acceptodds
Withdrawn from ICLR 2027

Endpoint-Consistent Continual Learning Under Evolving Label Ontologies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual learners can reach different endpoint predictors when the same final ontology is revealed through different legal split, merge, and refinement orders. We formulate this path dependence as two coupled failures: historical labels may lack an endpoint-compatible representation, and the resulting update operators may remain functionally noncommutative. ONTOLOGY PATH CONSISTENCY represents every observed label as a marginal constraint over shared latent leaves and penalizes the functional defect between endpoint-equivalent update orders. Across CIFAR-100 superclass/subclass paths, it reaches 73.4% mean leaf accuracy and 72.0% worst-path accuracy with predictive ontology-path variance (OPV) \(2.3\times10^-3\) and path-consistency defect (PCD) \(2.9\times10^-3\). CLEO (MoOn) reaches 72.7%, 69.9%, \(7.1\times10^-3\), and \(9.4\times10^-3\); the latent-leaf-only ablation reaches 73.0%, 70.5%, \(5.9\times10^-3\), and \(7.8\times10^-3\). The paired reductions in PCD and OPV isolate endpoint-order consistency beyond hierarchy-aware representation alone.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual learning, lifelong learning, evolving label ontologies, ontology evolution, hierarchical classification, path consistency, latent label representation, catastrophic forgetting

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8CoEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is fully generated by the AI, which is significantly not satisfied the standard for publishing, so i recommend to reject.

Critical strengths

This paper is fully generated by the AI, which is significantly not satisfied the standard for publishing, so i recommend to reject.

Critical weaknesses

This paper is fully generated by the AI, which is significantly not satisfied the standard for publishing, so i recommend to reject.

Questions for authors

This paper is fully generated by the AI, which is significantly not satisfied the standard for publishing, so i recommend to reject.

Reviewer_uUAHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether two different ontology edit histories (different orders of splits, merges and refinements) that end at the same label set lead to different final models. The method has two parts: old coarse labels supervise a set of shared latent leaves through a marginal likelihood, and a PCD regularizer penalizes disagreement between two reversed virtual update orders. On CIFAR-100 superclass/subclass paths, it gets lower OPV and better worst-path accuracy than LECO, CLEO and a latent-leaf-only ablation.

Critical strengths

1.The question is nice and I haven't seen it framed this way before. The latent-leaf-only ablation in Table 1 is the right control, since it separates "fixing what old labels mean" from "fixing order sensitivity".

2.The authors are honest about scope. Section 4 says lowering the local PCD does not prove that arbitrary paths commute, and they report worst-path accuracy so that low variance can't come from every path being equally bad.

Critical weaknesses

1.The method can't really be reproduced from the paper. Section 4 gives Eq. 4 but not the full training loss, the PCD weight, the virtual step size, whether gradients go through the virtual updates, or how often PCD is computed. It's also unclear whether the final leaf set Z is known from the start. This matters, because if it is, the method may get information the baselines don't. The reproducibility statement says no code, configs or logs are included, so I have no way to check.

2.The experiments are too thin for the claims. Section 5 doesn't say how many paths were used, what the edit schedules are, or the backbone, optimizer and replay settings. Table 1 has one number per cell, with no seeds or std, on a single dataset. The gain over latent-leaf-only is 0.4 points in mean accuracy, and without variance I can't tell if that's real. The paper is about 4 pages, so there was room to include all this.

Questions for authors

1.Are the full latent leaf set and the final map M_T available before training starts? If so, do the baselines get the same information? (clarification)

2.What is the exact training objective with PCD (weight, virtual step size, anchor sampling, and whether you backprop through the virtual steps)? (clarification)

3.Could you list the paths and schedules, and report mean ± std over several seeds for Table 1, together with the extra training cost of PCD? (suggestion)

Reviewer_rSD5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies path dependence in continual learning when class ontologies evolve through splits, merges, and refinements. It examines whether two legal edit histories that reach the same final ontology produce the same final predictor. The paper hypothesizes two sources of path dependence: historical labels that lack an endpoint-compatible representation, and update operators that do not commute. The proposed method represents each observed label as a marginal constraint over a shared latent leaf space through a marginalization matrix. It adds a regularizer that minimizes a Jensen-Shannon path-consistency defect (PCD) for sampled legal update orders with the same endpoint. PCD compares endpoint predictions after a current-data update and a replay update are applied in both orders. The paper also defines a predictive ontology-path variance metric (OPV) over endpoint-equivalent paths. On CIFAR-100, the paper reports higher accuracy and lower path-dependence metrics than continual learning and evolving-ontology baselines.

Critical strengths
  • Studying whether different edit histories with the same final ontology yield the same predictor is a useful problem, distinct from forgetting.

  • The evaluation protocol holds the training setup fixed and varies only the ontology edit order, which is a sensible way to isolate path dependence.

  • The latent-leaf-only ablation is a helpful comparison that separates label transport from the explicit order-consistency term.

Critical weaknesses
  • The technical contribution seems small relative to known components. Marginal likelihood over compatible fine labels is a standard way to learn from coarse or partial labels in hierarchical classification. The path-consistency term is a symmetric divergence between two update orders, which is a form of consistency regularization. The paper does not discuss related work on partial-label learning or on continual learning methods that use virtual updates, so it is unclear how its components differ from these lines of work.

  • The experimental evidence is too thin to support the claims. All results come from a single table of aggregate numbers. The method directly minimizes PCD during training, so a lower PCD is expected by construction and does not on its own confirm the noncommutativity explanation. Measuring OPV on held-out edit families would make the claim more convincing. The paper also omits important technical details, such as how the endpoint-equivalent paths are constructed, the backbone, and how the virtual updates are configured. Without these details, the results are hard to interpret, especially the small mean-accuracy gain over the latent-leaf-only ablation.

  • The generality of the results is narrow. The evaluation uses one dataset with a fixed two-level hierarchy (CIFAR-100 superclasses and subclasses). The introduction motivates the problem with splits, merges, refinements, and renaming, but it is not clear which of these edit types the evaluated paths cover. Results on deeper or less regular taxonomies and larger datasets would show whether the method transfers to more realistic ontology changes.

Questions for authors
  • How is PCD computed for the baselines in Table 1?

  • How many endpoint-equivalent paths are evaluated, and which edit types do they contain?

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

Related papers

Discussion (0)

Sign in to comment.