acceptodds
Under review as a conference paper at ICLR 2027

Age of Learning: Temporal Persistence of Prediction Errors as a Learning Signal

Abstract

Current machine learning algorithms primarily rely on instantaneous signals such as loss, margin, and prediction confidence to characterize model behavior. These signals indicate how difficult a prediction is at the current optimization step, but they do not capture how long the model has remained incorrect. We study this temporal dimension of learning and introduce Age of Learning (AoL), a learning-state variable that measures the persistence of prediction errors over time. AoL increases while an error remains unresolved and resets when a correct prediction is achieved, thereby distinguishing persistent under-learning from transient mistakes. We develop AoL-based training strategies for both offline and streaming settings. In offline learning, sample-level AoL is accumulated over training and aggregated into class-level states that guide adaptive reweighting and resampling. In streaming learning, where full historical access is unavailable, we maintain lightweight class-level AoL states using current and buffered observations. Across long-tailed classification settings, AoL improves or matches standard training baselines, with larger benefits when learning difficulty persists over time. Multi-seed streaming experiments further show reproducible gains under temporally stable imbalance. Analysis of class frequency, loss, and margin shows that AoL is related to conventional difficulty measures but captures additional information about error duration. These results suggest that temporal persistence provides a useful complementary signal for characterizing and controlling learning dynamics in imbalanced and non-stationary environments.

Primary area:
optimization
Keywords:
learning dynamics, age of learning, adaptive training

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vLKXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Age of Learning (AoL) for continual learning with long-tailed distribution. AoL increased when a sample or class is misclasified, and AoL becomes 0 if the sample or class is correctly classified. The reset ability is different from conventional class weights or loss functions for long-tailed classificaiton. Experimental results on long-tailed CIFAR10 and CIFAR100 demonstrate that the proposed AoL slightly imporves the acccuracy compared with Softmax + class balanced.

Critical strengths

The concept of AoL is interesting. Specifically, the reset ability is new idea. The AoL can be used for sample or class weight and sampling.

Critical weaknesses

Experiments are insufficient to show the effectiveness of the proposed method. 1-1. In Table 1, the auhors select the better result between sample and class-level AoL. This is unfair. The authors must show both results.

1-2. In Table 1, the accuracy of the proposed AoL is not so improved compared with class balanced softmax.

1-3. There is no comparison with conventional methods for continual learning or long-tailed classification. The readers can not judge the effectiveness of your method without the comparison.

1-4. In Table 2, the comparison with class balanced softmax is missing.

1-5. In section 4.3, only the accuracies on Places365 and Imagenet-1K datasets are shown. The experimental deitals are missing. The auhtors used buffer capacity of 500,000. What if you change the capacity size? The sensitivity analsis of the size is required. There is no result of the case that memory buffer is not used, though the authors define the method without buffer in equation (12).

Questions for authors

Please respond to the Weakness I pointed out.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1-1

Dear Reviewer vLKX,

Thank you very much for taking the time to review our submission and for providing your valuable comments and generous support. We first clarify a factual misunderstanding in the summary: AoL resets to 1, not 0, upon correct classification, as defined in Eq. (2) and illustrated in Figure 1. Below, we have carefully and fully addressed all mentioned weaknesses. The corresponding changes are highlighted in the revised PDF file for your easy reference. The additional implementation of the comparison experiments is also provided in the updated supplementary material as "main_new.py."

Regarding the concern that "experiments are insufficient to show the effectiveness of the proposed method," we have substantially expanded our experimental evaluation, including additional comparisons with existing methods, complete sample-level and class-level results, multi-seed experiments on large-scale datasets, and buffer-capacity sensitivity analyses. We address each concern below.


Weakness 1-1:

In Table 1, the auhors select the better result between sample and class-level AoL. This is unfair. The authors must show both results.

Reply:

We thank the reviewer for pointing out the potential ambiguity in Table 1. To ensure fair reporting, we have revised Table 1 to consistently report class-level AoL, rather than selecting the better result between sample-level and class-level formulations. The complete results for both formulations are provided in Appendix Table 4, with class-level results outside parentheses and sample-level results inside parentheses.

Table 1: Offline long-tail classification results on CIFAR-100 and CIFAR-10 under different IFs. Top-1 test accuracy (%). All AoL results use class-level weighting.

MethodCIFAR-100 IF-400IF-100IF-10IF-1CIFAR-10 IF-400IF-100IF-10IF-1
Softmax31.5040.1057.1571.5960.3272.2086.7292.98
Softmax + CB32.0440.4057.4871.5961.3173.5986.7492.98
Softmax + AoL32.4340.5157.9871.6762.3672.6286.8093.19
Sigmoid33.0040.6557.6371.3360.1771.8586.8893.00
Sigmoid + CB32.8640.8158.4471.3364.4574.9687.2893.00
Sigmoid + AoL33.0040.8458.0771.5666.6272.7687.2793.39

For example, on CIFAR-10/IF-400, class-level Softmax+AoL achieves 62.36%, compared with 60.32% for Softmax and 61.31% for Softmax+CB. The corresponding sample-level result is 63.94%, but the main table consistently reports 62.36%. Similarly, at IF-100, the main table reports the class-level result of 72.62%, rather than the higher sample-level result of 73.74%.

By consistently reporting class-level AoL in Table 1 and providing both formulations in Table 4, we eliminate the possibility of selectively reporting whichever formulation achieves higher accuracy.

Manuscript revision. We have revised Table 1 (page 8) to consistently report class-level AoL and added the following clarification to Section 4.2 (page 8):

Table 1 consistently reports class-level AoL results, while complete results for both sample-level and class-level AoL are provided in Appendix B.1.

We have also clarified the reporting convention in Appendix B.1, Table 4 (page 18):

For non-dagger AoL rows, values outside and inside parentheses denote class-level and sample-level results, respectively.


AuthorsOpenReview ↗

Response to Weakness 1-2

Weakness 1-2:

In Table 1, the accuracy of the proposed AoL is not so improved compared with class balanced softmax.

Reply:

We appreciate the reviewer's observation. Our objective is not to consistently outperform Class-Balanced (CB) Loss (Cui et al., 2019) but to demonstrate that consecutive prediction-error duration provides an effective learning signal without requiring global class-frequency information.

For example, on CIFAR-10/IF-400, Softmax+AoL achieves 62.36%, compared with 61.31% for Softmax+CB and 60.32% for Softmax. Thus, AoL improves the original Softmax baseline from 60.32% to 62.36% (+2.04 points) and exceeds CB by 1.05 points. On CIFAR-100/IF-100, AoL achieves 40.51% compared with 40.40% for CB, demonstrating comparable performance.

Unlike CB, which constructs weights from class-frequency statistics, AoL determines training emphasis from persistent prediction errors. This distinction is particularly relevant in streaming learning, where the global data distribution may be unknown or evolving. Our additional experiments further evaluate AoL alongside CB under these conditions.

Manuscript revision. We have revised the discussion in Section 4.2 (page 8) as follows:

AoL's relative performance against CB varies across datasets, imbalance factors, and loss functions. The offline results show benefits over unweighted training in several settings, without establishing uniform superiority over CB. We therefore complement these comparisons with multi-seed streaming experiments and evaluations incorporating AoL into existing methods.

Manuscript revision. We have added the following interpretation to Appendix C (page 22, Interpretation of "complementary"):

Complementarity in this sense does not imply that AoL must produce an additive gain when mechanically combined with every existing rebalancing method. The offline experiments compare AoL and CB independently to isolate their respective weighting signals, whereas the online comparisons in Section 4.4 evaluate AoL in combination with seven adapted baselines, including CB, to assess its additional contribution under matched conditions.


AuthorsOpenReview ↗

Response to Weakness 1-3

Weakness 1-3:

There is no comparison with conventional methods for continual learning or long-tailed classification. The readers can not judge the effectiveness of your method without the comparison.

Reply:

We thank the reviewer for raising this important concern. In view of the above comments, even though conventional long-tailed methods are already represented: Table 1 explicitly compares against Class-Balanced Loss (Cui et al., 2019); we have further expanded the evaluation by conducting matched comparisons with seven online-adapted methods: CB (Cui et al., 2019), Focal Loss (Lin et al., 2017), LDAM (Cao et al., 2019), ICDW/Cyclical Weighting (Chen et al., 2025), SBS/forgetting-inspired weighting (Hacohen & Tuytelaars, 2025), EMA Error Weighting, and EMA Loss Weighting.

Each method is evaluated both independently and with AoL under identical experimental conditions on CIFAR-10 and CIFAR-100 (IF=200, Period-10, ResNet-32, K=50,000), using three random seeds (42, 1024, and 2026).

Table 3: Comparison of seven online baselines with and without AoL on CIFAR-10 and CIFAR-100 under IF=200 and Period-10. Top-1 test accuracy (%) and mean ± sample SD are reported.

MethodVariantCIFAR-10 (42)CIFAR-10 (1024)CIFAR-10 (2026)CIFAR-10 (Mean ± SD)CIFAR-100 (42)CIFAR-100 (1024)CIFAR-100 (2026)CIFAR-100 (Mean ± SD)
CBBaseline76.7875.6978.2276.90 ± 1.2746.5945.7345.5545.96 ± 0.56
CB+ AoL81.1179.1679.3079.86 ± 1.0946.5945.7345.7946.04 ± 0.48
FocalBaseline76.7074.7175.5075.64 ± 1.0042.2741.1239.8841.09 ± 1.20
Focal+ AoL79.7879.5277.8479.05 ± 1.0545.4345.4546.0145.63 ± 0.33
LDAMBaseline78.7677.2477.6377.88 ± 0.7942.3242.4444.3443.03 ± 1.13
LDAM+ AoL79.6778.8678.8579.13 ± 0.4746.9445.0246.5646.17 ± 1.02
ICDWBaseline82.6482.8181.8282.42 ± 0.5352.4754.1253.7053.43 ± 0.86
ICDW+ AoL83.2983.9184.2383.81 ± 0.4852.4754.1253.7053.43 ± 0.86
SBSBaseline76.0476.5376.8176.46 ± 0.3942.6244.7043.0543.46 ± 1.10
SBS+ AoL79.7978.6278.5879.00 ± 0.6944.8845.1445.9545.32 ± 0.56
EMA ErrorBaseline78.4575.8278.9077.72 ± 1.6643.4243.7043.9843.70 ± 0.28
EMA Error+ AoL80.0279.0578.9079.32 ± 0.6145.9744.0044.5744.85 ± 1.01
EMA LossBaseline78.5476.1777.8577.52 ± 1.2244.1143.5644.0243.90 ± 0.30
EMA Loss+ AoL79.7479.3980.6979.94 ± 0.6744.8545.4944.9645.10 ± 0.34

Across all three seeds, adding AoL improves 36 of the 42 comparisons and matches the remaining six, without decreasing accuracy in any comparison. These matched experiments demonstrate that AoL can complement frequency-based, difficulty-based, and historical weighting methods.

We note that these are controlled online adaptations under a fixed-label-space streaming protocol, rather than full task-incremental or class-incremental continual-learning systems.

Manuscript revision. We have added the following comparison to Section 4.4 (page 9), with results reported in Table 3 (page 10):

To assess whether AoL complements existing learning signals more comprehensively, we further implement and compare seven online-adapted baselines based on CB (Cui et al., 2019), Focal (Lin et al., 2017), LDAM (Cao et al., 2019), ICDW (Chen et al., 2025), SBS (Hacohen & Tuytelaars, 2025), EMA Error Weighting, and EMA Loss Weighting.

Each method is evaluated with and without AoL under identical conditions on CIFAR-10 and CIFAR-100, using IF-200, Period-10, ResNet-32, SGD, a buffer of K = 50,000, and three random seeds.

We have also detailed the online baseline implementations in Appendix B.2 (pages 17–19) and provided the comparison implementation in the supplementary material as "main_new.py."


Reviewer_97sTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Age of Learning (AoL), which tracks consecutive prediction errors and uses smoothed sample or class-level states for reweighting or resampling. In streaming settings, it tracks how long class accuracy remains below a threshold. Experiments cover offline long-tailed CIFAR and synthetic streams built from CIFAR, Places365, and ImageNet, with the clearest gains under temporally persistent imbalance.

Critical strengths
  • Error duration is an intuitive, promising complement to instantaneous difficulty, with a clear motivation for persistent under-learning.
  • The AoL signal is simple to integrate; the streaming state requires only O(C) additional storage beyond the observation buffer.
Critical weaknesses
  • The distinct methodological contribution is not isolated. AoL adapts an error-duration idea related to AoII and learning-history signals such as example forgetting. Matched comparisons with EMA error/loss weighting and forgetting-based weighting are missing.

  • Performance evidence is mixed. In Table 1, CIFAR-100/IF-100 gains over CB are only 0.11 and 0.03 points for Softmax and Sigmoid; CIFAR-10/IF-100 Sigmoid is 2.20 points worse.

Questions for authors
  • Do existing comparisons distinguish consecutive-error duration from EMA error/loss or forgetting-based weighting?
  • How does the median-threshold online state relate to offline error persistence, and how sensitive is it to the threshold?
  • Were tables computed from final or test-selected checkpoints, and what held-out validation split was used?
  • Can existing logs quantify the extra AoL evaluation time and total runtime relative to the baseline?
7 comments
AuthorsOpenReview ↗

Response to Weakness 1

Dear Reviewer 97sT,

Thank you very much for taking the time to review our submission and for providing your valuable comments and generous support. We have carefully and fully addressed all mentioned weaknesses and questions as follows. The corresponding changes are highlighted in the revised PDF file for your easy reference. The additional implementation of the comparison experiments is also provided in the updated supplementary material as "main_new.py."


Weakness 1:

The distinct methodological contribution is not isolated. AoL adapts an error-duration idea related to AoII and learning-history signals such as example forgetting. Matched comparisons with EMA error/loss weighting and forgetting-based weighting are missing.

Reply:

We sincerely thank the reviewer for highlighting the need to distinguish AoL from existing temporal learning signals. AoL represents consecutive unresolved-error duration through an accumulation-and-reset mechanism, followed by temporal smoothing. EMA-error weighting summarizes recent error frequency, EMA-loss weighting summarizes recent loss magnitude, and forgetting statistics characterize transitions between correctness states. AoL instead emphasizes how long an incorrect or underperforming state has persisted.

To directly address the requested comparisons, we evaluate seven online-adapted methods with and without AoL under identical settings (IF=200, Period-10, ResNet-32, K=50,000) across three seeds. Among them, the forgetting-based Online SBS adaptation uses historical class accuracy, while Online EMA Error and EMA Loss maintain smoothed class-level error rates and losses, respectively. The three methods most relevant to the reviewer's concern show the following mean accuracy improvements:

  • Online EMA Error Weighting: AoL improves mean accuracy from 77.72 ± 1.66% to 79.32 ± 0.61% (+1.60 points) on CIFAR-10 and from 43.70 ± 0.28% to 44.85 ± 1.01% (+1.15 points) on CIFAR-100.

  • Online EMA Loss Weighting: AoL improves mean accuracy from 77.52 ± 1.22% to 79.94 ± 0.67% (+2.42 points) on CIFAR-10 and from 43.90 ± 0.30% to 45.10 ± 0.34% (+1.20 points) on CIFAR-100.

  • Forgetting-based weighting (Online SBS; Hacohen & Tuytelaars, ICML 2025, "Predicting the Susceptibility of Examples to Catastrophic Forgetting"): Our streaming adaptation uses historical class accuracy to construct class-level weights. Adding AoL improves mean accuracy from 76.46 ± 0.39% to 79.00 ± 0.69% (+2.54 points) on CIFAR-10 and from 43.46 ± 1.10% to 45.32 ± 0.56% (+1.87 points) on CIFAR-100.

Table 3: Comparison of seven online baselines with and without AoL on CIFAR-10 and CIFAR-100 under IF=200 and Period-10. Top-1 test accuracy (%) and mean ± sample SD are reported.

MethodVariantCIFAR-10 (42)CIFAR-10 (1024)CIFAR-10 (2026)CIFAR-10 (Mean ± SD)CIFAR-100 (42)CIFAR-100 (1024)CIFAR-100 (2026)CIFAR-100 (Mean ± SD)
CBBaseline76.7875.6978.2276.90 ± 1.2746.5945.7345.5545.96 ± 0.56
CB+ AoL81.1179.1679.3079.86 ± 1.0946.5945.7345.7946.04 ± 0.48
FocalBaseline76.7074.7175.5075.64 ± 1.0042.2741.1239.8841.09 ± 1.20
Focal+ AoL79.7879.5277.8479.05 ± 1.0545.4345.4546.0145.63 ± 0.33
LDAMBaseline78.7677.2477.6377.88 ± 0.7942.3242.4444.3443.03 ± 1.13
LDAM+ AoL79.6778.8678.8579.13 ± 0.4746.9445.0246.5646.17 ± 1.02
ICDWBaseline82.6482.8181.8282.42 ± 0.5352.4754.1253.7053.43 ± 0.86
ICDW+ AoL83.2983.9184.2383.81 ± 0.4852.4754.1253.7053.43 ± 0.86
SBSBaseline76.0476.5376.8176.46 ± 0.3942.6244.7043.0543.46 ± 1.10
SBS+ AoL79.7978.6278.5879.00 ± 0.6944.8845.1445.9545.32 ± 0.56
EMA ErrorBaseline78.4575.8278.9077.72 ± 1.6643.4243.7043.9843.70 ± 0.28
EMA Error+ AoL80.0279.0578.9079.32 ± 0.6145.9744.0044.5744.85 ± 1.01
EMA LossBaseline78.5476.1777.8577.52 ± 1.2244.1143.5644.0243.90 ± 0.30
EMA Loss+ AoL79.7479.3980.6979.94 ± 0.6744.8545.4944.9645.10 ± 0.34

As shown in Table 3, across all seven methods, AoL improves 36 of the 42 comparisons and matches the remaining six. The consistent improvements over EMA error/loss weighting and historical-accuracy weighting provide direct evidence supporting the usefulness of duration-based emphasis alongside existing learning-history signals.

AuthorsOpenReview ↗

Response to Weakness 1

Manuscript revision. We have added the following discussion to Section 2 (page 3):

Whereas forgetting statistics count correctness transitions and Cyclical Weighting (Chen et al., 2025) combines momentum-based class accuracy with instance loss, AoL tracks consecutive-error duration through accumulation and reset, with EMA retaining information from earlier episodes. Thus, two prediction histories can have identical error rates and current correctness yet different AoL states.

Manuscript revision. We have added the following comparison to Section 4.4 (page 9), with results reported in Table 3 (page 10):

Focal operates on incoming mini-batches; ICDW replaces instance histories with current losses; and SBS uses historical class accuracy. The two EMA baselines track class-level error rates and losses, respectively. Each method is evaluated with and without AoL under identical conditions on CIFAR-10 and CIFAR-100, using IF-200, Period-10, ResNet-32, SGD, a buffer of K = 50,000, and three random seeds.

We have also detailed the online adaptations in Appendix B.2 (pages 17–19).


AuthorsOpenReview ↗

Response to Weakness 2

Weakness 2:

Performance evidence is mixed. In Table 1, CIFAR-100/IF-100 gains over CB are only 0.11 and 0.03 points for Softmax and Sigmoid; CIFAR-10/IF-100 Sigmoid is 2.20 points worse.

Reply:

We sincerely appreciate the reviewer's careful examination of our results. Table 1 evaluates whether temporal error persistence provides an effective training signal, with CB serving as an established frequency-based reference. Our goal is to demonstrate that AoL provides useful training emphasis, rather than uniformly outperforming CB in every configuration.

We acknowledge that the magnitude of improvement varies across datasets, imbalance levels, and loss functions. For instance, on CIFAR-100/IF-100, Softmax+AoL achieves 40.51%, compared with 40.40% for Softmax+CB (+0.11 points), while Sigmoid+AoL achieves 40.84%, compared with 40.81% for Sigmoid+CB (+0.03 points). On CIFAR-10/IF-100, although Sigmoid+AoL (72.76%) is below Sigmoid+CB (74.96%), it improves the original Sigmoid baseline from 71.85% to 72.76% (+0.91 points). Under stronger imbalance (IF=400), AoL improves Sigmoid from 60.17% to 66.62% (+6.45 points), exceeding Sigmoid+CB (64.45%) by 2.17 points.

More importantly, our expanded experiments in Table 3 further demonstrate that AoL can complement CB rather than compete with it. On CIFAR-10/IF-200 under Period-10, adding AoL improves mean Online CB accuracy from 76.90 ± 1.27% to 79.86 ± 1.09% (+2.96 points) across three seeds. Unlike CB, AoL derives its weights from persistent prediction errors without requiring prior knowledge of the global class distribution.

These results demonstrate that AoL provides a useful complementary signal, with benefits depending on the dataset and imbalance level.

Manuscript revision. We have revised the discussion in Section 4.2 (page 8) as follows:

AoL's relative performance against CB varies across datasets, imbalance factors, and loss functions. The offline results show benefits over unweighted training in several settings, without establishing uniform superiority over CB. We therefore complement these comparisons with multi-seed streaming experiments and evaluations incorporating AoL into existing methods.

We have also added paired Online CB comparisons in Section 4.4 (pages 9–10, Table 3), evaluating whether AoL provides complementary benefits alongside class-frequency weighting.


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

The paper introduces Age of Learning (AoL), a novel learning signal that measures the persistence of prediction errors during training. AoL captures how long samples or classes remain incorrectly classified. Moreover, the authors propose AoL-based reweighting and resampling strategies for offline learning and extend the approach to streaming settings using class-level states. Experiments on long-tailed classification benchmarks demonstrate performance improvements, especially in cases where learning difficulties persist over time.

Critical strengths

The paper proposed an interesting and fundamental research question for machine learning, and the paper presentation is quite good. The experimental results support the author’s claim.

Critical weaknesses

The paper does not sufficiently show that the error duration captures meaningful information beyond conventional difficulty measures [1] or historical error rates [2]. [1] Focal Loss for Dense Object Detection, ICCV 2017. [2] Cyclical weighting: A new training strategy for long-tailed recognition, EAAI 2025.

Questions for authors
  1. Is the evolution of prediction correctness during training approximately Markovian? Or is anything in the learning process Markovian? I think it’s a fundamental question for AoL.
  2. If you think the prediction correctness during training is approximately Markovian, can you further explain how AoL provides a useful learning signal?
  3. If you think the prediction correctness during training isn’t Markovian, can you further prove it?
3 comments
AuthorsOpenReview ↗

Response to Weakness 1

Dear Reviewer CGYp,

Thank you very much for taking the time to review our submission and for providing your valuable comments and generous support. We have carefully and fully addressed all mentioned weaknesses and questions as follows. The corresponding changes are highlighted in the revised PDF file for your easy reference. The additional implementation of the comparison experiments is also provided in the updated supplementary material as "main_new.py."


Weakness 1:

The paper does not sufficiently show that the error duration captures meaningful information beyond conventional difficulty measures [1] or historical error rates [2]. [1] Focal Loss for Dense Object Detection, ICCV 2017. [2] Cyclical weighting: A new training strategy for long-tailed recognition, EAAI 2025.

Reply:

We sincerely thank the reviewer for this important suggestion. To directly address this concern, we have conducted matched experiments with seven online-adapted methods, including the two specifically cited works:

  • [1] Focal Loss (Lin et al., ICCV 2017) corresponds to Online Focal. Unlike Focal Loss, which emphasizes instantaneous prediction difficulty, AoL measures consecutive error duration. Across three seeds, adding AoL improves mean accuracy from 75.64 ± 1.00% to 79.05 ± 1.05% (+3.41 points) on CIFAR-10 and from 41.09 ± 1.20% to 45.63 ± 0.33% (+4.54 points) on CIFAR-100.

  • [2] Cyclical Weighting (Chen et al., EAAI 2025) corresponds to Online ICDW. Our streaming adaptation retains class- and instance-difficulty weighting and the cyclical schedule, replacing persistent instance histories with current losses. Adding AoL improves mean accuracy from 82.42 ± 0.53% to 83.81 ± 0.48% (+1.39 points) on CIFAR-10 and maintains 53.43 ± 0.86% on CIFAR-100 across three seeds.

All seven methods are evaluated with and without AoL under identical settings (IF=200, Period-10, ResNet-32, K=50,000) across three random seeds.

Table 3: Comparison of seven online baselines with and without AoL on CIFAR-10 and CIFAR-100 under IF=200 and Period-10. Top-1 test accuracy (%) and mean ± sample SD are reported.

MethodVariantCIFAR-10 (42)CIFAR-10 (1024)CIFAR-10 (2026)CIFAR-10 (Mean ± SD)CIFAR-100 (42)CIFAR-100 (1024)CIFAR-100 (2026)CIFAR-100 (Mean ± SD)
CBBaseline76.7875.6978.2276.90 ± 1.2746.5945.7345.5545.96 ± 0.56
CB+ AoL81.1179.1679.3079.86 ± 1.0946.5945.7345.7946.04 ± 0.48
FocalBaseline76.7074.7175.5075.64 ± 1.0042.2741.1239.8841.09 ± 1.20
Focal+ AoL79.7879.5277.8479.05 ± 1.0545.4345.4546.0145.63 ± 0.33
LDAMBaseline78.7677.2477.6377.88 ± 0.7942.3242.4444.3443.03 ± 1.13
LDAM+ AoL79.6778.8678.8579.13 ± 0.4746.9445.0246.5646.17 ± 1.02
ICDWBaseline82.6482.8181.8282.42 ± 0.5352.4754.1253.7053.43 ± 0.86
ICDW+ AoL83.2983.9184.2383.81 ± 0.4852.4754.1253.7053.43 ± 0.86
SBSBaseline76.0476.5376.8176.46 ± 0.3942.6244.7043.0543.46 ± 1.10
SBS+ AoL79.7978.6278.5879.00 ± 0.6944.8845.1445.9545.32 ± 0.56
EMA ErrorBaseline78.4575.8278.9077.72 ± 1.6643.4243.7043.9843.70 ± 0.28
EMA Error+ AoL80.0279.0578.9079.32 ± 0.6145.9744.0044.5744.85 ± 1.01
EMA LossBaseline78.5476.1777.8577.52 ± 1.2244.1143.5644.0243.90 ± 0.30
EMA Loss+ AoL79.7479.3980.6979.94 ± 0.6744.8545.4944.9645.10 ± 0.34

Overall, incorporating AoL improves 36 of 42 comparisons and matches the remaining six. Although Online ICDW already achieves the strongest baseline performance on CIFAR-100, AoL preserves this accuracy while improving ICDW on CIFAR-10 and substantially improving Online Focal on both datasets. These results demonstrate the complementary value of error persistence beyond the two learning signals identified by the reviewer.

AuthorsOpenReview ↗

Response to Weakness 1

Manuscript revision. We have added the following discussion to Section 2 (page 3):

Whereas forgetting statistics count correctness transitions and Cyclical Weighting (Chen et al., 2025) combines momentum-based class accuracy with instance loss, AoL tracks consecutive-error duration through accumulation and reset, with EMA retaining information from earlier episodes. Thus, two prediction histories can have identical error rates and current correctness yet different AoL states.

Manuscript revision. We have added the following comparison to Section 4.4 (page 9), with the complete results reported in Table 3 (page 10):

To assess whether AoL complements existing learning signals more comprehensively, we further implement and compare seven online-adapted baselines based on CB (Cui et al., 2019), Focal (Lin et al., 2017), LDAM (Cao et al., 2019), ICDW (Chen et al., 2025), SBS (Hacohen & Tuytelaars, 2025), EMA Error Weighting, and EMA Loss Weighting.

Each method is evaluated with and without AoL under identical conditions on CIFAR-10 and CIFAR-100, using IF-200, Period-10, ResNet-32, SGD, a buffer of K = 50,000, and three random seeds.

We have also detailed the Online Focal and Online ICDW adaptations in Appendix B.2 (pages 17–19).


AuthorsOpenReview ↗

Response to Questions 1-3

Question 1:

Is the evolution of prediction correctness during training approximately Markovian? Or is anything in the learning process Markovian? I think it’s a fundamental question for AoL.

Reply:

We thank the reviewer for raising this fundamental question.

The answer depends on the state considered. Under suitable assumptions about data generation and sampling, training can admit a Markov description when the state includes all variables needed for subsequent updates, such as model parameters, optimizer state, and relevant sampling and adaptive-weighting states (Dieuleveut et al., 2020). For AoL-guided training, these include the AoL and EMA states.

Binary prediction correctness is only a coarse observation of this training state and need not inherit its Markov property. The same correctness value can correspond to different model parameters, margins, and optimizer states, with different subsequent behavior. Thus, Markovianity of the full training state does not establish Markovianity of the correctness sequence. Our formulation assumes neither Markovianity nor non-Markovianity of that sequence.


Question 2:

If you think the prediction correctness during training is approximately Markovian, can you further explain how AoL provides a useful learning signal?

Reply:

We thank the reviewer for this important follow-up question.

Under the exact first-order Markov assumption for a fixed sample and training policy, next-step correctness depends on current correctness rather than the preceding correctness history. Thus, at a given training step, consecutive-error duration would then provide no additional information for predicting that sample’s next-step correctness under the first-order Markov assumption, once its current correctness is known.

AoL, however, is used to modify training priorities, rather than merely predict correctness under an unchanged policy. Markovianity of the observed correctness sequence under a fixed policy does not establish that current correctness alone is sufficient for choosing training weights. Nor does it imply that the same transition dynamics hold after those weights change.

AoL, therefore, does not require non-Markovianity as a justification. Its usefulness as a training signal is assessed through the resulting learning outcomes. Our expanded comparisons do demonstrate that duration-based emphasis improves training that is already guided by alternative difficulty and historical signals.


Question 3:

If you think the prediction correctness during training isn’t Markovian, can you further prove it?

Reply:

We thank the reviewer for the opportunity to clarify this point.

We do not claim that prediction correctness is necessarily non-Markovian, and a proof of non-Markovianity is not a prerequisite of the proposed method. AoL is well defined whether or not correctness follows a Markov process.

Our contribution is to represent consecutive unresolved-error duration explicitly and use it to guide optimization. The relevant empirical question is whether this duration-based state improves training, which we investigate through the expanded comparisons.

Manuscript revision. To address Questions 1–3, we have added the following clarification to Section 3.1 (page 4):

AoL does not require prediction correctness to be non-Markovian. A sufficiently specified training state can admit Markov dynamics (Dieuleveut et al., 2020), even when its binary correctness observations do not. If correctness is first-order Markov under a fixed training policy, duration adds no next-step predictive information beyond current correctness at a given training step. This does not determine the effect of changing training priorities using AoL; we evaluate its usefulness through training interventions.

Reference: Dieuleveut, A., Durmus, A., and Bach, F. (2020). Bridging the gap between constant step-size stochastic gradient descent and Markov chains. The Annals of Statistics, 48(3), 1348–1382.


Thank you again for your time in reviewing our manuscript and for recognizing our work. If you find our clarifications and improvements satisfactory, we would sincerely appreciate your consideration of a higher rating.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Weakness 1-4

Weakness 1-4:

In Table 2, the comparison with class balanced softmax is missing.

Reply:

We thank the reviewer for this suggestion. Conventional CB relies on global class-frequency statistics, which are not available in advance under our streaming protocol. To provide a direct comparison, we have implemented Online CB (Cui et al., 2019), using current-epoch observed class counts to compute effective-number weights.

Under CIFAR-10/IF-200/Period-10, adding AoL improves Online CB mean accuracy from 76.90 ± 1.27% to 79.86 ± 1.09% (+2.96 points) across three seeds. Specifically, the improvements are 76.78% to 81.11%, 75.69% to 79.16%, and 78.22% to 79.30% for seeds 42, 1024, and 2026, respectively.

On CIFAR-100, mean accuracy changes from 45.96 ± 0.56% to 46.04 ± 0.48% (+0.08 points), with two seeds matching the baseline and one showing a small improvement. This indicates that AoL can substantially enhance Online CB on CIFAR-10 while preserving comparable performance on CIFAR-100.

Manuscript revision. We have added the Online CB and Online CB+AoL comparisons to Section 4.4 (pages 9–10, Table 3) and provided the following implementation details in Appendix B.2 (pages 17–19):

We adapt class-balanced weighting to unknown and evolving class distributions by replacing full-dataset class frequencies with the observed class counts of the current epoch. Effective-number weights are computed with β=0.9\beta=0.9 and normalized to unit mean. Thus, Online CB does not require advance knowledge of the global training distribution.


AuthorsOpenReview ↗

Response to Official Comment by Reviewer vLKX

Official Comment by Reviewer vLKX:

Do you have comparison results with conventional continual learning methods?

Reply:

Dear Reviewer vLKX,

Thank you very much for your timely and valuable follow-up. Yes, the expanded comparisons in Table 3 include an online implementation of SBS (Speed-Based Sampling), following Hacohen and Tuytelaars (Predicting the Susceptibility of Examples to Catastrophic Forgetting, ICML 2025). SBS was developed specifically to address catastrophic forgetting in continual learning.

Under our streaming protocol, we evaluate SBS with and without AoL using identical experimental conditions across three seeds. As shown in Table 3, adding AoL improves mean accuracy from 76.46 ± 0.39% to 79.00 ± 0.69% on CIFAR-10, with a mean paired gain of 2.54 percentage points, and from 43.46 ± 1.10% to 45.32 ± 0.56% on CIFAR-100, with a mean paired gain of 1.87 points. These results demonstrate that AoL can further improve an established conventional continual-learning method.

Thank you again for your time in reviewing our response and the additional experiments. If you find our clarifications and improvements satisfactory, we would sincerely appreciate your consideration of a higher rating.

Sincerely,
The Authors of Submission 4190

AuthorsOpenReview ↗

Summary of Revisions and Additional Evidence

Dear Area Chairs,

Thank you for overseeing the review of our submission. We would like to briefly summarize the principal revisions and additional evidence addressing the reviewers’ concerns.

1. Broader comparisons with existing learning signals. We added three-seed comparisons with seven online-adapted methods, including Focal, ICDW, SBS, and EMA-error/loss weighting. AoL improves 36 of 42 paired comparisons and matches the remaining six (Section 4.4, Table 3). In particular, the mean gains over EMA-error and EMA-loss weighting are 1.60/1.15 and 2.42/1.20 percentage points on CIFAR-10/CIFAR-100. These experiments strengthen the evidence that duration-based emphasis remains useful alongside existing difficulty and historical signals.

2. Large-scale reproducibility and buffer-free operation. We expanded Places365 and ImageNet-1k to three seeds across three imbalance factors. At IF=200, mean gains are 2.74 and 1.46 points, respectively (Table 2(b)). We also added CIFAR-10 buffer-capacity comparisons: without a buffer, AoL improves accuracy by 3.61 and 5.60 points at IF=200 and IF=100 (Table 5(a)).

3. Clearer presentation and technical explanations. Table 1 now consistently reports class-level AoL, with both formulations provided separately in Table 4. We clarified the Markovianity questions, online threshold, evaluation protocol, and measured runtime.

The revised evidence supports our central contribution: consecutive error persistence provides a useful training signal beyond static frequency and instantaneous difficulty, with reproducible benefits under persistent streaming imbalance. We respectfully ask that these substantial additions be considered in the final assessment of the paper.

Thank you for your consideration.

Sincerely,
The Authors of Submission 4190

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.