acceptodds
Withdrawn from ICLR 2027

One-Step Curvature Probes Miss the Fitting Operator: Retained Capacity and Terminal Null-Space Correction for Continual Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

A one-step curvature probe evaluates an initial direction, whereas continual learners are judged after reaching comparable new-task fit. These quantities need not agree. In an overparameterized linearization, projected gradient descent converges to , and its squared-displacement inflation is exactly the reciprocal of the direction-specific retained fitting capacity . More generally, the terminal old-task quadratic ratio factorizes as , where compares curvature along the normalized endpoint directions. In the rank-one case, equals the one-step probe gain; for multiple outputs, the two gains need not agree. The terminal quadratic further decomposes into a common curvature-optimal fitting floor and an algorithm-dependent null-space excess. This geometry identifies the variable omitted by a matched-norm probe and yields terminal null-space correction, a post-hoc operation that preserves linearized new-task outputs on its Jacobian batch. In separate controlled checks, the local quadratic has slope (); projection reduces the matched-norm quadratic by while terminal forgetting changes by only . Across five common-threshold configurations, projection produces the larger signed old-task loss change in matched pairs. In the eligible-subset probe study, median gain ranges from about to (eight pairs at the upper end). The common-threshold rank sweep shows median retained fitting capacity falling from to . A fixed-rank comparison separates terminal forgetting when probe gain is approximately matched, whereas the capacity-matched contrast remains compatible with effects of either sign. On Permuted MNIST and Split CIFAR-100, terminal correction decreases signed old-task loss in method–dataset–seed pairs under the all-seed intention-to-correct analysis. Because acceptance and outcome reporting use the same held-out split, this result is descriptive and test-conditioned. Together, the results replace gain-only interpretation with a matched-fit account: in the fixed-Jacobian quadratic model, terminal cost depends jointly on retained fitting capacity, endpoint-direction curvature, and the null-space component selected by optimization.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual learning, catastrophic forgetting, stability-plasticity, curvature, gradient projection, generalized Gauss-Newton, neural tangent kernel, learning theory

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_yGGzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper argues that one-step curvature probes do not reliably predict terminal forgetting in continual learning. It proposes a theoretical analysis showing that the final old-task cost depends on retained fitting capacity, endpoint-direction curvature, and the null-space component selected by optimization.

Critical strengths

This paper provides a theoretical analysis of forgetting in continual learning and validates the theoretical predictions through experiments.

Critical weaknesses
  1. I find this paper very hard to read. Large parts of the text read like writing generated by AI and the logic is often disjointed. AI can certainly help with theoretical proofs, but a reasonable approach would be for the authors to understand the material themselves and rewrite the AI-generated exposition into something readers can follow. Otherwise, the authors' contribution is close to nonexistent.

  2. The experimental section is also confusing. The paper does not describe the experimental setup at all. For example, it does not say how many samples ER uses or what sampling strategy is adopted. It also does not say how many samples GPM uses to construct its subspace. In my own experiments, with the same number of samples, GPM forgets less than ER, which seems to contradict Table 3. This makes me question the credibility of the results.

Questions for authors

Please see Weaknesses.

Reviewer_WaorOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is completely AI generated and frankly unreadable

Critical strengths

None

Critical weaknesses

This paper is completely AI generated and unreadable. The abstract is not an abstract, the introduction is not an introduction.

Questions for authors

Why would you submit this and waste everyone's time?

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

This paper argues that conventional "one-step curvature probes" (e.g., gradient projection gains under Fisher/Hessian approximations) are unsuitable for evaluating continual learners. The authors show that when retained capacity collapses, high probe gains paradoxically accelerate catastrophic forgetting. They further decompose the terminal quadratic into a common curvature-optimal floor and an algorithm-dependent null-space excess, leading to Terminal Null-Space Correction (TNC), a post-hoc operation that removes curvature energy along the new-task Jacobian null space. The theoretical derivations are corroborated on Permuted MNIST and Split CIFAR-100 across five architecture-dataset configurations.

Critical strengths
  1. By mathematically linking the fitting operator, retained capacity, and parameter displacement inflation, the work offers a potential explanation for why locally protective orthogonal projections often fail in long-run practice.

  2. The authors design a range of controlled experiments that isolate retained capacity from probe gain.

Critical weaknesses
  1. The paper is poorly written. For example, the abstract and introduction are filled with mathematical notation, some of which is undefined, and they lack an intuitive explanation of the theory's practical implications. In addition, the experimental analysis tends to repeat the reported results rather than providing in-depth analysis and insights.

  2. The practical role of the proposed theory and method is unclear. The core propositions and proofs rely strictly on a fixed new-task Jacobian and a quadratic old-task cost, which may not transfer directly to practical deep neural network architectures. Moreover, TNC requires retaining and operating on rank-approximated old-task curvature matrices and new-task Jacobian products, which increases peak memory overhead and training time, making it difficult to apply to large-scale modern neural networks.

Questions for authors
  1. In deep neural network architectures that experience substantial representation drift during new-task training, how significantly does Jacobian drift degrade the validity of cP(r)c_P(r) and the null-space invariance of TNC?

  2. How does TNC scale as the model size or the number of tasks increases?

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

Related papers

Discussion (0)

Sign in to comment.