.net/forum? id=aEFaE0W5pAd

    Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of- distribution detection? In: The Eleventh Inter- national Conference on Learning Represen- tations (2023). https://openreview.net/forum? id=aEFaE0W5pAd

    2023

  45. [53]

    Advances in neural information processing systems 30 (2017)

    Nickel, M., Kiela, D.: Poincar´ e embeddings for learning hierarchical representations. Advances in neural information processing systems 30 (2017)

    2017

  46. [54]

    In: International Conference on Machine Learning, pp

    Nickel, M., Kiela, D.: Learning continuous hierar- chies in the lorentz model of hyperbolic geom- etry. In: International Conference on Machine Learning, pp. 3779–3788 (2018). PMLR

    2018

  47. [55]

    In: International Conference on Machine Learning (2019) 22

    Nagano, Y., Yamaguchi, S., Fujita, Y., Koyama, M.: A wrapped normal distribution on hyper- bolic space for gradient-based learning. In: International Conference on Machine Learning (2019) 22

    2019

  48. [56]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Park, J., Jung, Y.G., Teoh, A.B.J.: Nearest neigh- bor guidance for out-of-distribution detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1686–1695 (2023)

    2023

  49. [57]

    IEEE Transactions on pattern analysis and machine intelligence 44(12), 10023–10044 (2021)

    Zhao, G.: Hyperbolic deep neural networks: A survey. IEEE Transactions on pattern analysis and machine intelligence 44(12), 10023–10044 (2021)

    2021

  50. [58]

    In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp

    Ryu, S., Koo, S., Yu, H., Lee, G.G.: Out-of-domain detection based on generative adversarial net- work. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 714–718 (2018)

    2018

  51. [59]

    University of Pennsyl- vania, ??? (2005)

    Schuler, K.K.: VerbNet: A Broad-coverage, Com- prehensive Verb Lexicon. University of Pennsyl- vania, ??? (2005)

    2005

  52. [60]

    arXiv preprint arXiv:2103.12051 (2021)

    Sehwag, V., Chiang, M., Mittal, P.: Ssd: A unified framework for self-supervised outlier detection. arXiv preprint arXiv:2103.12051 (2021)

    2021 arXiv

  53. [61]

    In: Inter- national Conference on Machine Learning, pp

    Sala, F., De Sa, C., Gu, A., R´ e, C.: Representation tradeoffs for hyperbolic embeddings. In: Inter- national Conference on Machine Learning, pp. 4460–4469 (2018). PMLR

    2018

  54. [62]

    In: Euro- pean Conference on Computer Vision (2022)

    Sun, Y., Li, Y.: Dice: Leveraging sparsifica- tion for out-of-distribution detection. In: Euro- pean Conference on Computer Vision (2022). Springer Sur ´ ıs, D., Liu, R., Vondrick, C.: Learning the pre- dictability of the future. In: Computer Vision and Pattern Recognition (2021)

    2022

  55. [63]

    In: International Conference on Machine Learning, pp

    Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of- distribution detection with deep nearest neigh- bors. In: International Conference on Machine Learning, pp. 20827–20840 (2022). PMLR

    2022

  56. [64]

    In: International Conference on Machine Learning, pp

    Sastry, C.S., Oore, S.: Detecting out-of- distribution examples with gram matrices. In: International Conference on Machine Learning, pp. 8491–8501 (2020). PMLR

    2020

  57. [65]

    Advances in Neural Information Process- ing Systems 35 (2022)

    Song, Y., Sebe, N., Wang, W.: Rankfeat: Rank-1 feature removal for out-of-distribution detec- tion. Advances in Neural Information Process- ing Systems 35 (2022)

    2022

  58. [66]

    In: The Eleventh International Conference on Learning Representations (2022)

    Tao, L., Du, X., Zhu, J., Li, Y.: Non-parametric outlier synthesis. In: The Eleventh International Conference on Learning Representations (2022)

    2022

  59. [67]

    Springer, ??? (2022) Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y.,

    Ungar, A.: A Gyrovector Space Approach to Hyperbolic Geometry. Springer, ??? (2022) Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y.,

    2022

  60. [68]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Belongie, S.: The inaturalist species classifica- tion and detection dataset. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 8769–8778 (2018)

    2018

  61. [69]

    In: International Conference on Learning Representations (2021)

    Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classi- fier is all you need. In: International Conference on Learning Representations (2021)

    2021

  62. [70]

    In: ICCV (2023)

    Spengler, M., Berkhout, E., Mettes, P.: Poincar´ e resnet. In: ICCV (2023)

    2023

  63. [71]

    arXiv preprint arXiv:2007.05566 (2020)

    Winkens, J., Bunel, R., Roy, A.G., Stanforth, R., Natarajan, V., Ledsam, J.R., MacWilliams, P., Kohli, P., Karthikesalingam, A., Kohl, S., et al.: Contrastive training for improved out-of-distribution detection. arXiv preprint arXiv:2007.05566 (2020)

    2020 arXiv

  64. [72]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out- of-distribution with virtual-logit matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4921–4930 (2022)

    2022

  65. [73]

    Wei, H., Xie, R., Cheng, H., Feng, L., An, B., Li, Y.: Mitigating neural network overconfidence with logit normalization (2022)

    2022

  66. [74]

    arXiv preprint arXiv:2310.00227 (2023)

    Xu, K., Chen, R., Franchi, G., Yao, A.: Scal- ing for training time and post-hoc out- of-distribution detection enhancement. arXiv preprint arXiv:2310.00227 (2023)

    2023 arXiv

  67. [75]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, 23 pp

    Yu, Q., Aizawa, K.: Unsupervised out-of- distribution detection by maximum classifier discrepancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, 23 pp. 9518–9526 (2019)

    2019

  68. [76]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Yan, J., Luo, L., Deng, C., Huang, H.: Unsuper- vised hyperbolic metric learning. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12465– 12474 (2021)

    2021

  69. [77]

    In: International Conference on Medical Image Computing and Computer-Assisted Interven- tion (2022)

    Yu, Z., Nguyen, T., Gal, Y., Ju, L., Chandra, S.S., Zhang, L., Bonnington, P., Mar, V., Wang, Z., Ge, Z.: Skin lesion recognition with class- hierarchy regularized hyperbolic embeddings. In: International Conference on Medical Image Computing and Computer-Assisted Interven- ti...

    2022

  70. [78]

    In: Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)

    Yuval, N.: Reading digits in natural images with unsupervised feature learning. In: Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)

    2011

  71. [79]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Yang, J., Wang, H., Feng, L., Yan, X., Zheng, H., Zhang, W., Liu, Z.: Semantically coherent out-of-distribution detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 8301–8309 (2021)

    2021

  72. [80]

    Advances in Neural Information Processing Systems 35, 32598–32611 (2022)

    Sun, Y., et al.: Openood: Benchmarking gen- eralized out-of-distribution detection. Advances in Neural Information Processing Systems 35, 32598–32611 (2022)

    2022

  73. [81]

    In: International Conference on Learning Representations (2023)

    Zeng, S., Combes, R.T., Zhao, H.: Learn- ing structured representations by embedding class hierarchy. In: International Conference on Learning Representations (2023). https:// openreview.net/forum?id=7J-30ilaUZM

    2023

  74. [82]

    In: International Conference on Learning Representations (2022)

    Wang, G., Han, S., Zhang, D., et al.: Out-of- distribution detection based on in-distribution data patterns memorization with modern hop- field energy. In: International Conference on Learning Representations (2022)

    2022

  75. [83]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Chen, Y., Li, H.: Mixture outlier exposure: Towards out-of-distribution detection in fine- grained environments. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 5531–5540 (2023)

    2023

  76. [84]

    In: Interna- tional Joint Conference on Artificial Intelligence (2022)

    Zhang, B., Jiang, H., Feng, S., Li, X., Ye, Y., Ye, R.: Hyperbolic knowledge transfer with class hierarchy for few-shot learning. In: Interna- tional Joint Conference on Artificial Intelligence (2022)

    2022

  77. [85]

    IEEE transac- tions on pattern analysis and machine intelli- gence 40(6), 1452–1464 (2017)

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transac- tions on pattern analysis and machine intelli- gence 40(6), 1452–1464 (2017)

    2017

  78. [86]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zhang, Z., Xiang, X.: Decoupling maxlogit for out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3388–3397 (2023)

    2023

  79. [87]

    arXiv preprint arXiv:2306.09301 (2023) 24

    Zhang, W., Li, Y., Liu, Z., Chen, Y., Hai, L.: Openood v1.5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301 (2023) 24

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.