acceptodds
Under review as a conference paper at ICLR 2027

HyGLoA: Dual-Granularity Hyperbolic Modeling with Local Geodesic-Ray Alignment for Zero-Shot Learning

Abstract

Zero-shot learning (ZSL) recognizes unseen classes by transferring knowledge from auxiliary semantics. Existing methods have explored various visual-semantic alignment schemes at either attribute or class granularity, yet most perform in Euclidean spaces. The polynomial volume growth of Euclidean space makes it difficult to allocate increasing capacity to finer semantic levels, while its geometry does not naturally encode the asymmetric specificity relations among classes, attributes, and visual evidence. We introduce HyGLoA (Hyperbolic Global–Local Alignment), a dual-granularity framework that represents global class identity and local attribute evidence in two independent hyperbolic spaces under the Lorentz model. The global branch aligns the image CLS token with class-name prototypes, whereas the local branch projects patch tokens, aggregates them via Lorentzian centroid pooling, and aligns the resulting representation with attribute prototypes. Both branches use branch-specific direction–radius decoupled projections to control semantic direction and geodesic radius separately. To capture the asymmetric relation between partial local evidence and complete attribute semantics, we derive local geodesic-ray alignment from the zero-aperture limit of hyperbolic entailment. It removes cone-angle tolerance and penalizes deviation of the pooled local representation from the matched prototype's origin-facing geodesic ray. We further perform prototype bundle expansion on normalized semantic prototypes before hyperbolic lifting, improving angular discrimination without interfering with radial modeling. The two branch distributions undergo probability-level fusion for final prediction. Experiments on three popular benchmarks yield the highest generalized-ZSL harmonic mean among the compared methods, while ablations validate the principal design choices.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
zero-shot learning, hyperbolic spacce, visual-semantic alignment.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3BmnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper deals with zero-shot learning (ZSL): object classes are associated to a set of attributes, a visual/attributes mapping is learned on seen classes such that at test time, the model is able to classify samples from unseen classes that are associated to the same attributes as the seen classes but have not been used at training time.

Previous works proposed to learn a hyperbolic (visual) embedding for ZSL. The first claimed novelty of the current paper is to propose two mappings: the CLS visual token to the the class name on the one hand (global branch) and the patch tokens to the attrbute prototypes on the other hand (local branch). The visual representations are projectd (lifted) to the hyperbolic manifold. Both benefit from an aditional projection, the radius offset head (ROH), which modulated the direct representation. In addition, the local representation is aggregated. This lorentz centroid pooling (LCP) consists to compute the average o gthe local listed points in the ambiant Euclidean space then project back to the hyperbolic manifold. Before the projection (hyperbolic lifting), the author propose to ``spread'' the semantic representations (class name or attribute) in order to improve the ability to distinguish them (process named PBE: Prototype bundle expansion). After projection, the resulting logits then probabilities are combined linearly. Last, on the local branch, the method include a regularization (LGRA: local geodesic-ray alignment) that force the local representation to align to the attribute prototype.

The method is compared to several recent approaches on CUB, SUN and AWA2, which are usual benchmarks in the ZSL literature. A quite significant ablation study is also reported, allowing to compare the relative merits of each module of the full model. The stability of the performance is also reported for two important hyperparameters. Last, the paper reports qualitative results, in particular to illustrate the effect of the ROH.

Critical strengths
  • [S1] the protoype bundle expansion is an original contribution, that seems efficient if one compare (Table 2) the model with all the proposed modules and that without PBE (line f). One also appreciate that it construction is motivated with some details in appendix A.2

  • [S2] another original contribution is the radius-offset head that acts as a sort of attention on the normalized local and global representations. A study of its specific interest would have been interesting (see [Q2])?

The paper reports good performance on three datasets that are usual in the ZSL literature. However, I consider that the protocol is unfair, as explained in [W1] below, thus it can not be considered as a strength at the same time.

Critical weaknesses

[W1] the evaluations are unfair, because most hyperparameters were optimzed directly on the test set. To the authors’ credit, they are quite transparent on this point, in appendix B.3, on line 952 and in Appendix B.4 on line 991 and in section 4.4 (line 422 and 424). It is particularly stricking for the setup of w0=0.6w_0=0.6 on SUN (Table 3, line 855) that correspond to "peak of performance" on the test set (Figure 6b), but it seems quite general, including λLGRA\lambda_{LGRA} and fusion weights. One must nevertheless recognize that it is a common practice in the ZSL literature, SVIP and VSPCN having been published (ICCV and CVPR 2025 respectively) with likely the same problem, and that reporteing better results than them with a fair protocol is difficult. It nevertheless remains a fundamentaly bad practice that should be banned as soon as possible. In 2017, Xian et al. published a famous paper to highlight ``ugly'' practices in ZSL. It seems that there is still work to do on this point, althought it is not specific to the current paper.

  • another good practice, not applied in this paper nor in the usual ZSL literature, consists to run several times the experiments and report either a standard deviation or applying a statistical test to assesss the significance of the differences of performance.
  • [W1a] regarding the performances reported, it is also surprising that the baseline model of the proposed approach without any of the proposed contribution (but PBE) (line 'a' of Table 2) has performance on-par (CUB, SUN) or even much better (AWA2) than recent state of the art methods (Table 1). It is likely it explains most of the good performance reported in Table 1. However, it is diffcult to understand and not commented in section 4.3: See questions.

[W2] regarding the novelty of the contribution, it looks like it essentially consists to aggregate existing approaches, or their utility is arguable.

  • as mentioned in the paper, hyperbolic projection has already been proposed in the ZSL literature (Liu et al, 2020). While not mentioned in the related works, multi-branch mapping (image/class + attributes/token) has also been exhaustivly explored in the prototype network literature (e.g [a] and subsequent works; a paper such as [b,c] have similar dual-branch architecture as the current paper). The first claim, that consist to mix these two ideas, has thus a limited novelty.
  • the LCP is a naive average followed by a projection on the manifold (equation 6). This choise seems odd: why the average is not directly computed in the manifold with its natural metric? While efficient in practice (Table 2; lines h versus full) computing a naive average to agregate can not be considered as particularly inovative.
  • LGRA seems to very very close, if not identical to the Lentail\mathcal{L}_{entail} of (Desai et al, 2023). In particular, equation (11) of the current manuscript is essentially similar to equations (11) and (12) of (Desai et al, 2023). Here again, the novelty seems marginal in comparison to this previous work.

[a] Xu et al (2020) Attribute Prototype Network for Zero-Shot Learning, NeurIPS

[b] Xiang et al. (2023) Dual Feature Augmentation Network for Generalized Zero-shot Learning, BMVC

[c] Shen et al. (2025) Zero-shot image classification based on class representation learning and attribute embedding learning, PLoS One

Questions for authors
  • [Q1] The ablation contains several interesting comparison but it misses the simplests, namely that with P and C onlyn without PBE (and possibly "C only" and "P only") wich is is the most basic baseline. On line 'a', with PBE only, the performances are already around the state of the art on CUB and SUN and much better on AWA2. It seems to explain most of the top results reported with the contributions. However the baseline results, withour most of the proposed contributions, are diffcult to understand. Is it due to PBE: what are the results without this module at all (line (a) without PBE)? How the very good baseline results can be explained?

    • comparing lines (f) and (full), it seems that the specific contribution of PBE is not enough important to explain the baseline good results
  • [Q2] did the author conducted ablation to evaluate the specific merit of the ROH? In other words, what would be the results with rk=1r_k=1 in equation (4) and rc=1r^c=1 in equation (5)?

Minor questions:

  • [MQ1] in equation (7), Does "clip(.,.,.)" a cosine similarity between the CLIP embeddings? what is ϵ\epsilon? It seems (from appendix A.2) that it allows to stabilize the normalization: reporting the exact formulation in this appendix, and studying its importance would have been appreciable.

Minor remarks (not to be addressed in the rebuttal)

  • DRF and LCP are used on line 203 before they are defined. At this level, it would be better to give an overview without using acronym, even BPE and LGRA which meaning has been given in section 1
  • the projections ϕ\phi on line 215 (attr and cls) have not the same typography as that on Fig. 2
  • in equation 4 and 5, it should be clarified that hkh_k and hch^c are in Rda+1\mathbb{R}^{d_a+1} and Rdc+1\mathbb{R}^{d_c+1} e.g with Lda⊂Rda+1\mathbb{L}^{d_a}\subset\mathbb{R}^{d_a+1}
  • on line 246, 247, the text alternates between notation that relate to yy ( eybe_y^b, θb,y\theta_{b,y}, νb,y\nu_{b,y}) and other that do not (μb\mu_b, θbmax\theta^{max}_b), which makes reading tedious.
    • at least it should start with "let eybe_y^b be the normalized descriptor for class yy..."
  • introducing the notation ece^c and eae^a while they are only the l2l_2 normalized version of aa and cc does not help either.
    • by the way, the ea/ce^{a/c} are not explicitly defined, it is only the Fig. 2 that indicates they are aa and cc whici are l2l_2-normalized.
  • the motivation for PBE (line 105) should be reminded and detailled in the dedicated section (3.4)
  • What means "ROH" on line 258?? It is only explained in the caption of Figure 4, line 405! Possibly, one can guess it from Figure 2. It should be introduced on line 211, or line 221 (after Δ\Delta introduction). In any case, it would be nice to explain what it is (in which range it lies...)
  • In section 4.5, the experiment leading to Figure 4 should be explained in the text before refeering to its results
Reviewer_Yv1rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents HyGLoA, a dual-granularity zero-shot learning framework operating in two independent hyperbolic (Lorentz) spaces. The global branch aligns an image's CLS token with class-name prototypes, while the local branch aggregates image patch tokens via Lorentzian centroid pooling and aligns them with attribute prototypes. The authors propose Local Geodesic-Ray Alignment (LGRA), derived from hyperbolic entailment cones, to capture the asymmetric relationship between local visual patches and complete attribute prototypes. Additionally, the method employs Prototype Bundle Expansion (PBE) and direction-radius decoupled projection. The two branches ultimately make fused decisions at the probability level.

Critical strengths

The paper proposes a dual-granularity hyperbolic space modeling framework (HyGLoA), which models global class-level semantics and local attribute-level evidence separately in two independent Lorentzian hyperbolic spaces. This approach of combining granularity-specific semantic information with the high-capacity properties of hyperbolic geometry possesses a certain degree of novelty in the ZSL domain.

The proposed Local Geodesic-Ray Alignment (LGRA) mechanism is derived from the zero-aperture limit of the hyperbolic entailment cone. By penalizing the deviation from the prototype's origin-facing ray, it elegantly captures the asymmetric relationship between local evidence and complete semantics in a way that forces local visual features closer to the origin without needing a predefined cone angle.

The experimental section conducts exhaustive ablation studies on multiple proposed components (such as hyperbolic geometry, PBE, DRF, LCP, and the direction of LGRA), as shown in Tables 2 and 4, effectively validating the contribution of each component to the overall system performance.

Critical weaknesses

The paper emphasizes the importance of the radius-offset head (ROH) and direction-radius factorization (DRF) in controlling spatial distribution. However, in the highly critical ablation experiments, removing the ROH (variant g in Table 4) only causes the GZSL Harmonic mean (H) on the CUB dataset to drop from 78.3 to 78.1, on SUN from 67.7 to 67.5, and on AWA2 from 95.2 to 95.0 (a mere 0.2 percentage point drop across the board). The authors also admit in the qualitative analysis that the classification gain can remain small. In ZSL experiments, a 0.2% difference falls entirely within the range of random seed fluctuations or deep learning model training noise. For an ICLR-level conference, if a core module claiming significant architectural innovation only yields marginal improvements without the support of error bars, its effectiveness is highly questionable.

In Table 1, the frozen CLIP baseline achieves a CZSL accuracy of 97.1% on AWA2, while HyGLoA achieves 97.0%. More critically, in the ablations of Table 4, if only the CLS branch is retained using Euclidean distance (variant k), the GZSL H value on AWA2 reaches 94.6%, whereas the full complex hyperbolic dual-branch model scores 95.2%. This implies that the stacking of complex mechanisms like dual branches, hyperbolic mappings, PBE, and LGRA provides almost no meaningful gain on AWA2 compared to a simple Euclidean global feature (only a +0.6% H value improvement). This strongly suggests that the model heavily relies on CLIP's strong pre-trained priors on this dataset, rather than the proposed hyperbolic and local attribute alignment mechanisms.

The motivation for LGRA comes from hyperbolic entailment cones. The paper simplifies the computation by approximating the cone angle to zero (limit derivation), thus turning it into an alignment along a ray. Geometrically, however, this means completely stripping away the tolerance of the cone originally intended to accommodate semantic diversity or intra-class variance. For local image patches with rich diversity (e.g., various shapes of bird wings or beaks), forcing them to converge onto an extremely strict 1D ray intuitively contradicts the need to preserve feature diversity in zero-shot learning. Moreover, as mentioned in the first point, even when the ROH controlling radius flexibility is removed (variant g), the model's performance hardly drops. This further exposes that this strict ray constraint may not actually play the claimed role in hierarchical modeling, but merely exists as a generic regularization term.

Questions for authors

Regarding the minuscule 0.2% performance drop caused by variant g (removing ROH) in Table 4, can the authors provide the mean and standard deviation across multiple random seed runs? If this difference is not statistically significant, can we conclude that direction-radius factorization (DRF) is numerically unnecessary?

On the AWA2 dataset, why does an extremely simplified Euclidean baseline with only the CLS branch (variant k) achieve an H value of 94.6%? Does this imply that the features of AWA2 are completely dominated by CLIP's pre-trained global features, rendering the proposed complex local attribute branch and hyperbolic geometry redundant?

Does strictly compressing the aperture of the hyperbolic entailment cone to zero (the basis for deriving LGRA) overly penalize reasonable intra-class variance of local features? Have the authors tried retaining a learnable or very small non-zero cone angle, and what are the specific performance differences between that and the current zero-aperture limit approach?

Reviewer_c6FdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyGLoA, a dual-granularity hyperbolic framework for zero-shot learning. The authors use two separate Lorentz spaces to model global class semantics and local attribute evidence. The global branch aligns CLS tokens with class-name prototypes. The local branch aligns pooled patch features with attribute prototypes. The authors further introduce Local Geodesic-Ray Alignment (LGRA), direction-radius factorization, Lorentzian centroid pooling, and prototype bundle expansion. The method combines both branches through probability-level fusion. Experiments on CUB, SUN, and AWA2 show strong GZSL results and improvements over the compared methods.

Critical strengths

The authors explore hyperbolic geometry for both global and local semantic alignment. The separation of class-name and attribute semantics makes sense for ZSL. The proposed LGRA also offers an interesting way to model asymmetric visual-semantic relations.

Critical weaknesses
  1. The overall novelty is somewhat limited. The authors mainly combine existing ideas such as hyperbolic embeddings, dual-branch learning, prototype transformation, and probability fusion. LGRA is the main new component, but it is basically a zero-aperture modification of existing hyperbolic entailment constraints.
  2. The paper does not clearly isolate the benefit of hyperbolic geometry itself. The Euclidean baseline also removes direction-radius factorization and Lorentzian pooling, so several important changes happen together. The authors should compare Euclidean and hyperbolic spaces with more closely matched architectures and training settings.
  3. The improvement from several proposed components is quite small. Removing direction-radius factorization only reduces the harmonic mean by 0.2 points on all three datasets. Removing prototype bundle expansion also causes less than one point drop, which makes their practical contribution less convincing.
  4. The experimental comparison is not fully fair. Most conventional ZSL baselines use different backbones and pretraining settings, while HyGLoA fine-tunes a pretrained CLIP model. The frozen CLIP baseline also does not account for the benefit of finetuning, so the reported gains cannot directly show how much the proposed geometry actually contributes. The authors should include a fine-tuned CLIP baseline under the same training budget.
  5. The geometric motivation still lacks strong justification. The authors assume pooled local evidence should lie closer to the origin than its complete attribute prototype, but this assumption may not hold for all attributes or images. LGRA also does not strictly guarantee that visual features stay between the origin and prototype. The paper needs stronger evidence that the learned radial structure captures meaningful semantic specificity rather than simply improving classification through an additional regularization loss.
Questions for authors

Please address above issues.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.