Pith. sign in

REVIEW 3 major objections 4 minor 73 references

Hyperbolic Multimodal Continual Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Preventing forgetting in hyperbolic multimodal learning is equivalent to preserving old-task embeddings under one shared spatial rotation, and the paper derives a closed-form null-space projection that implements this constraint.

arxiv 2608.09572 v1 pith:ROKKQYWA submitted 2026-08-10 cs.LG

Jiahong Liu , Ming Shen , Xiaohao Liu , Rex Ying , Menglin Yang , Tat-Seng Chua , Irwin King This is my paper · ORCID
classification cs.LG MSC 68T0753B3015A63
keywords hyperbolicgeometrycontinuallearningmultimodalcatastrophicforgettingLorentzmodelcontrastiverepresentationpreservationnull-spaceprojection
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

This paper studies what happens to multimodal representations that live in hyperbolic space when a model is trained on a stream of tasks. It claims that preventing catastrophic forgetting in this setting requires old-task embeddings to stay invariant under one shared hyperbolic isometry: the same spatial rotation applied to every modality. The paper proves a necessary-and-sufficient characterization: under a non-degeneracy condition, preserving intra-modal, inter-modal, and hierarchical structure across continual stages is equivalent to the existence of a shared rotation $R=\mathrm{diag}(1,\tilde R)$ with $\tilde R\in SO(d)$ applied to all old-task embeddings. From this, it derives HMCL, which projects each parameter update onto the null space of old-task spatial representations, and reports large reductions in backward transfer on classification and retrieval benchmarks. If the proof holds, it gives a geometric first principle for designing continual learners for hyperbolic multimodal models rather than adapting Euclidean strategies.

What carries the argument

The central object is the extended Lorentz Gram matrix $G^s_{t,\mathrm{ext}}=Z^s_{t,\mathrm{all}}G(Z^s_{t,\mathrm{all}})^{\top}$ formed from the joint embedding matrix of all modalities; preserving it encodes both intra-modal and inter-modal invariance. The paper's argument runs through three geometric tools: Witt's extension theorem to show that equal Gram matrices force a global isometry, a uniqueness result for the proper orthochronous Lorentz group $SO^+(1,d)$ acting identically across modalities, and a boost-elimination argument showing that the spatial-norm condition (P3) forces that isometry to be a pure spatial rotation. The algorithmic mechanism that carries the result is the null-space projector $I-P^{t-1}$ onto the orthogonal complement of the old-task spatial subspace, computed once per stage from a low-rank basis; it is what turns the theorem into a closed-form update rule.

What would settle it

Train a hyperbolic multimodal model on a task stream using pure replay of old samples, with no geometric constraints, and measure both old-task accuracy and the extended Lorentz Gram matrices of old-task embeddings; if accuracy on previous tasks remains flat while those Gram matrices change, the paper's necessity claim would be refuted, whereas if accuracy drops whenever the Gram matrices change, the claim would be supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the preservation conditions (P1), (P2), and (P3) — exact preservation of intra-modal and inter-modal Lorentzian inner products and of spatial norms for old-task embeddings — hold between stage $t-1$ and stage $t$ if and only if there is a single spatial rotation $R=\mathrm{diag}(1,\tilde R)$, $\tilde R\in SO(d)$, with $Z^{m,t}_{t-1}=Z^{m,t-1}_{t-1}R^{\top}$ for every modality $m$. This means a Lorentz boost, which changes the spatial norm and therefore hierarchical specificity, is incompatible with preserving old-task knowledge, while a rotation shared across modalities preserves both cross-modal similarity and entailment-based hierarchy. The proof combines preservation of the extended Lorentz Gram matrix, uniqueness of the proper orthochronous Lorentz isometry under full-rank embeddings, and elimination of boosts via the spatial-norm condition. The paper then derives the first-order condition that admissible parameter updates are exactly those whose spatial part lies in the null space of the old-task spatial representations, with the time-like block fixed, yielding the closed-form update $\Delta W^m_s = \delta^m_s(I-P^{t-1})$.

Load-bearing premise

The load-bearing premise is that avoiding forgetting is exactly the same as preserving all pairwise Lorentzian inner products and spatial norms of old-task embeddings; if task accuracy could survive while these geometric quantities change, the shared-isometry requirement would be overkill.

Editorial extensions

If this is right

  • Any continual update that preserves old-task hyperbolic knowledge must act as one shared spatial rotation across modalities; a boost that changes spatial norms will break the hierarchy condition.
  • The admissible-update rule is a closed-form projection: unconstrained gradients are multiplied by $I-P^{t-1}$, so no extra parameters or per-task bases are needed and memory stays constant.
  • Geometry-agnostic continual methods such as EWC, GEM, and C-FLAT leave the Lorentz invariants unprotected, which explains their limited gains in the reported experiments.
  • In the reported benchmarks HMCL improves overall performance and reduces backward transfer substantially; for example classification BWT on MERU-L moves from $-6.46$ to $-0.74$.
  • Preserving the invariants also preserves the entailment cone partial order, so the image-to-root traversals remain semantically coherent after subsequent tasks.

Reading between the lines

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

  • The necessity claim is conditional on the paper's choice to define forgetting as exact preservation of all pairwise Lorentzian inner products and spatial norms; replay-based or task-specific-head methods could in principle maintain task accuracy while letting those invariants drift.
  • The same Gram-matrix-plus-null-space argument should transfer to other pseudo-Riemannian representation spaces; a natural test is to replace the Lorentz metric with another nondegenerate bilinear form and check whether the shared-isometry theorem still holds.
  • The first-order projection is designed for a linear Lorentz transformation layer; extending the projection to full-backbone hyperbolic fine-tuning would require an approximate tangent-space projection, and the paper's own appendix notes such higher-order residuals are not exactly controlled.
  • A cheap diagnostic suggested by the analysis: monitor time-like (radial) drift alone as an early warning for hierarchy collapse, since first-order hierarchy preservation forces spatial updates orthogonal to the current direction.
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

3 major / 4 minor

Summary. The paper studies continual learning for multimodal representations embedded in hyperbolic (Lorentz) space. It defines stability as exact preservation of three geometric invariants: intra-modal Lorentzian inner products (P1), inter-modal Lorentzian inner products (P2), and spatial norms related to hierarchy (P3). The main theoretical result, Theorem 1, states that under a non-degeneracy condition these preservation conditions hold if and only if the old-task embeddings are transformed by a single shared spatial rotation acting identically across modalities. From this characterization the authors derive a first-order constraint on parameter updates, leading to a closed-form null-space projection method called HMCL. Experiments on classification and retrieval benchmarks with three hyperbolic backbones report substantially reduced backward transfer and geometric drift relative to Euclidean continual-learning baselines.

Significance. If the theoretical characterization and the experimental isolation were both sound, this would be a valuable first geometric study of hyperbolic multimodal continual learning: the update rule is simple, introduces no additional learnable parameters, and has memory overhead constant in the number of tasks. The paper also goes beyond aggregate metrics by measuring radial, angular, cross-modal, and paired-distance drift, which is a useful diagnostic connection between geometry and forgetting. However, the central empirical claim is currently confounded by the step-pullback regularizer, and the proof of the main theorem contains a gap concerning the use of the hyperbolic origin. The theoretical framework is also conditional on the modeling assumption that preservation of P1–P3 is the correct formalization of 'preventing forgetting.' These issues are locally addressable but are load-bearing for the paper's main claims.

major comments (3)
  1. [Appendix E.1, Theorem 3] The proof of Theorem 3 ('Boost Elimination') applies condition (P3) to the specific embedding z=(1,0,...,0), i.e., the hyperbolic origin. The origin is not an old-task embedding and is not guaranteed by the non-degeneracy condition, which only ensures that the rows of the joint embedding matrix span R^{1,d}. Because (P3) is a per-point nonlinear constraint, showing that a boost fails to preserve the origin's spatial norm does not imply that the boost fails on a spanning set of old-task embeddings. Consequently the 'only if' direction of Theorem 1 is not established as stated. A correct proof needs a lemma showing directly that if L in SO+(1,d) preserves spatial norms on a spanning subset of H^d, then L is a pure spatial rotation; if such a lemma is false, the theorem needs an explicit additional assumption.
  2. [Section 6.1, Table 1, Figure 5] The step pullback W←(1−β)W+βW_{t,0} is applied to HMCL but is not reported for the baselines, and the ablation 'HMCL w/o block' in Figure 5 still contains the pullback. This pullback is a strong weight-anchoring regularizer that can reduce backward transfer independently of the geometric projection. The baseline paragraph states that baselines share the backbone, task stream, optimizer, and learning-rate schedule, but not the pullback. Therefore the experiments do not isolate the null-space projection derived from Theorem 1, and the claim that 'geometry-aware constraints are key to stable multimodal continual learning' is not yet supported. The authors should add control baselines trained with the same pullback and an HMCL variant without the pullback.
  3. [Section 3.2 and Theorem 1] The paper defines stability as the preservation conditions (P1)–(P3) and then concludes that preventing forgetting requires a shared hyperbolic isometry. This conclusion is relative to that definition: exact preservation of all pairwise Lorentzian inner products and spatial norms is sufficient but not shown to be necessary for maintaining task performance, which could in principle be preserved by replay, task-specific heads, or other mechanisms. The abstract's stronger phrasing ('preventing forgetting requires cross-modal invariance') should be qualified, or the paper should supplement Theorem 1 with an argument linking (P1)–(P3) to downstream task performance.
minor comments (4)
  1. [Section 4.1, Theorem 1] The 'non-degeneracy condition' is not defined in the main text; it appears only in the appendix proofs as a rank condition on the joint embedding matrix. Please state the condition explicitly in the theorem statement.
  2. [Section 4.2, Corollary 1] Corollary 1 says 'If the time-like coordinate is required to be first-order invariant, i.e. (P3),' but (P3) is an exact equality, not a first-order condition. Please rephrase to distinguish the exact preservation condition from its first-order implication.
  3. [Section 6.3, Figure 3, Table 7] The drift reductions are reported as aggregate percentages with a p-value but without confidence intervals in the main text. Please report per-dataset error bars or a table with confidence intervals so the reader can assess the variability across old tasks.
  4. [Appendix G.2, Figure 5] The caption of Figure 5 should state explicitly that the 'w/o block' variant retains the step pullback; currently the reader must infer this from Section 6.1, and the caption could easily be misread as an ablation of the full geometric mechanism.

Circularity Check

1 steps flagged · score 2.0 of 10

The theorem chain is internally self-contained; the main definitional overclaim is equating (P1)-(P3) with preventing forgetting, and the step-pullback regularizer is not isolated from the geometric update in the ablation.

  1. self definitional [Section 3.2 (P1)-(P3), compared with Theorem 1 in Section 4.1 and the Abstract]
    "In hyperbolic representation learning, the preservation of historical knowledge fundamentally depends on maintaining the geometric structure of the embedding space. This requires maintaining three geometric properties of the learned representations."

    The abstract's headline claim that 'preventing forgetting requires cross-modal invariance under a shared hyperbolic isometry' is obtained by chaining two equivalences: Section 3.2 defines preservation of historical knowledge as the three geometric invariants (P1)-(P3), and Theorem 1 states that (P1)-(P3) hold if and only if there exists a shared spatial rotation R=diag(1,\tilde R) with Z^{m,t}_{t-1}=Z^{m,t-1}_{t-1}R^\top. Hence the 'requirement' is true by construction relative to that definition; Theorem 1 does not establish that task-level forgetting (the BWT measured later) forces P1-P3. The algebra in the proof is nontrivial, so this is a framing circularity and an overclaim rather than a fabricated derivation.

full rationale

The formal derivation chain (Theorem 1 -> Proposition 1 -> Corollary 2) is self-contained and not fitted to data: Theorem 1 is an algebraic equivalence proved via Witt's extension theorem and a boost-elimination argument, and the null-space projection update in Corollary 2 follows from an explicit minimal-rotation regularization (lambda -> infinity). No parameter is fitted and then relabeled as a prediction. The only definitional circularity-adjacent issue is the bridge between the P1-P3 stability conditions and the abstract's 'preventing forgetting': P1-P3 are assumed to be the formal content of stability, and Theorem 1 then characterizes them as a shared rotation, so the headline necessity claim is conditional on that modeling assumption and does not by itself prove necessity for task-level performance. This is an overclaim, not a circular fit. On the empirical side, Section 6.1 describes a step pullback W <- (1-beta)W + beta W_{t,0} with beta=0.10 and beta=0.20, and Table 1's caption says 'HMCL uses the reported step-pullback configuration', while the baselines paragraph does not state that baselines receive this pullback; this weakens the attribution of the BWT gains to the geometric projection, but it is an experimental confound rather than a circular derivation. The self-citation to Liu et al. 2025b for the continual multimodal framework is not load-bearing for the theorem. Overall, the mathematics is not circular; the score reflects the minor definitional overclaim and the unisolated regularizer, not a constructed circularity.

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

The central derivation relies on standard results in quadratic forms and the Lorentz group, plus two paper-specific assumptions: that P1-P3 encode all that matters for preventing forgetting, and that the proof's origin-based test can be replaced by the rank and spanning argument. The method's practical success also depends on hand-tuned hyperparameters such as beta, retained variance, loss weight, and curvature.

free parameters (4)
  • step pullback coefficient beta = 0.10 (MERU-L/MERU-B), 0.20 (HyCoCLIP-B)
    Chosen per backbone to stabilize finite-step AdamW; not derived from theory and not applied to baselines.
  • memory projector retained variance ratio = 0.9 (MERU-L), 0.8 (MERU-B), 0.9 (HyCoCLIP-B)
    Controls how much of the old-task spatial subspace is protected via PCA/SVD; hand-tuned per backbone.
  • entailment loss weight = 0.2
    Set for all tasks; no sensitivity analysis is reported.
  • curvature magnitude K = 0.1
    Fixed curvature for all experiments; affects cone apertures and distances.
assumptions (6)
  • standard math Witt's extension theorem for non-degenerate quadratic forms
    Used in Theorem 2 and Appendix E.1 to extend an isometry between spanning sets to a global isometry in O(1,d).
  • standard math Lorentz group polar decomposition: every SO+(1,d) element is a spatial rotation composed with a boost
    Used in Theorem 3 to reduce L to a rotation times a boost and test the boost component.
  • domain assumption Old-task embedding matrix has full rank d+1
    Non-degeneracy condition used for uniqueness of the isometry; may fail in practice with low-rank or collapsed representations.
  • ad hoc to paper P1-P3 preservation is the correct formalization of preventing task-level forgetting
    The paper equates geometric invariant preservation with stability, but does not prove that task performance requires these exact invariants.
  • ad hoc to paper Old-task embeddings include the hyperbolic origin, or the proof's origin test is valid
    Theorem 3 in Appendix E.1 applies P3 to z=[1,0,...,0]; the origin need not be present, so the written proof is incomplete and must be repaired with the rank condition.
  • domain assumption Finite-step AdamW with step pullback approximates the infinitesimal admissible-update analysis
    Appendix F.3 acknowledges higher-order and optimizer-induced residuals; the pullback is an extra stabilizer not derived from the geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyperbolic Multimodal Continual Learning." pith.science (2026). https://pith.science/paper/ROKKQYWA

@misc{pith2026260809572,
  author       = {Pith},
  title        = {Pith review of: Hyperbolic Multimodal Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROKKQYWA}},
  note         = {Machine review of arXiv:2608.09572}
}
read the original abstract

Hyperbolic geometry has recently emerged as a powerful representation space for multimodal learning, as it naturally captures hierarchical semantic structure across modalities. Despite this progress, how such representations behave under continual learning poses fundamentally different challenges that remain underexplored. This work provides a geometric perspective on this problem and establishes a theoretical foundation for representation preservation in hyperbolic space, showing that preventing forgetting requires cross-modal invariance under a shared hyperbolic isometry. We further show that forgetting in hyperbolic continual learning involves both semantic relation drift and hierarchy-related distortion, motivating preservation of both cross-modal relational structure and hierarchical geometry. Guided by these insights, a principled continual learning framework is derived that preserves essential geometric structure while allowing effective adaptation to new tasks. Experiments on continual multimodal benchmarks corroborate the effectiveness of the proposed approach.

Figures

Figures reproduced from arXiv: 2608.09572 by the authors.

Figure 1
Figure 1. Continual learning in hyperbolic representation space. Multimodal representations are embedded in hyperbolic space to capture semantic hierarchy across tasks. Under naive continual updates, distortions in the hyperbolic geometry lead to catastrophic forgetting and collapse of previously learned structures, whereas structure-aware hyperbolic updates significantly reduce forgetting. modal data and contrastive objectiv… view at source ↗
Figure 2
Figure 2. Geometric interpretation of hyperbolic multimodal continual learning. Multimodal representations are embedded in hyperbolic space, where distance encodes semantic similarity and distance to the origin encodes semantic hierarchy. Preserving old-task knowledge across tasks requires maintaining these geometric relations under a shared hyperbolic isometry. The top-down view illustrates that first-order hierarchy preserv… view at source ↗
Figure 3
Figure 3. Dataset-wise old-task representation drift from the task-end checkpoint to the final checkpoint. Gray circles denote Vanilla and blue diamonds denote HMCL; shaded rows indicate retrieval tasks. We measure four complementary changes: radial drift for hierarchy-related radius changes, angular drift for within-modality directional distortion, cross-modal drift for image–text relational structure, and paired-distance dr… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Case study: image traversals and hierarchical concept retrieval. For three Flickr30K images, we traverse each embedding toward [ROOT] and retrieve concepts from specific to generic. HMCL yields coherent hierarchies (e.g., hat → fashion → style), whereas Vanilla drifts …
Figure 5
Figure 5. Figure 5: Ablation of block-wise admissible updates. Results comparing Vanilla, HMCL (w/ block), and a variant without block-wise decomposition (w/o block). Removing the separation between time-like and space-like updates leads to consistently worse backward transfer, highlighti…
Figure 6
Figure 6. Figure 6: Flickr30K geodesic-interpolation case studies, page 1. Cases (1)–(3) retrieve over a mixed pool of captions, nouns, and adjectives. Blue rows show the checkpoint immediately after learning Flickr30K, and green rows show the final checkpoint after subsequent continual-l…
Figure 7
Figure 7. Figure 7: Flickr30K geodesic-interpolation case studies, page 2. Cases (4)–(6) restrict retrieval to the caption pool. Blue rows show the checkpoint immediately after learning Flickr30K, and green rows show the final checkpoint after subsequent continual-learning tasks. 29 [PIT…
Figure 8
Figure 8. Figure 8: COCO geodesic-interpolation case studies, page 1. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]
Figure 9
Figure 9. Figure 9: COCO geodesic-interpolation case studies, page 2. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_9.png]
Figure 10
Figure 10. Figure 10: COCO geodesic-interpolation case studies, page 3. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_10.png]
Figure 11
Figure 11. Figure 11: COCO geodesic-interpolation case studies, page 4. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 32 canonical work pages

  1. [1]

    International conference on machine learning , pages=

    Learning continuous hierarchies in the lorentz model of hyperbolic geometry , author=. International conference on machine learning , pages=. 2018 , organization=

    work page 2018

  2. [2]

    Advances in neural information processing systems , volume=

    Hyperbolic graph convolutional neural networks , author=. Advances in neural information processing systems , volume=

  3. [3]

    Learning multiple layers of features from tiny images , author=

  4. [4]

    Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition , volume=

    Learning generative visual models from few training examples: An incremental Bayesian approach tested on 101 object categories , author=. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition , volume=

    work page

  5. [5]

    European Conference on Computer Vision , pages=

    Food-101 -- Mining discriminative components with random forests , author=. European Conference on Computer Vision , pages=. 2014 , organization=

    work page 2014

  6. [6]

    European Conference on Computer Vision , pages=

    Microsoft COCO: Common objects in context , author=. European Conference on Computer Vision , pages=. 2014 , organization=

    2014

  7. [7]

    International Conference on Learning Representations , year=

    Adam: A method for stochastic optimization , author=. International Conference on Learning Representations , year=

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    PyTorch: An imperative style, high-performance deep learning library , author=. Advances in Neural Information Processing Systems , volume=

Show all 73 references
  1. [9]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    arXiv

  2. [10]

    Transactions of the association for computational linguistics , volume=

    From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions , author=. Transactions of the association for computational linguistics , volume=

  3. [11]

    2013 , eprint=

    Fine-Grained Visual Classification of Aircraft , author=. 2013 , eprint=

    2013

  4. [12]

    Lawrence and Girshick, Ross , year=

    Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens and Fei-Fei, Li and Zitnick, C. Lawrence and Girshick, Ross , year=. 1612.06890 , archivePrefix=

    arXiv

  5. [13]

    2013 , eprint=

    Describing Textures in the Wild , author=. 2013 , eprint=

    2013

  6. [14]

    1709.00029 , archivePrefix=

    Helber, Patrick and Bischke, Benjamin and Dengel, Andreas and Borth, Damian , year=. 1709.00029 , archivePrefix=

    arXiv

  7. [15]

    2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing , pages=

    Automated Flower Classification over a Large Number of Classes , author=. 2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing , pages=. 2008 , doi=

    2008

  8. [16]

    Proceedings of the IEEE , volume=

    Gradient-based learning applied to document recognition , author=. Proceedings of the IEEE , volume=. 1998 , doi=

    1998

  9. [17]

    and Linmans, Jasper and Winkens, Jim and Cohen, Taco and Welling, Max , year=

    Veeling, Bastiaan S. and Linmans, Jasper and Winkens, Jim and Cohen, Taco and Welling, Max , year=. Rotation Equivariant. 1806.03962 , archivePrefix=

    arXiv

  10. [18]

    Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages=

    Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank , author=. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages=. 2013 , publisher=

    2013

  11. [19]

    Advances in neural information processing systems , volume=

    Hyperbolic neural networks , author=. Advances in neural information processing systems , volume=

  12. [20]

    Proceedings of the 35th International Conference on Machine Learning , pages=

    Hyperbolic Entailment Cones for Learning Hierarchical Embeddings , author=. Proceedings of the 35th International Conference on Machine Learning , pages=. 2018 , editor=

    2018

  13. [21]

    Nickel, Maximillian and Kiela, Douwe , journal=. Poincar

  14. [22]

    Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=

    Hyperbolic geometry of complex networks , author=. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=. 2010 , publisher=

    2010

  15. [23]

    arXiv preprint arXiv:2504.08896 , year=

    Position: Beyond Euclidean--Foundation Models Should Embrace Non-Euclidean Geometries , author=. arXiv preprint arXiv:2504.08896 , year=

  16. [24]

    2025 , eprint=

    Hyperbolic Graph Neural Networks: A Review of Methods and Applications , author=. 2025 , eprint=

    2025

  17. [25]

    Proceedings of the ACM web conference 2022 , pages=

    HRCF: Enhancing collaborative filtering via hyperbolic geometric regularization , author=. Proceedings of the ACM web conference 2022 , pages=

    2022

  18. [26]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    HICF: Hyperbolic Informative Collaborative Filtering , author=. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=. 2022 , doi=

    2022

  19. [27]

    2022 , eprint=

    Enhancing Hyperbolic Graph Embeddings via Contrastive Learning , author=. 2022 , eprint=

    2022

  20. [28]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Fully hyperbolic neural networks , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  21. [29]

    International Conference on Machine Learning , pages=

    Hyperbolic image-text representations , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  22. [30]

    European Conference on Computer Vision , pages=

    Hyperbolic learning with multimodal large language models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  23. [31]

    European Conference on Computer Vision , pages=

    Hype: Hyperbolic entailment filtering for underspecified images and texts , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  24. [32]

    Transactions on Machine Learning Research , year=

    Intriguing properties of hyperbolic embeddings in vision-language models , author=. Transactions on Machine Learning Research , year=

  25. [33]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Hypformer: Exploring efficient transformer fully in hyperbolic space , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  26. [34]

    Advances in Neural Information Processing Systems , year=

    Hyperbolic Fine-Tuning for Large Language Models , author=. Advances in Neural Information Processing Systems , year=

  27. [35]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    HiHPQ: Hierarchical Hyperbolic Product Quantization for Unsupervised Image Retrieval , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2024 , doi=

    2024

  28. [36]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Understanding and Mitigating Hyperbolic Dimensional Collapse in Graph Contrastive Learning , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=. 2025 , doi=

    2025

  29. [37]

    2024 , eprint=

    HC-GLAD: Dual Hyperbolic Contrastive Learning for Unsupervised Graph-Level Anomaly Detection , author=. 2024 , eprint=

    2024

  30. [38]

    FedKDD: International Joint Workshop on Federated Learning for Data Mining and Graph Analytics , year=

    FlatLand: Personalized Federated Learning via Tailored Lorentz Space , author=. FedKDD: International Joint Workshop on Federated Learning for Data Mining and Graph Analytics , year=

  31. [39]

    2024 , url=

    Yang, Menglin and Liu, Jiahong and King, Irwin and Ying, Rex , booktitle=. 2024 , url=

    2024

  32. [40]

    2024 , url=

    Enhancing Multimodal Survival Prediction with Pathology Reports in Hyperbolic Space , author=. 2024 , url=

    2024

  33. [41]

    arXiv preprint arXiv:2006.08210 , year=

    Hyperbolic neural networks++ , author=. arXiv preprint arXiv:2006.08210 , year=

    2006 arXiv

  34. [42]

    The Thirteenth International Conference on Learning Representations , year=

    Compositional Entailment Learning for Hyperbolic Vision-Language Models , author=. The Thirteenth International Conference on Learning Representations , year=

  35. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Accept the modality gap: An exploration in the hyperbolic space , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  36. [44]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Hyperbolic Safety-Aware Vision-Language Models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  37. [45]

    arXiv preprint arXiv:2506.10710 , year=

    Continual Hyperbolic Learning of Instances and Classes , author=. arXiv preprint arXiv:2506.10710 , year=

    arXiv

  38. [46]

    arXiv preprint arXiv:1606.04671 , year=

    Progressive neural networks , author=. arXiv preprint arXiv:1606.04671 , year=

    arXiv

  39. [47]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Piggyback: Adapting a single network to multiple tasks by learning to mask weights , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  40. [48]

    Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages=

    Packnet: Adding multiple tasks to a single network by iterative pruning , author=. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages=

  41. [49]

    Advances in Neural Information Processing Systems , volume=

    Make continual learning stronger via c-flat , author=. Advances in Neural Information Processing Systems , volume=

  42. [50]

    arXiv preprint arXiv:2403.06295 , year=

    A streamlined approach to multimodal few-shot class incremental learning for fine-grained datasets , author=. arXiv preprint arXiv:2403.06295 , year=

    arXiv

  43. [51]

    Euclidean Multimodal Contrastive Learning: Adapting Alignment Calibration to MERU , author=

    Machine Unlearning in Hyperbolic vs. Euclidean Multimodal Contrastive Learning: Adapting Alignment Calibration to MERU , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  44. [52]

    arXiv preprint arXiv:2504.14280 , year=

    CLIP-Powered Domain Generalization and Domain Adaptation: A Comprehensive Survey , author=. arXiv preprint arXiv:2504.14280 , year=

    arXiv

  45. [53]

    Advances in neural information processing systems , volume=

    Gradient episodic memory for continual learning , author=. Advances in neural information processing systems , volume=

  46. [54]

    Advances in neural information processing systems , volume=

    Dark experience for general continual learning: a strong, simple baseline , author=. Advances in neural information processing systems , volume=

  47. [55]

    International Conference on Machine Learning , pages=

    Continual vision-language representation learning with off-diagonal information , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  48. [56]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Boosting continual learning of vision-language models via mixture-of-experts adapters , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  49. [57]

    Proceedings of the national academy of sciences , volume=

    Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the national academy of sciences , volume=. 2017 , publisher=

    2017

  50. [58]

    International Conference on Machine Learning , pages=

    The numerical stability of hyperbolic representation learning , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  51. [59]

    Advances in neural information processing systems , volume=

    Clap4clip: Continual learning with probabilistic finetuning for vision-language models , author=. Advances in neural information processing systems , volume=

  52. [60]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Preventing zero-shot transfer degradation in continual learning of vision-language models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  53. [61]

    arXiv preprint arXiv:2410.05352 , year=

    Recent advances of multimodal continual learning: A comprehensive survey , author=. arXiv preprint arXiv:2410.05352 , year=

  54. [62]

    arXiv preprint arXiv:2507.17343 , year=

    Principled Multimodal Representation Learning , author=. arXiv preprint arXiv:2507.17343 , year=

  55. [63]

    arXiv preprint arXiv:2511.12034 , year=

    Calibrated Multimodal Representation Learning with Missing Modalities , author=. arXiv preprint arXiv:2511.12034 , year=

    arXiv

  56. [64]

    arXiv preprint arXiv:2503.14963 , year=

    Continual multimodal contrastive learning , author=. arXiv preprint arXiv:2503.14963 , year=

  57. [65]

    The Thirteenth International Conference on Learning Representations , year=

    C-CLIP: Multimodal continual learning for vision-language model , author=. The Thirteenth International Conference on Learning Representations , year=

  58. [66]

    arXiv preprint arXiv:2409.07402 , year=

    What to align in multimodal contrastive learning? , author=. arXiv preprint arXiv:2409.07402 , year=

    arXiv

  59. [67]

    2005 , publisher=

    Lam, Tsit-Yuen , series=. 2005 , publisher=

    2005

  60. [68]

    arXiv preprint math-ph/0211047 , year=

    The interplay of the polar decomposition theorem and the Lorentz group , author=. arXiv preprint math-ph/0211047 , year=

    arXiv

  61. [69]

    ICML , pages=

    Learning transferable visual models from natural language supervision , author=. ICML , pages=

  62. [70]

    CVPR , pages=

    Momentum contrast for unsupervised visual representation learning , author=. CVPR , pages=

  63. [71]

    arXiv preprint arXiv:2309.03905 , year=

    Imagebind-llm: Multi-modality instruction tuning , author=. arXiv preprint arXiv:2309.03905 , year=

    arXiv

  64. [72]

    NeurIPS , volume=

    Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset , author=. NeurIPS , volume=

  65. [73]

    International symposium on graph drawing , pages=

    Low distortion delaunay embedding of trees in hyperbolic plane , author=. International symposium on graph drawing , pages=. 2011 , organization=

    2011

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.