Pith. sign in

REVIEW 4 major objections 6 minor 48 references

Global and Local Entailment Learning for Natural World Imagery

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A transitivity-aware loss enforces taxonomic order in vision-language models.

desk verdict Interesting global-entailment loss for hierarchical VLMs, but the iNaturalist evaluation likely leaks training species and needs de-duplication before the zero-shot claims hold. read the letter →

arxiv 2506.21476 v1 pith:SAK7CHYA submitted 2025-06-26 cs.CV

Srikumar Sastry , Aayush Dhakal , Eric Xing , Subash Khanal , Nathan Jacobs This is my paper
classification cs.CV
keywords hierarchicalclassificationvision-languagemodelsentailmentlearningtransitivitytaxonomyTreeofLiferadialembeddingszero-shot
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 argues that vision-language models should learn hierarchies the way taxonomies actually behave: transitively, so that if 'Mammalia' entails 'Carnivora' and 'Chordata' entails 'Mammalia', then 'Chordata' also entails 'Carnivora'. Previous entailment-learning methods enforced only local parent-child relationships. The authors introduce Radial Cross-Modal Embeddings (RCME), a training objective that imposes this partial order directly on normalized radial embeddings without predefined cone angles, and build a Tree of Life foundation model from it. Across rank-wise zero-shot classification, hierarchical retrieval, and image-to-image retrieval, RCME outperforms the baselines it is compared with, with the largest gains at fine ranks. If the claim is right, the same objective can be lifted to any domain with a known is-a hierarchy.

What carries the argument

The load-bearing object is the transitivity-enforced entailment constraint on normalized radial embeddings. The paper defines global entailment as $S(T^i_{j-1}, T^i_{j+1}) \geq S(T^i_{j-1}, T^i_j)\,S(T^i_j, T^i_{j+1})$, derives from it a margin-based global loss $L_{\mathrm{GE}}$, and combines it with the local radial-embedding loss and a cross-modal alignment loss that jointly fine-tunes the vision and text encoders. The theory is anchored by Lemma 1: in a transitivity-enforced entailment, fine-grained concepts are progressively projected away from the entailment root into smaller subregions, which follows from the aperture-angle relation $\psi(T^i_j) \propto \arcsin(1/r(T^i_j, T_0))$. The framework needs no predefined cone aperture angles, removing the main tunable assumption of earlier entailment-cone methods.

What would settle it

Train RCME on TreeOfLife-10M with an entire phylum held out from training, then measure rank-wise zero-shot classification and Kendall's tau_d on species from that held-out phylum: if ordering accuracy on the unseen phylum stays close to in-distribution accuracy, transitivity generalizes, while a large drop would show the model memorized the training hierarchy instead of learning a transitive partial order.

Watch

Extended reading notes

Core claim

RCME's central claim is that global entailment, not just local parent-child entailment, can be imposed as a margin loss in vision-language embedding space. Concretely, with similarity $S$ between embeddings, the constraint $S(T^i_{j-1}, T^i_{j+1}) \geq S(T^i_{j-1}, T^i_j)\,S(T^i_j, T^i_{j+1})$ must hold for every consecutive triple in the hierarchy, and the paper turns this into the loss $L_{\mathrm{GE}} = \max(0, \Xi(T^i_{j-1}, T^i_{j+1}) - \arccos(S(T^i_{j-1}, T^i_j)\,S(T^i_j, T^i_{j+1})) + \alpha)$ with margin $\alpha = \pi/2$. Combined with a local entailment term and a cross-modal alignment term that fine-tunes both vision and text encoders, the objective pushes coarse concepts toward an entailment root and fine-grained species outward into smaller regions. Trained on TreeOfLife-10M, the model reaches an average 65.09% zero-shot accuracy across ranks on iNaturalist-2021 and Kendall's $\tau_d$ of 0.993 on hierarchical ordering, outperforming the baselines it compares against.

Load-bearing premise

The load-bearing assumption is that the species used for evaluation are effectively unseen during training, but both the training data and the evaluation sets come from the same iNaturalist platform and no species-level de-duplication is reported, so the apparent gains from transitive entailment could partly reflect memory of the training hierarchy rather than generalization to new branches of the Tree of Life.

Editorial extensions

If this is right

  • A single model can classify at any taxonomic rank, not just species, because distance from the entailment root carries semantic granularity.
  • Jointly fine-tuning the vision and text encoders improves fine-grained image-to-image retrieval at genus and species rank, indicating better separation of visually similar taxa.
  • The objective transfers outside biology: on HierarCaps, caption embeddings are ordered by caption granularity and retrieval improves relative to radial embeddings.
  • The representations expose ranks and kingdoms where the taxonomic system is brittle, such as plant families and orders, offering a tool for auditing taxonomic labels.

Reading between the lines

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

  • An untested but direct extension: the same partial-order loss should apply to medical ontologies, geographic region hierarchies, or product catalogs, since nothing in the objective is specific to biology.
  • A cleaner test of transitivity generalization would train on one large clade and evaluate on a taxonomically distant clade; because the training and evaluation data share the iNaturalist platform, a species-level de-duplication check is needed before attributing all gains to learned transitivity.
  • Combining the transitivity loss with hyperbolic geometry might yield embeddings with both the partial order and the extra capacity of curved spaces, since RCME currently works in Euclidean radial space.
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 proposes Radial Cross-Modal Embeddings (RCME), a fine-tuning framework for CLIP-style vision-language models that adds a global entailment loss enforcing a transitivity inequality over ancestor-parent-child text embeddings, a cross-modal alignment loss, and a hard-negative mining strategy. The authors train hierarchical models for the Tree of Life on TreeOfLife-10M and report gains over CLIP, OpenCLIP, BioCLIP, TaxaBind, Radial Embeddings, MERU, and ATMG in taxonomic-label ordering, zero-shot rank-wise classification, and image-to-image retrieval, plus a generalization experiment on HierarCaps.

Significance. If the empirical claims hold, the paper makes a useful incremental contribution: a simple, cone-free loss that imposes a global order on text embeddings, with open-sourced code and models and an ablation showing that the global loss and cross-modal alignment each help. The HierarCaps experiment also suggests the objective transfers beyond the Tree of Life. However, the principal evidence for zero-shot generalization to novel taxa is weakened by the lack of a reported species-level de-duplication between the iNaturalist-derived training and evaluation sets, and by a baseline-initialization confound. These issues are fixable but need to be resolved before the significance of the empirical gains can be assessed.

major comments (4)
  1. [§5.1 (Evaluation Datasets), Tables 1–4] The training set TreeOfLife-10M is iNaturalist-derived (via BioCLIP [36]), and both evaluation sets (iNaturalist-2021 and BioCLIP-Rare) are also iNaturalist-derived. The paper never reports a species-level or label-level de-duplication between training and evaluation. Since the zero-shot protocol evaluates exactly the species and taxonomic labels whose text embeddings were optimized during training, the high Kendall tau_d and accuracy numbers could reflect memorization of seen label positions rather than generalization of transitive entailment to novel taxa. This is load-bearing for the Section 1 motivation (reasoning about undescribed or novel species). Please quantify the label overlap and re-run the evaluation on a split that excludes all training species, or on species absent from TreeOfLife-10M.
  2. [§5.1 (Implementation Details and Baselines), Tables 2–3] The headline RCME model is initialized from OpenCLIP, while the fine-tuned baselines (Radial Emb., MERU, ATMG) start from BioCLIP. This confounds the method contribution with the initialization choice. The FT variant, which starts from BioCLIP, is below ATMG on both datasets (iNaturalist-2021: 57.71 vs 61.89; BioCLIP-Rare: 54.62 vs 61.39), while the OpenCLIP-initialized RCME is above. To support the "outperforms state-of-the-art" claim, all methods should be compared under matched initializations (both OpenCLIP and BioCLIP starts), or the discussion should be limited to the matched- initialization comparison.
  3. [§5.2 (Table 1) and Lemma 1/§4.1] The Kendall tau_d ordering metric evaluates how well labels are ordered by distance from the entailment root, and the proposed LGE loss is explicitly designed to enforce exactly that ordering (Lemma 1 and Eq. (6)). Consequently, a high tau_d on labels seen in training is largely a check that the objective was optimized, not independent evidence that the model has learned transitive entailment in a way that generalizes. The more informative results are the hierarchical retrieval precision/recall and the classification numbers, which are the ones affected by the overlap issue above. Please report tau_d on held-out taxa or re-frame the ordering result as a diagnostic of the training objective.
  4. [§5.2 (Tables 2–4), Appendix B] No standard deviations, confidence intervals, or number of seeds are reported for any experiment, and the models are trained for a single epoch. At several ranks the differences between RCME and the best baseline are small or negative (e.g., Table 3, Order: RCME 50.10 vs ATMG 53.03), so without variance estimates the aggregate gains cannot be assessed. Please report multiple runs/seeds or at least error bars on the headline averages.
minor comments (6)
  1. [§1, §2.3, §5.2] There are typos and grammatical slips: "Estevaet al." should be "Esteva et al.", "mislabeled A combination" needs punctuation, and "While these methods improves adaptability" should be "improve adaptability".
  2. [Table 5] The ablation table is hard to read: please label the first row explicitly (presumably BioCLIP) and clarify the initialization used in each row, especially whether the final row with LCMA is trained from OpenCLIP or BioCLIP.
  3. [Appendix B] The beta weight for LCMA is set to 0.1 for the BioCLIP-initialized model and 1.0 for the OpenCLIP-initialized model, but no sensitivity analysis for beta is reported; please add an ablation or at least justify the choice.
  4. [§5.2] The phrase "minimum absolute gain of +0.168 in correlation" is unclear; please report the comparison to the best baseline directly rather than as a minimum across all baselines.
  5. [Appendix A] The Euclidean case in the Lemma 1 proof assumes an unnormalized embedding space with the root at the origin, while the radial case assumes unit-norm embeddings; please clarify which geometry is used in the main RCME model and whether L2 normalization is applied before or after the proposed losses.
  6. [Appendix D (Table 9)] The text says performance improves with increasing margins, but the Kendall tau_d values are essentially identical across alpha values; please rephrase to avoid overclaiming and report other metrics if the effect is in precision/recall.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: Kendall's τd ordering metric measures exactly the distance-from-root property that the paper's own Lemma 1 proves follows from the transitivity constraint the loss directly enforces; classification and cross-domain results remain independent.

  1. self definitional [Section 3 (Lemma 1), Section 4.1 (Eq. 6), Appendix A (Eqs. 19-23), Section 5.2/Table 1]
    "Global entailment is enforced when Equation 4 holds for all possible sub-hierarchies in addition to local entailment... This ensures that Lemma 1 is satisfied. ... Firstly, we check whether the taxonomic labels are correctly ordered according to their distance from the entailment root using Kendall's Tau (τd)."

    L_GE (Eq. 6) is stated to enforce the transitivity inequality Eq. 4, and Appendix A (Eqs. 19-23) proves from Eq. 4 the consequence r(T^i_{j+1}, T_0) ≥ r(T^i_j, T_0): distance from the entailment root increases monotonically with rank. The Section 5.2 'ordering of taxonomic labels' experiment then measures, via Kendall's τd, exactly this theorem-consequence. The reported τd = 0.993 therefore verifies that the soft-margin optimization satisfied the loss's own target property; it is a self-check of the objective, not an independent demonstration that transitive entailment was learned. The retrieval-by-radius protocol (Appendix C) exploits the same enforced ordering, so Table 1's precision/recall inherit the alignment.

full rationale

Walking the derivation chain: the central contribution is a loss (L_GE, Eq. 6) that directly optimizes the transitivity inequality Eq. 4, which the paper's own Lemma 1 and Appendix A show is sufficient (given Ganea et al.'s cone premise, an external citation) for monotonic distance-from-root ordering. The headline 'ordering of taxonomic labels' evaluation (Table 1, Kendall's τd = 0.993) measures exactly that distance-from-root ordering, so this particular result is substantially a restatement of the objective: a partial, quotable self-definitionality, though not a hard identity, since the soft-margin loss does not guarantee constraint satisfaction (RCME_FT scores 0.963 while BioCLIP scores 0.012, so the metric is not vacuous). No parameters are fitted to the evaluation sets, no prediction is a renamed fit, and no load-bearing self-citation exists: TaxaBind and BirdSAT appear only as baselines/related work, and the entailment formalism is imported from external works (Ganea et al. [14], Esteva et al. [13]). The classification gains, the hard-negative ablation, and the out-of-domain HierarCaps table provide independent empirical content. The paper's self-flagged limitation (abnormally low plant accuracy at family/order ranks, Section 5.2) is a performance caveat with no circularity bearing. The reader's species-overlap concern (TreeOfLife-10M, iNaturalist-2021, and BioCLIP-Rare all iNaturalist-derived, with no reported species-level de-duplication) is a real data-validity risk for the zero-shot-generalization claim, but it is a leakage/correctness issue, not circularity, and therefore does not raise the circularity score. Overall: one headline metric reduces to the objective; the rest of the evaluation is self-contained, giving a partial circularity score of 4.

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

The method introduces no new physical entities. It depends on one tuned margin alpha, a tuned alignment weight beta, a chosen root token, and several domain assumptions about the correctness of the taxonomy and the validity of the cone-geometry inequalities used in the proof of Lemma 1.

free parameters (4)
  • alpha (margin) in LGE = pi/2
    Set to the maximum possible value pi/2; the ablation (Table 9) shows performance improves with larger margins, so this is a tuned hyperparameter, not a derived constant.
  • beta (cross-modal loss weight) = 0.1 (BioCLIP init), 1.0 (OpenCLIP init)
    Chosen separately for the two model variants without reported sensitivity analysis; affects the balance between entailment and alignment.
  • Entailment root token = 'Eukarya'
    A design choice; the root anchors all distance ordering and is not validated against alternatives.
  • Number of retrieval steps in ordering evaluation = 50
    Evaluation hyperparameter; the appendix shows precision and recall vary with step count, though Kendall's tau is stable.
assumptions (4)
  • domain assumption Transitivity is correctly characterized by Eq 4: S(T^j-1, T^j+1) >= S(T^j-1, T^j) * S(T^j, T^j+1).
    Stated as inspired by Esteva et al. [13] but not derived from cone geometry; the paper uses it as the definition of global entailment.
  • domain assumption In a transitivity-enforced entailment, Xi(T^j, T^j+1) <= psi(T^j) <= pi/2 holds (Eq 11, from Ganea et al. [14]).
    Used in the Appendix proof of Lemma 1; the paper does not verify this holds for the optimized RCME embeddings, and the loss with alpha=pi/2 does not guarantee it.
  • domain assumption The unit-hypersphere radial geometry is an appropriate space for the entailment partial order.
    Follows Alper et al. [1]; the proof of Lemma 1 relies on embeddings lying on a unit sphere.
  • domain assumption The Tree of Life taxonomy in TreeOfLife-10M is a correct, complete, and noise-free ground truth.
    All positive and negative triplets are sampled from this taxonomy; label noise (especially in plants) is acknowledged in the paper but not modeled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Global and Local Entailment Learning for Natural World Imagery." pith.science (2026). https://pith.science/paper/SAK7CHYA

@misc{pith2026250621476,
  author       = {Pith},
  title        = {Pith review of: Global and Local Entailment Learning for Natural World Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SAK7CHYA}},
  note         = {Machine review of arXiv:2506.21476}
}
read the original abstract

Learning the hierarchical structure of data in vision-language models is a significant challenge. Previous works have attempted to address this challenge by employing entailment learning. However, these approaches fail to model the transitive nature of entailment explicitly, which establishes the relationship between order and semantics within a representation space. In this work, we introduce Radial Cross-Modal Embeddings (RCME), a framework that enables the explicit modeling of transitivity-enforced entailment. Our proposed framework optimizes for the partial order of concepts within vision-language models. By leveraging our framework, we develop a hierarchical vision-language foundation model capable of representing the hierarchy in the Tree of Life. Our experiments on hierarchical species classification and hierarchical retrieval tasks demonstrate the enhanced performance of our models compared to the existing state-of-the-art models. Our code and models are open-sourced at https://vishu26.github.io/RCME/index.html.

Figures

Figures reproduced from arXiv: 2506.21476 by the authors.

Figure 1
Figure 1. Conceptual overview of our method focusing on pre [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Transitivity in Entailment. In an ideal transitivity-imposed entailment, textual embeddings satisfy partial order conditions. 3. Preliminaries We begin with the arguments on the conditions for entail￾ment as proposed in Esteva et al. [13] and Ganea et al. [14]. For our purposes, we consider the tree of life hierarchy and develop the logic for entailment with respect to it. Let R{j=0,1...N} represent the sets of the … view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: UMAP Visualization of Textual Embeddings. We visualize the textual embeddings using 2-D UMAP to show our model learns to preserve the partial order of taxonomic labels based on their distance from the entailment root. (R@1) for this task. Our method outperforms all oth…
Figure 5
Figure 5. Figure 5: UMAP Visualization of Textual Embeddings. The visualizations show our model has successfully imparted partial order in the textual embeddings. 3 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 37 canonical work pages

  1. [36]

    Bioclip: A vision foundation model for the tree of life

    Samuel Stevens, Jiaman Wu, Matthew J Thompson, Eliza- beth G Campolongo, Chan Hee Song, David Edward Carlyn, Li Dong, Wasila M Dahdul, Charles Stewart, Tanya Berger- Wolf, et al. Bioclip: A vision foundation model for the tree of life. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 19412–19424,

  2. [1]

    Emergent visual- semantic hierarchies in image-text representations

    Morris Alper and Hadar Averbuch-Elor. Emergent visual- semantic hierarchies in image-text representations. Euro- pean Conference on Computer Vision , 2024. 2, 4, 5, 6, 7, 8

    work page 2024

  3. [2]

    Multi-relational poincar ´e graph embeddings

    Ivana Balazevic, Carl Allen, and Timothy Hospedales. Multi-relational poincar ´e graph embeddings. Advances in Neural Information Processing Systems, 32, 2019. 2

    work page 2019

  4. [3]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European confer- ence on computer vision (ECCV), pages 456–473, 2018. 2

    2018

  5. [4]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher R ´e, and Jure Leskovec. Hyperbolic graph convolutional neural networks. Advances in neural information processing systems , 32,

    work page

  6. [5]

    Low-dimensional hy- perbolic knowledge graph embeddings

    Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher R ´e. Low-dimensional hy- perbolic knowledge graph embeddings. arXiv preprint arXiv:2005.00545, 2020. 2

    arXiv 2005

  7. [6]

    Mammalnet: A large-scale video benchmark for mammal recognition and behavior understanding

    Jun Chen, Ming Hu, Darren J Coker, Michael L Berumen, Blair Costelloe, Sara Beery, Anna Rohrbach, and Mohamed Elhoseiny. Mammalnet: A large-scale video benchmark for mammal recognition and behavior understanding. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13052–13061, 2023. 1, 2

    work page 2023

  8. [7]

    Mitree: Multi-input transformer ecoregion encoder for species distribution mod- elling

    Theresa Chen and Yao-Yi Chiang. Mitree: Multi-input transformer ecoregion encoder for species distribution mod- elling. In Proceedings of the 7th ACM SIGSPATIAL Interna- tional Workshop on AI for Geographic Knowledge Discov- ery, pages 110–120, 2024. 2

    work page 2024

Show all 48 references
  1. [8]

    Revis- iting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens

    Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revis- iting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. InPro- ceedings of the IEEE/CVF Conference on C...

    2023

  2. [9]

    Probabilistic language-image pre-training

    Sanghyuk Chun, Wonjae Kim, Song Park, and Sangdoo Yun. Probabilistic language-image pre-training. arXiv preprint arXiv:2410.18857, 2024. 2

    2024

  3. [10]

    Wildsat: Learning satel- lite image representations from wildlife observations

    Rangel Daroya, Elijah Cole, Oisin Mac Aodha, Grant Van Horn, and Subhransu Maji. Wildsat: Learning satel- lite image representations from wildlife observations. arXiv preprint arXiv:2412.14428, 2024. 2

    2024 arXiv

  4. [11]

    Hyper- bolic image-text representations

    Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyper- bolic image-text representations. In International Confer- ence on Machine Learning, pages 7694–7731. PMLR, 2023. 2, 5, 6, 7

    2023

  5. [12]

    Embedding text in hyperbolic spaces

    Bhuwan Dhingra, Christopher Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. Embedding text in hyperbolic spaces. In Proceedings of the Twelfth Workshop on Graph-Based Methods for Natural Language Processing (TextGraphs-12), pages 59–69, New Orleans, Louisiana, USA, 201...

    2018

  6. [13]

    Logics for approximate and strong en- tailments

    Francesc Esteva, Llu ´ıs Godo, Ricardo O Rodr ´ıguez, and Thomas Vetterlein. Logics for approximate and strong en- tailments. Fuzzy Sets and Systems, 197:59–70, 2012. 3

    2012

  7. [14]

    Hyperbolic entailment cones for learning hierarchical em- beddings

    Octavian Ganea, Gary B ´ecigneul, and Thomas Hofmann. Hyperbolic entailment cones for learning hierarchical em- beddings. In International conference on machine learning, pages 1646–1655. PMLR, 2018. 2, 3, 1

    2018

  8. [15]

    Lowe, Graham W

    ZeMing Gong, Austin Wang, Xiaoliang Huo, Joakim Brus- lund Haurum, Scott C. Lowe, Graham W. Taylor, and An- gel X Chang. CLIBD: Bridging vision and genomics for biodiversity monitoring at scale. In The Thirteenth Interna- tional Conference on Learning Representations, 2025. 1, 2

    2025

  9. [16]

    Pigeon: Predicting image geolocations

    Lukas Haas, Michal Skreta, Silas Alberti, and Chelsea Finn. Pigeon: Predicting image geolocations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12893–12902, 2024. 1

    2024

  10. [17]

    Contrastive ground-level image and remote sensing pre- training improves representation learning for natural world imagery

    Andy V Huynh, Lauren E Gillespie, Jael Lopez-Saucedo, Claire Tang, Rohan Sikand, and Mois ´es Exp ´osito-Alonso. Contrastive ground-level image and remote sensing pre- training improves representation learning for natural world imagery. In European Conference on Computer Visio...

    2024

  11. [18]

    Open- clip, 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Han- naneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Open- clip, 2021. 5, 6, 7, 2

    2021

  12. [19]

    Scaling up visual and vision-language representa- tion learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representa- tion learning with noisy text supervision. In International conference on machine learning, pages 4904–4916. PMLR,

  13. [20]

    Inferring concept hierarchies from text corpora via hyperbolic embeddings.arXiv preprint arXiv:1902.00913, 2019

    Matt Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximilian Nickel. Inferring concept hierarchies from text corpora via hyperbolic embeddings.arXiv preprint arXiv:1902.00913, 2019. 2

    1902 arXiv

  14. [21]

    9 Align before fuse: Vision and language representation learn- ing with momentum distillation

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. 9 Align before fuse: Vision and language representation learn- ing with momentum distillation. Advances in neural infor- mation processing systems, 34:9694–9705, 2021. 2

    2021

  15. [22]

    Fine-grained semantically aligned vision-language pre-training

    Juncheng Li, Xin He, Longhui Wei, Long Qian, Linchao Zhu, Lingxi Xie, Yueting Zhuang, Qi Tian, and Siliang Tang. Fine-grained semantically aligned vision-language pre-training. Advances in neural information processing sys- tems, 35:7290–7303, 2022. 2

    2022

  16. [23]

    Maruf, Arka Daw, Kazi Sajeed Mehrab, Harish Babu Manogaran, Abhilash Neog, Medha Sawhney, Mridul Khu- rana, James P

    M. Maruf, Arka Daw, Kazi Sajeed Mehrab, Harish Babu Manogaran, Abhilash Neog, Medha Sawhney, Mridul Khu- rana, James P. Balhoff, Yasin Bakis, Bahadir Altintas, Matthew J. Thompson, Elizabeth G. Campolongo, Josef C. Uyeda, Hilmar Lapp, Henry L. Bart, Paula M. Mabee, Yu Su, Wei-...

    2024

  17. [24]

    Cross the gap: Exposing the intra-modal misalignment in clip via modality inversion

    Marco Mistretta, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Andrew D Bagdanov. Cross the gap: Exposing the intra-modal misalignment in clip via modality inversion. arXiv preprint arXiv:2502.04263, 2025. 2

    2025 arXiv

  18. [25]

    Animal kingdom: A large and diverse dataset for animal behavior understanding

    Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023–19034, 2022. 1, 2

    2022

  19. [26]

    Poincar ´e embeddings for learning hierarchical representations

    Maximillian Nickel and Douwe Kiela. Poincar ´e embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 2

    2017

  20. [27]

    Learning continuous hierarchies in the lorentz model of hyperbolic geometry

    Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, pages 3779–

  21. [28]

    Compositional entailment learning for hyperbolic vision-language models

    Avik Pal, Max van Spengler, Guido Maria D’Amely di Me- lendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. The Thirteenth International Con- ference on Learning Representations, 2025. 2

    2025

  22. [29]

    Har- nessing artificial intelligence to fill global shortfalls in biodi- versity knowledge

    Laura J Pollock, Justin Kitzes, Sara Beery, Kaitlyn M Gaynor, Marta A Jarzyna, Oisin Mac Aodha, Bernd Meyer, David Rolnick, Graham W Taylor, Devis Tuia, et al. Har- nessing artificial intelligence to fill global shortfalls in biodi- versity knowledge. Nature Reviews Biodiversi...

  23. [30]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  24. [31]

    Accept the modality gap: An exploration in the hyperbolic space

    Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27263–27272, 2024. 2, 5, 6, 7

    2024

  25. [32]

    Accelerating ocean species discovery and laying the foundations for the future of marine biodiver- sity research and monitoring

    Alex David Rogers, Hannah Appiah-Madson, Jeff A Ardron, Nicholas J Bax, Punyasloke Bhadury, Angelika Brandt, Pier-Luigi Buttigieg, Olivier De Clerck, Claudia Delgado, Daniel L Distel, et al. Accelerating ocean species discovery and laying the foundations for the future of mari...

    2023

  26. [33]

    Birdsat: Cross-view contrastive masked autoencoders for bird species classification and mapping

    Srikumar Sastry, Subash Khanal, Aayush Dhakal, Di Huang, and Nathan Jacobs. Birdsat: Cross-view contrastive masked autoencoders for bird species classification and mapping. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 7136–7145, 2024. 2

    2024

  27. [34]

    Taxabind: A unified embedding space for ecological applications

    Srikumar Sastry, Subash Khanal, Aayush Dhakal, Adeel Ah- mad, and Nathan Jacobs. Taxabind: A unified embedding space for ecological applications. In Winter Conference on Applications of Computer Vision. IEEE/CVF, 2025. 1, 2, 5, 6, 7

    2025

  28. [35]

    Deepwild: Wildlife identification, localisation and estima- tion on camera trap videos using deep learning

    Fanny Sim ˜oes, Charles Bouveyron, and Fr ´ed´eric Precioso. Deepwild: Wildlife identification, localisation and estima- tion on camera trap videos using deep learning. Ecological Informatics, 75:102095, 2023. 2

    2023

  29. [37]

    Semi-supervised learn- ing with taxonomic labels.arXiv preprint arXiv:2111.11595,

    Jong-Chyi Su and Subhransu Maji. Semi-supervised learn- ing with taxonomic labels.arXiv preprint arXiv:2111.11595,

    arXiv

  30. [38]

    Poincar \’e glove: Hyperbolic word embeddings

    Alexandru Tifrea, Gary B ´ecigneul, and Octavian-Eugen Ganea. Poincar \’e glove: Hyperbolic word embeddings. arXiv preprint arXiv:1810.06546, 2018. 2

    2018 arXiv

  31. [39]

    The inaturalist species classification and de- tection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and de- tection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages...

  32. [40]

    Order-embeddings of images and language

    Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. arXiv preprint arXiv:1511.06361, 2015. 2

    2015 arXiv

  33. [41]

    Jones, Oisin Mac Aodha, Sara Beery, and Grant Van Horn

    Edward Vendrow, Omiros Pantazis, Alexander Shepard, Gabriel Brostow, Kate E. Jones, Oisin Mac Aodha, Sara Beery, and Grant Van Horn. Inquire: A natural world text- to-image retrieval benchmark, 2024. 2

    2024

  34. [42]

    Inquire: A natural world text-to-image retrieval benchmark

    Edward Vendrow, Omiros Pantazis, Alexander Shepard, Gabriel Brostow, Kate Jones, Oisin Mac Aodha, Sara Beery, and Grant Van Horn. Inquire: A natural world text-to-image retrieval benchmark. Advances in Neural Information Pro- cessing Systems, 37:126500–126514, 2025. 1

    2025

  35. [43]

    Geoclip: Clip-inspired alignment be- tween locations and images for effective worldwide geo- localization

    Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment be- tween locations and images for effective worldwide geo- localization. Advances in Neural Information Processing Systems, 36, 2023. 1

    2023

  36. [44]

    10 Learning visual hierarchies with hyperbolic embeddings

    Ziwei Wang, Sameera Ramasinghe, Chenchen Xu, Julien Monteil, Loris Bazzani, and Thalaiyasingam Ajanthan. 10 Learning visual hierarchies with hyperbolic embeddings. arXiv preprint arXiv:2411.17490, 2024. 2

    2024

  37. [45]

    Biotrove: A large curated image dataset enabling ai for bio- diversity

    Chih-Hsuan Yang, Benjamin Feuer, Talukder Jubery, Zi Deng, Andre Nakkab, Md Zahid Hasan, Shivani Chiranjeevi, Kelly Marshall, Nirmal Baishnab, Asheesh Singh, et al. Biotrove: A large curated image dataset enabling ai for bio- diversity. Advances in Neural Information Processin...

    2025

  38. [46]

    Filip: Fine-grained interactive language-image pre-training

    Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2

    2021 arXiv

  39. [47]

    Shadow cones: A generalized framework for partial order embeddings

    Tao Yu, Toni JB Liu, Albert Tseng, and Christopher De Sa. Shadow cones: A generalized framework for partial order embeddings. arXiv preprint arXiv:2305.15215, 2023. 2 11 Global and Local Entailment Learning for Natural World Imagery Supplementary Material A. Proof for Lemma 1 ...

    2023 arXiv

  40. [48]

    into smaller subregions in a transitivity-enforced entail- ment. We begin with the definition of distance in an entail- ment configuration: Ξ(T i j , Tk l ) = arccos ⟨(T i j − T0), (T k l − T i j )⟩ ||T i j − T0||.||T k l − T i j || ! (10) where ⟨·,·⟩ is an inner product betwe...

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.