acceptodds
Under review as a conference paper at ICLR 2027

AT-LPC: Reliability-Tempered Log-Partition Calibration for Continual Local Learning

Abstract

Continual local learning provides a modular alternative to end-to-end optimization by assigning learning objectives to individual components. However, the class-score channels produced by these components can lose a consistent comparison reference as learning progresses through a nonstationary stream. We identify this failure mode as temporal score incomparability: within-task discrimination can remain useful while scores learned at different stages become unreliable under global all-class comparison. We diagnose this phenomenon in score space and introduce Automatic Tempered Log-Partition Calibration (AT-LPC), a replay-based post-update calibration method that restores a shared score reference without retraining the classifier. AT-LPC evaluates every class-score channel on a common replay memory, estimates class-wise tempered log partitions, and selects a shared temperature using robust score scale and effective sample size (ESS). This reliability-aware tempering suppresses poorly supported score-tail effects while retaining stronger distributional correction when replay evidence is sufficient. In five-seed one-pass experiments, AT-LPC improves final accuracy from 15.64% to 26.90% on CIFAR-10 and from 4.57% to 8.48% on CIFAR-100, while remaining compatible with fixed-schedule updates, blurry arrivals, recent-prior adjustment, and alternative local score models. These results establish score comparability and replay reliability as complementary factors in the stability-plasticity behavior of continual local learning.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Temporal Score Incomparability, Log-Partition Calibration, Replay-Based Calibration, Reliability Tempering, Class-Score Alignment, Online Adaptation

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

The paper studies "temporal score incomparability" in continual local learning, assuming a Forward-Forward-style training as new classes arrive in a class-incremental stream. It proposes AT-LPC, a replay-based post-hoc calibration that subtracts a per-class tempered log-partition statistic from scores, with the temperature automatically selected. Experiments are run on CIFAR-10/100 and Tiny-ImageNet-200 in a one-pass class-incremental streaming. It shows accuracy gains over uncalibrated local experience replay.

Critical strengths

The paper studies continual learning in the specific case of forward-forward training. Even if it's not clear whether it's the way to go, it advances the field in an original direction

Critical weaknesses

I see two key critical weaknesses:

  1. Novelty and significance are not established. Classifier score incomparability is a well-known problem in continual learning, and the paper's claimed novelty rests on it being specific to the Forward-Forward/local-learning case but it never clarifies what is actually specific to this setting, nor why naively adapting existing classifier-correction methods (LUCIR, SDC, AANets, FOSTER, DATS, none of which are discussed) would be insufficient. This is consequential because, without this, we cannot determine whether the paper offers a genuine contribution or is applying a known technique to a narrow variant of an already well-studied problem. My ability to judge this is further limited by the overall clarity of the writing, which obscures the paper's positioning and makes a confident assessment difficult.
  2. The evidence needed to substantiate the central claim is weak, hidden, or missing. The one ablation that would justify the method's necessity that existing BiC/Weight-Alignment-style corrections barely help in this setting (Table 7) is relegated to the supplementary material with no seeds or error bars (unlike the five-seed main results) and no explanation for this behavior. Similarly, the one direct comparison to the closest prior work (SFFA) is single-seed, hidden in the appendix, and reaches near-chance accuracy without clear discussion. No classical regularization-based CL baselines (EWC, SI, GEM/A-GEM) or recent classifier-correction methods are tested either. This is consequential because the paper's core empirical claim that this specific correction is necessary and effective in this setting is not adequately supported by complete, well-explained, and properly powered evidence.
Questions for authors

I have numerous comments that are not all critical, but altogether they make the paper clearly not ready for publication

  • Writing. Overall, the writing is unclear, and substantial work is needed to make the paper easily understandable. This may stem from heavy reliance on AI for writing. I have tried to give my best assessment of the paper, but that assessment is bounded by how much I could understand of the method and its positioning given the current state of the writing.
  • Background (Sec. 3.1). The paper does not provide the necessary background to understand the fundamentals of Forward-Forward techniques. Describing the class-incremental stream setting explicitly would also help.
  • Motivation for the setting. The paper targets a very niche setting: class-incremental + streaming + Forward-Forward. It fails to explain why this specific combination is particularly relevant.
  • Figure 2 is not comprehensive:
    • The axis captions are not clear to me.
    • What does each point represent? At what point in training is it taken?
    • What is the meaning of the different colors?
    • I assume the colored lines below the subtitle have no meaning — if so, I would drop them.
  • Figure 3 is not very informative:
    • It lists several parts of the method without connecting them, so the overall pipeline is missing.
    • Listing all the equations without explanation doesn't help; I'd recommend dropping most of the equations from the figure and instead focusing on explaining the overall pipeline and its motivation.
    • I don't see the point of showing the three Gaussians in panel 4. What are we meant to see in the histogram in panel 5?
    • The role of caching is not clear, and the text doesn't clarify this point either.
  • Section 3.2. The motivation for this formulation is unclear. Is the formulation itself novel, or is the novelty in applying it to this task? What is the actual motivation here?
  • Missing related work on classifier calibration. Classifier calibration is a well-studied problem in continual learning, and several relevant works are missing from the discussion and potentially from the experiments:
    • LUCIR — Hou, Pan, Loy, Wang & Lin, "Learning a Unified Classifier Incrementally via Rebalancing" (CVPR 2019)
    • Semantic Drift Compensation (SDC) — Yu et al., CVPR 2020
    • AANets — Liu, Schiele & Sun, "Adaptive Aggregation Networks for Class-Incremental Learning" (CVPR 2021)
    • FOSTER — "Feature Boosting and Compression for Class-Incremental Learning" (ECCV 2022)
    • DATS — "Distance-Aware Temperature Scaling for Calibrated Class-Incremental Learning" (2025)
  • Figure 4: please include as a vector format.
  • Experimental evaluation is weak for several reasons:
    • The one baseline from the closest prior work (SFFA), which is cited, is tested only on a single seed and hidden in the appendix. It reaches near-chance accuracy, which questions the validity of this comparison, as it does not properly establish a real comparison to SFFA's own mechanism.
    • The ablation that motivates the proposed method's necessity — namely that existing BiC/Weight-Alignment corrections barely help in this setting (Table 7) — is also hidden in the supplementary material, with no seeds/error bars, unlike the five-seed main results. There is no detailed explanation for this unexpected behavior. This point and the previous one are not minor details: understanding why these baselines underperform is exactly what would justify and motivate the development of the proposed method. This should be clearly explained in the main paper.
    • Even granting that the setting is different (Forward-Forward-specific), no classical regularization-based CL baselines (EWC, SI, GEM/A-GEM) are tested, despite these being cheap to adapt given that each local block already computes its own gradient. If there is a reason this is not straightforward, the authors should include that discussion in the paper.
    • No comparison is made to the closest recent classifier-correction methods (LUCIR, SDC, FOSTER, and especially DATS, which is also a temperature-based calibration method for CIL).
    • Evaluation is performed only on small and not very diverse datasets, which seems limited.

Suggestions for future work :

  • Add background on Forward-Forward fundamentals and the class-incremental stream setting in Section 3.1.
  • Motivate why the class-incremental + streaming + Forward-Forward combination is a practically relevant setting.
  • Discuss classical CL classifier-correction methods (LUCIR, SDC, AANets, FOSTER, DATS) in related work.
  • Provide Figure 4 in vector format.
  • Broaden the experimental evaluation beyond CIFAR and Tiny-ImageNet.
Reviewer_csz8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates class-incremental learning with local learning methods, where gradients are confined within each block and predictions compares class scores aggregated across blocks. The authors identify a failure model they call temporal score incomparability where class-score channels learned at different stages of the stream drift in offset and shape. In this way, classes remain distinguishable within their task but become unreliable under all-class comparison. To solve this issue, they propose AT_LPC, a post-update calibration method that 1) evaluates all class-score channels on the replay memory, 2) computes a class-wise tempered log-partition, and 3) subtracts it as a cached bias. Experiments on online one-pass scenarios on CIFAR10, CIFAR100 and Tiny-ImageNet show improvements in final accuracy and forgetting over different baselines and settings.

Critical strengths
  1. Interesting and well-motivated problem: the distinction between within-task discrimination and all-class comparison is clearly illustrated in Figure 2.

  2. Simple and cheap method: AT-LPC does not require retraining and only add a cached bias vector at inference. Conveniently, it can also be combined with other methods (e.g., LAS).

Critical weaknesses
  1. Limited evaluation and low absolute accuracy: the main results are reported only on CIFAR10 and CIFAR100, and most of the additional analysis are carried over on CIFAR10 only. Tiny-ImageNet results are reported in the appendix, but only with a single seed. This is too limited to support the generalisation abilities of the proposed method. Furthermore, the results are obtained on top of very weak baselines. Could you please clarify how did you perform hyperparameter selection and training?

  2. Relation to recency bias: A bias toward recent classes is a well-known issue in class-incremental learning. As the authors also mention in Section 2, this is the reason why methods such as BiC have been proposed. However, the paper does not show whether temporal score incomparability differs from recency bias. What happens if you evaluate AT-LPC on a BiC-corrected trained network?

  3. Stability–plasticity trade-off: the improvement in final accuracy comes with a large drop in current-task accuracy, while the gains in anytime accuracy are smaller. This suggests the the method mainly shifts predictions toward older classes. Could you please elaborate on this?

  4. Limited dataset selection and class-balanced assumption: CIFAR10, CIFAR100 and TinyImageNet are all balanced, natural image datasets, and all experiments use a class-balance replay memory. However, in more realistic cases, the stream (and/or the memory) can be imbalanced. What happens to AT-LPC in such cases? To strengthen the paper, I would suggest the authors include datasets with different characteristics. For example, to start, you can use the long-tailed (LT) version of CIFAR. Furthermore, it would be interesting to evaluate the method on more realistic datasets from other domains, such as the ones in the MedMNIST collection.

  5. Experimental details are missing: mini-batch size, replay batch size, learning rate, and optimiser are not reported. The first two in particular are really important in online settings. Could you please specify and ablate different values of these two hyperparameters to see how the method is dependent on such factors?

Overall, I vote for rejecting the paper. The problem is interesting, but the evaluation is too limited to support the general claims of the paper. Addressing the concerns above would require substantial changes to the paper that go beyond the scope of a rebuttal. Nevertheless, I would appreciate clarifications about the points and I believe that including such suggestions would considerably strengthen the next version of the paper.

Questions for authors

See above.

Reviewer_d9FPOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper addresses continual local learning by identifying temporal score incomparability. It proposes Automatic Tempered Log-Partition Calibration (AT-LPC), a replay-based post-update calibration method that restores a shared score reference without retraining the classifier. Experiments on CIFAR-10 and CIFAR-100 demonstrate promising performance.

Critical strengths
  1. The paper offers a valuable perspective on continual local learning by investigating cross-stage class-score comparability.

  2. The proposed method achieves promising performance on CIFAR-10 and CIFAR-100.

Critical weaknesses
  1. The authors should provide appropriate citations for the baseline methods in the experimental section to adhere to academic writing standards, as the absence of references makes it unclear which methods are being compared.

2.The paper identifies score incomparability as a key failure in continual local learning, but similar biases also exist in continual learning. The authors do not sufficiently clarify what makes this phenomenon specific to local learning.

3.The experiments are conducted on CIFAR-10 and CIFAR-100. The authors should evaluate the proposed method on larger-scale datasets (e.g., ImageNet-1K) to demonstrate the robustness of the proposed method.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.