REVIEW 4 major objections 4 minor 1 cited by
Continual Hyperbolic Learning of Instances and Classes
T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read By embedding the instance–class hierarchy in hyperbolic space, HyperCLIC lifts continual instance accuracy from 20.05% to 41.76% on EgoObjects, with stronger class-level recognition and milder mistakes.
desk verdict A genuinely new task and large empirical gains on EgoObjects, but the hyperbolic-specific claim is under-supported because the method bundles a fixed future-class hierarchy with hyperbolic geometry and then tests with Euclidean inference. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the precomputed set of hyperbolic prototypes for every node of the class–instance tree. It is built in three passes: the Poincaré distance loss of Nickel & Kiela places nodes so that hyperbolic distance tracks graph distance, the max-margin entailment-cone loss of Ganea et al. forces each child's embedding to lie inside its parent's cone, and a separation loss, borrowed from Long et al., stops the instance prototypes from collapsing together. During continual training the Euclidean backbone output is mapped into the ball by the exponential map, the hyperbolic logit for label $y$ is $h(z,y) = -d_{\mathbb{B}}(z, P_y)/\tau$, and both the classification and distillation losses are cross-entropies over these hyperbolic logits. This mechanism is what is supposed to let one feature space recognize instances, classes, and superclasses simultaneously while resisting catastrophic forgetting.
What would settle it
Train the same iCaRL-style pipeline on EgoObjects with the hyperbolic prototype loss replaced by an ordinary Euclidean prototype loss — identical distillation, exemplars, and nearest-mean classifier — and compare instance-, class-, and superclass-level accuracy plus LCA distance; if the Euclidean variant matches HyperCLIC, the reported gains are not attributable to hyperbolic geometry. A second check is to run nearest-exemplar-mean classification in hyperbolic space at test time and see whether it at least ties the Euclidean version; if it does not, the Euclidean representations the deployed model uses are not in fact aligned with the hierarchy.
Extended reading notes
Core claim
HyperCLIC treats instances, classes, superclasses, and ancestors as nodes of one tree and embeds that tree in advance into the Poincaré ball, a model of hyperbolic space in which the distance between two prototypes mirrors the graph distance between their nodes. Visual features from a Euclidean backbone are pushed into the same ball by the exponential map $\exp_0$, and the logit for a label is the negative hyperbolic distance to that label's prototype, scaled by temperature $\tau = 0.1$. Training combines a hyperbolic classification loss for the current task with a hyperbolic distillation loss over exemplars of previous tasks, following the iCaRL recipe but with every logit computed hyperbolically. On EgoObjects this yields 41.76% instance-level accuracy from scratch versus 20.05% for iCaRL, 45.91% versus 21.39% at class level, and 48.04% versus 22.24% at superclass level, with the lowest common-ancestor distance (4.93) and lowest forgetting (4.17%); with a pretrained backbone the instance accuracy is 84.81% versus 81.63%, and the paper reports $p < 0.05$ against iCaRL and DER over five seeds. The paper's central claim is that aligning continual classification and distillation with a precomputed hyperbolic hierarchy produces this joint multi-granularity performance, and that even the mistakes are hierarchically mild.
Load-bearing premise
The method assumes that Euclidean nearest-exemplar-mean classification at test time is already aligned with the hyperbolic prototypes used during training, so the hyperbolic geometry never needs to be computed at inference — an alignment the paper asserts in Section 4.2 but does not quantitatively verify.
Editorial extensions
If this is right
- If a single model can classify at instance and class level at once, systems like robotics and self-driving perception gain a graceful-degradation property: a wrong instance prediction can still land in the correct class, which is precisely what the lower lowest-common-ancestor distance — the average edge distance to the deepest shared ancestor of prediction and ground truth — quantifies.
- The continual hierarchical metrics proposed here — instance-, class-, and superclass-level accuracy plus LCA distance — give the field a standard way to report not only how often a model is right but how far wrong it is when it errs.
- Because the hierarchy is embedded once before training, HyperCLIC needs no hierarchical information at test time and adds no extra inference computation; the paper shows the gains hold across backbones from ResNet34 to ViT-B/16.
- HyperCLIC's stability across 15- and 37-task streams, with 4.17% forgetting from scratch, suggests that prototype-anchored hyperbolic distillation slows the erosion of fine-grained instance knowledge that class-level continual learners typically sacrifice.
Reading between the lines
- The paper compares against baselines that were given no hierarchy at all; a test the authors leave implicit is whether iCaRL or DER closes the gap once they receive the same WordNet tree in Euclidean form, which would separate the contribution of hyperbolic geometry from the mere use of hierarchical information.
- Because the hierarchy is built automatically by matching class names to WordNet synsets, the two-stage scheme transfers to any dataset with a taxonomy; a natural extension is to test it on deeper ImageNet-style trees, where the class- and superclass-level metrics have more levels to climb and hierarchy errors can be scored more finely.
- The paper positions its work as complementary to hierarchical task-ordering methods; combining a coarse-to-fine task schedule with hyperbolic prototypes could plausibly compound the gains, since ordering and geometry would then push representations toward the same tree.
- If the Euclidean-alignment assumption used at test time were verified explicitly — say, by reporting the correlation between Euclidean feature distances and hyperbolic prototype distances — the method would become easier to adopt, since the exponential map could be dropped at inference with confidence.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper defines a new continual learning task in which a model must recognize both object instances and object classes across a sequence of tasks. It proposes HyperCLIC, which first embeds a WordNet-derived class-instance hierarchy as fixed prototypes in the Poincaré ball, then trains a feature extractor with a hyperbolic prototype-based classification loss and a hyperbolic distillation loss. Evaluation on EgoObjects with 15, 6, and 37 tasks reports large gains over iCaRL, DER, and other baselines on instance-, class-, and superclass-level accuracy, LCA distance, and forgetting. The appendix provides 5-seed averages and significance tests against iCaRL and DER.
Significance. If the causal role of hyperbolic geometry were established, this would be a meaningful contribution: a new joint instance-and-class continual learning formulation, a practical method that works across multiple granularities, and evaluation metrics for hierarchical mistake severity. The paper ships code and metadata, runs multi-seed experiments with p-values, and provides thorough ablations over temperature, loss weighting, distillation losses, and backbone architectures. The main weakness is that the experiments do not currently isolate the hyperbolic mechanism from the hierarchical prototype prior, so the central mechanistic claim is supported only partially by the empirical results.
major comments (4)
- [Sections 4.1-4.2, Tables 1-2] The comparison does not isolate the hyperbolic mechanism. HyperCLIC differs from the baselines in at least two bundled ways: it uses fixed prototypes for every node in the full EgoObjects hierarchy, including instance classes that will only appear in future tasks (Eq. (7) computes logits against all |C| prototypes in every task), and it places those prototypes in hyperbolic space. A Euclidean control that uses the same hierarchical prototypes, the same softmax over all prototypes, and the same distillation loss is needed to determine whether the from-scratch gain of 41.76% versus 20.05% for iCaRL comes from hyperbolic geometry or from the hierarchical prototype prior alone. The statement in Section 5 that 'there is no non-trivial way to integrate hierarchical structure into their existing frameworks' is an argument, not a substitute for such a control.
- [Section 4.2, inference paragraph] Test-time predictions use Euclidean nearest-exemplar-mean classification, not the hyperbolic logits used in training. The paper asserts that 'the Euclidean representations are already aligned with the hierarchy' but provides no quantitative evidence for this alignment. Since the deployed decision rule is Euclidean, the reported gains cannot be attributed to the hyperbolic training objective unless this transfer is measured, for example by comparing hyperbolic-versus-Euclidean logit accuracy on the test set or by measuring the correlation between Euclidean distances and graph distances in the hierarchy.
- [Section 3 and Algorithm 2] The method uses the complete set of future instance labels during every task: the hyperbolic prototypes for all instance nodes are precomputed from the full hierarchy in Section 4.1, and Algorithm 2 computes classification logits against all prototypes, including classes not yet observed. This is a protocol advantage over the class-incremental baselines, which do not have access to future class labels or their geometric layout. Please clarify whether this is intended as a permitted prior, and ideally include an ablation in which prototypes are added only when a class first appears, to show that the gains persist without future-class information.
- [Appendix A.3, Tables 8-9] The main text (Table 1) reports a from-scratch instance accuracy of 41.76% for HyperCLIC, while Table 9 reports a 5-seed average of 38.55±2.37 (and 21.49±0.91 for iCaRL). The paper does not state whether Table 1 is a single seed or the best seed, nor does it explain how the p-values in Table 8 relate to these numbers. Please report mean±std in the main table or otherwise reconcile the headline number with the averaged result.
minor comments (4)
- [Section 4.1, Eq. (3)] The formula for the Poincaré loss L_Poincaré is missing from the main text; it appears only in Algorithm 1. Please insert the equation and its definition of the negative set N(u) at the appropriate place.
- [Equation (8)] The notation in Eq. (8) uses y_i both as the target for sample i and as the summation index over classes; please use separate symbols for the sample index and the class index to avoid confusion.
- [Section 5, Hierarchies paragraph] There is a typo in 'benhcmark'; additionally, the description of hierarchy construction could state more explicitly how the 91 classes without instances were removed and how the resulting tree depth of 12 relates to the 15-task split.
- [Section 5, Qualitative examples] The sentence 'HyperCLIC makes more instance-level mistakes' appears to contradict the higher instance accuracy reported elsewhere; please rephrase to clarify that this refers to the specific examples shown rather than the overall accuracy.
Circularity Check
No significant circularity: HyperCLIC's central instance-level results are external benchmark outcomes, and the hierarchical-metric concern is a design confound rather than a circular derivation.
full rationale
The paper's derivation chain is not circular. Stage 1 constructs fixed hyperbolic prototypes by minimizing Poincaré, entailment, and separation losses over the WordNet-derived tree T (Eqs. 3-5, Algorithm 1); Stage 2 trains a feature extractor with hyperbolic classification and distillation losses (Eqs. 6-10, Algorithm 2); and the final test-time decision rule is Euclidean nearest-exemplar-mean classification, which does not even use the hyperbolic prototypes. The headline empirical claims are instance-level accuracies (41.76% vs 20.05% from scratch; 84.81% vs 81.63% pretrained) against non-hierarchical baselines iCaRL and DER on EgoObjects—these are external benchmark numbers, not algebraic consequences of the method's inputs. The reader's concern that class-, superclass-, and LCA metrics are computed from the same WordNet hierarchy used to construct the prototypes is a legitimate experimental-design confound: any hierarchy-aware method would be expected to score higher on those metrics, so they do not independently prove that hyperbolic geometry, rather than the hierarchical prototype prior, causes the gains. However, confounding is not circularity: the reported metric values are empirical outcomes of trained models, not identities forced by the equations. Also, Section 4.2 asserts without quantitative evidence that 'the Euclidean representations are already aligned with the hierarchy,' which weakens the causal attribution to hyperbolic geometry, but this is an unsupported shortcut rather than a circular step. Self-citations (e.g., Ghadimi Atigh et al. 2021 for sibling/cousin accuracy, Long et al. 2020 for separation loss) are not load-bearing because the relevant equations are given explicitly in the paper. No fitted parameter is renamed as a prediction, no uniqueness result is imported from the authors, and no ansatz is smuggled in solely via self-citation.
Assumptions & free parameters
free parameters (4)
- Temperature tau =
0.1
- Loss balancing weight lambda =
0.5
- Poincare curvature c =
1
- Fine-tuned layers =
all layers for HyperCLIC; classifier-only for iCaRL in pretrained comparisons
assumptions (5)
- domain assumption The joint label structure of instances, classes, superclasses, and other nodes is a tree with each instance a leaf having exactly one parent class.
- domain assumption The WordNet-derived hierarchy for EgoObjects is an accurate ground-truth hierarchy, and manual insertions plus pruning of single-child nodes do not introduce harmful noise.
- standard math Poincare ball embeddings with entailment cones can embed the object tree with low distortion and preserve parent-child partial order.
- domain assumption Backbone features lie in the tangent space at the origin, so the exponential map produces hyperbolic representations comparable to the precomputed prototypes.
- ad hoc to paper Euclidean nearest-exemplar-mean inference is consistent with the hyperbolic training objective because Euclidean representations are already aligned with the hierarchy.
Cite this review
Pith. "Pith review of Continual Hyperbolic Learning of Instances and Classes." pith.science (2026). https://pith.science/paper/4W77OL6B
@misc{pith2026250610710,
author = {Pith},
title = {Pith review of: Continual Hyperbolic Learning of Instances and Classes},
year = {2026},
howpublished = {\url{https://pith.science/paper/4W77OL6B}},
note = {Machine review of arXiv:2506.10710}
}
read the original abstract
Continual learning has traditionally focused on classifying either instances or classes, but real-world applications, such as robotics and self-driving cars, require models to handle both simultaneously. To mirror real-life scenarios, we introduce the task of continual learning of instances and classes, at the same time. This task challenges models to adapt to multiple levels of granularity over time, which requires balancing fine-grained instance recognition with coarse-grained class generalization. In this paper, we identify that classes and instances naturally form a hierarchical structure. To model these hierarchical relationships, we propose HyperCLIC, a continual learning algorithm that leverages hyperbolic space, which is uniquely suited for hierarchical data due to its ability to represent tree-like structures with low distortion and compact embeddings. Our framework incorporates hyperbolic classification and distillation objectives, enabling the continual embedding of hierarchical relations. To evaluate performance across multiple granularities, we introduce continual hierarchical metrics. We validate our approach on EgoObjects, the only dataset that captures the complexity of hierarchical object recognition in dynamic real-world environments. Empirical results show that HyperCLIC operates effectively at multiple granularities with improved hierarchical generalization.
Figures
Figures from the paper (2 more)
Forward citations
Cited by 1 Pith paper
-
After the Euclidean Highway: Hyperbolic Expert AI as the Next Innovation
Hyperbolic losses at the loss layer alone train stably at 18M-sample scale, while the same geometry inside a trainable adapter is claimed to collapse training.
Reference graph
Works this paper leans on
-
[6]
Skip-gram word embeddings in hyperbolic space.arXiv preprint arXiv:1809.01498,
Matthias Leimeister and Benjamin J Wilson. Skip-gram word embeddings in hyperbolic space.arXiv preprint arXiv:1809.01498,
-
[7]
Disentangling the causes of plasticity loss in neural networks.arXiv preprint arXiv:2402.18762,
Clare Lyle, Zeyu Zheng, Khimya Khetarpal, Hado van Hasselt, Razvan Pascanu, James Martens, and Will Dabney. Disentangling the causes of plasticity loss in neural networks.arXiv preprint arXiv:2402.18762,
-
[9]
Lorenzo Pellegrini. clvision-challenge-2022 reports.https://github.com/ContinualAI/ clvision-challenge-2022/tree/main/reports,
work page 2022
-
[11]
Hyperbolic neural networks++.arXiv preprint arXiv:2006.08210,
Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++.arXiv preprint arXiv:2006.08210,
arXiv 2006
-
[12]
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,
-
[14]
Hyperbolic contrastive learning.arXiv preprint arXiv:2302.01409,
Yun Yue, Fangzhou Lin, Kazunori D Yamada, and Ziming Zhang. Hyperbolic contrastive learning.arXiv preprint arXiv:2302.01409,
-
[15]
Forward compatible few-shot class-incremental learning
Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. InCVPR, 2022a. Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning.arXiv preprint arXiv:2205.13218, 2022b. Da-Wei Zhou, Qi-Wei Wang, Zhi-H...
-
[2017]
Geoopt: Riemannian optimization in pytorch.arXiv preprint arXiv:2005.02819,
Max Kochurov, Rasul Karimov, and Serge Kozlukov. Geoopt: Riemannian optimization in pytorch.arXiv preprint arXiv:2005.02819,
arXiv 2005
Show all 15 references
-
[2018]
Leveraging hierarchical taxonomies in prompt-based continual learning.arXiv preprint arXiv:2410.04327,
Quyen Tran, Hoang Phan, Minh Le, Tuan Truong, Dinh Phung, Linh Ngo, Thien Nguyen, Nhat Ho, and Trung Le. Leveraging hierarchical taxonomies in prompt-based continual learning.arXiv preprint arXiv:2410.04327,
-
[2019]
Riemannian walk for incremental learning: Understanding forgetting and intransigence
Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. InECCV, 2018a. Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelon...
-
[2020]
Embedding text in hyperbolic spaces.arXiv preprint arXiv:1806.04313,
Bhuwan Dhingra, Christopher J Shallue, Mohammad Norouzi, Andrew M Dai, and George E Dahl. Embedding text in hyperbolic spaces.arXiv preprint arXiv:1806.04313,
-
[2021]
Apo-vae: Text generation in hyperbolic space.arXiv preprint arXiv:2005.00054,
Shuyang Dai, Zhe Gan, Yu Cheng, Chenyang Tao, Lawrence Carin, and Jingjing Liu. Apo-vae: Text generation in hyperbolic space.arXiv preprint arXiv:2005.00054,
2005 arXiv
-
[2022]
3rd continual learning workshop challenge on egocentric category and instance level object understanding.arXiv preprint arXiv:2212.06833,
Lorenzo Pellegrini, Chenchen Zhu, Fanyi Xiao, Zhicheng Yan, Antonio Carta, Matthias De Lange, Vincenzo Lomonaco, Roshan Sumbaly, Pau Rodriguez, and David Vazquez. 3rd continual learning workshop challenge on egocentric category and instance level object understanding.arXiv pre...
-
[2024]
An empirical analysis of forgetting in pre-trained models with incremental low-rank updates.arXiv preprint arXiv:2405.18069,
Simone Magistri, Joost van de Weijer, Andew D Bagdanov, et al. An empirical analysis of forgetting in pre-trained models with incremental low-rank updates.arXiv preprint arXiv:2405.18069,
-
[2025]
Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531,
Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531,
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.