acceptodds
Under review as a conference paper at ICLR 2027

RIDER: SUPPORT-AWARE ANALYTICAL EVIDENCE CONSTRUCTION FOR REHEARSAL-FREE CLASS- INCREMENTAL LEARNING

Abstract

Recently, rehearsal-free class-incremental learning (CIL) based on pre-trained models has attracted increasing attention, and low-rank adaptation further enables parameter-efficient continual learning. However, when task identity is unavailable at test time and class scores from different task branches participate in a unified decision, the model must resolve both competition across tasks and ambiguity among classes within each task. To better understand this issue, we diagnose the final prediction errors by decomposing them into cross-task errors and within-task classification errors, and find that cross-task errors dominate the final errors in three of the four evaluated settings. Motivated by this finding, we propose RIDER with a Support-Aware Analytical Evidence Construction mechanism that addresses the problem from the perspective of evidence construction.Specifically, RIDER combines calibrated prototype evidence, an all-class analytical global reference, and support-valid branch-native evidence to construct global predictions, without completing unsupported branch–class statistics. In addition,we constrain shared-update drift through importance-weighted regularization, decouple task-specific directions from their magnitudes, and progressively reduce the ranks of newly introduced directions to limit incremental parameter growth. Extensive experiments across multiple CIL benchmarks demonstrate the effectiveness of RIDER. In particular, both final accuracy and average incremental accuracy improve in all 55 strictly paired runs.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Class-Incremental Learning, Rehearsal-Free Learning, Parameter-Efficient Adaptation, Low-Rank Adaptation, Task-Agnostic Inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RBRGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This work addresses rehearsal-free class-incremental learning by constructing and aggregating three sources of classification evidence to mitigate cross-task errors. The experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The proposed branch-native classifier aggregation is a promising approach to mitigating cross-task errors.
  2. The reported performance improvements are consistent across the evaluated settings and paired experimental runs.
Critical weaknesses
  1. The aggregation of branch-native classifiers may introduce unequal reliability across tasks. Classes introduced in earlier tasks are supported by fewer native classifiers, which may lead to less reliable score estimates. Although the method normalizes by the number of supporting classifiers, this does not necessarily address differences in estimation reliability.
  2. The methodology section contains excessive implementation details, making it difficult to identify and understand the key novelty of the proposed method. A clearer presentation that emphasizes the central contribution and explains the roles of the individual components would substantially improve the paper.
  3. Although mitigating cross-task errors is the central motivation, the decomposition of cross-task and within-task errors is presented only in the diagnostic analysis in Figure 1. The subsequent component ablations report aggregate accuracy without showing how calibration, global-reference evidence, and branch-native evidence individually affect cross-task and within-task errors. Reporting this decomposition across components and random seeds would provide more direct evidence for the claimed mechanism.
Questions for authors
  1. The random-feature ridge classifier in Eq. 3 follows the approach used in RanPAC. What is the specific motivation for introducing this classifier in RIDER, and what advantages does it provide over alternative classification methods? RanPAC should also be explicitly cited when introducing the random-feature classifier.
  2. DUKAE and CKAA are cited as closely related methods addressing decision-level accumulation, missing-distribution handling, cross-subspace alignment, and adapter aggregation. Why are these methods not included as experimental baselines? A comparison with these methods would help clarify the advantages of RIDER over closely related approaches.
7 comments

W1: Concern about unequal reliability across tasks.

We evaluate Full (the complete RIDER) and Full−G (RIDER removes only the global reference) after all 10 tasks, grouping test examples by the task that introduced their true class. The comparison uses identical learned branches, fitted classifiers, and normalization statistics, with only the global-reference term removed. We report three seeds on ImageNet-R and CIFAR-100 with seeds 1993, 1994, 1995.

The table reports paired accuracy gains and cross-task error reductions from adding the global reference, in percentage points. ImageNet-R and CIFAR-100 entries are mean ± sample standard deviation across seeds. Positive values indicate improvement.

DatasetIntroduction task(s)Native heads per classAccuracy gainCross-task-error reduction
ImageNet-R111.52 ± 0.241.52 ± 0.24
ImageNet-R220.30 ± 0.400.34 ± 0.36
ImageNet-R1–2 combined1–20.92 ± 0.220.94 ± 0.26
ImageNet-R9–109–10−0.43 ± 0.61−0.32 ± 0.63
CIFAR-100111.301.40
CIFAR-100221.001.00
CIFAR-1001–2 combined1–21.151.20
CIFAR-1009–109–10−0.20−0.20

The earliest, single-head classes benefit in all three ImageNet-R seeds. Their accuracy gains are +1.31/+1.47/+1.78 points, with cross-task-error reductions of 1.45/1.32/1.78 points. The combined task-1–2 group also improves in every seed: its mean accuracy rises from 83.46% to 84.37%, while cross-task error falls from 15.68% to 14.74%. CIFAR-100 provides supporting single-run evidence, with early-group accuracy increasing from 90.95% to 92.10% and cross-task error decreasing from 8.60% to 7.40%. Thus, this evidence establishes a repeatable benefit for the earliest low-support classes in, rather than a benefit confined to later classes. .

W2: Cross-task and within-task error decomposition in component ablations.

We have extended the component ablation to report cross-task and within-task errors e<sub>cross</sub> and e<sub>within</sub>, both in % alongside accuracy. Using identical learned branches, this decomposition reveals how the prediction components affect the two error types. As shown in Table 1 below (Table 4 in the revised manuscript), on CIFAR-100/ImageNet-R, calibration alone improves final accuracy over the prototype baseline by 0.56/0.45 percentage points, compared with 6.14/5.63 points for full fusion. These gains primarily arise from reduced cross-task confusion: full fusion decreases e<sub>cross</sub> from 14.10%/25.31% to 7.82%/19.69%, corresponding to reductions of 6.28/5.62 points. Meanwhile, e<sub>within</sub> changes from 0.12%/0.22% to 0.27%/0.21% and remains small across all variants. Full fusion also exceeds the best variant in both final accuracy and average accuracy. Removing calibration, native evidence, or the global reference consistently increases cross-task error and decreases both accuracy metrics. Together, these results support the complementary contributions of calibrated prototypes, branch-native evidence, and the global reference to resolving cross-task confusion.

Table 1: Evidence ablation on fixed learned RIDER branches (three seeds).* All values are reported as mean ± std. B denotes the prototype baseline; C denotes prototype calibration; N denotes branch-native ridge evidence; G denotes the global reference. B+X indicates adding only component X to the prototype baseline. Full denotes the complete RIDER framework, while Full-X denotes removing component X from RIDER.*

VariantCIFAR-100 A_20CIFAR-100 Avg.CIFAR-100 e_crossCIFAR-100 e_withinImageNet-R A_40ImageNet-R Avg.ImageNet-R e_crossImageNet-R e_within
B85.77 ± 0.9991.45 ± 0.7514.10 ± 0.720.12 ± 0.1574.47 ± 0.6981.74 ± 0.4825.31 ± 0.570.22 ± 0.26
B+C86.33 ± 0.8392.03 ± 1.0913.40 ± 0.680.27 ± 0.0874.92 ± 0.6782.05 ± 0.5024.80 ± 0.530.28 ± 0.19
B+N89.97 ± 0.7492.95 ± 0.839.83 ± 0.620.20 ± 0.0778.03 ± 0.7583.13 ± 0.8221.74 ± 0.510.23 ± 0.09
B+G89.89 ± 0.6893.11 ± 0.779.90 ± 0.490.21 ± 0.1075.33 ± 0.4582.57 ± 0.6224.43 ± 0.450.24 ± 0.11
Full91.91 ± 0.1995.02 ± 0.287.82 ± 0.210.27 ± 0.1380.10 ± 0.3485.51 ± 0.5219.69 ± 0.380.21 ± 0.21
Full-C90.51 ± 0.2993.65 ± 0.379.37 ± 0.260.12 ± 0.0578.70 ± 0.5584.18 ± 0.7021.11 ± 0.460.19 ± 0.07
Full-N90.39 ± 0.4593.63 ± 0.589.50 ± 0.480.11 ± 0.1777.97 ± 0.3683.65 ± 0.5321.77 ± 0.350.26 ± 0.16
Full-G90.07 ± 0.2393.52 ± 0.309.76 ± 0.230.17 ± 0.0878.62 ± 0.3984.16 ± 0.5821.15 ± 0.360.23 ± 0.14

Q1: Motivation and advantages of the random-feature ridge classifier.

We adopt this classifier from RanPAC (McDonnell et al., 2023) and will explicitly cite it when introducing Eqs. (2)–(3).

Specific motivation in RIDER. The diagonal prototype rule scores each class only in its owner branch, leaving evidence from other frozen branches unused. RIDER processes each arriving task’s training data through all available frozen branches, so branch bb has observed Cb:t\mathcal C_{b:t} at stage tt. Each branch-native random-feature ridge head uses sufficient statistics accumulated from these observations and contributes scores only for supported classes, without reconstructing unobserved earlier-class statistics. To complement native heads with different class coverage, RIDER fits a separate ridge reference head in the first branch’s fixed feature space using an independent fixed random projection. Because this branch observes every task, the reference provides all-class scores in a common representation. Together, these heads supplement owner-prototype evidence using directly observed data.

Advantages over alternative classifiers. Compared with cosine prototypes, random-feature ridge uses both class-wise feature sums and the Gram matrix, incorporating second-order feature structure beyond prototypes alone. Compared with linear ridge on original features, a fixed random projection followed by ReLU provides nonlinear expressivity without learning projection parameters. Compared with iterative fitting, additive updates to the Gram matrix and feature–class sums recover the corresponding batch ridge solution over observed data for the same fixed feature map and regularization. RIDER maintains this condition by freezing each completed branch’s full shared and task-specific state while keeping its random projection and feature-extraction preprocessing fixed (Appendix C.2). Thus, statistics remain compatible across tasks, enabling classifier updates without retaining historical images or per-example features.

Our evidence ablations use identical learned branches. Removing native evidence reduces mean final accuracy by 1.52/2.13 percentage points on CIFAR-100(T=20)/ImageNet-R(T=40), while removing the global reference reduces it by 1.84/1.48 points, respectively.

Q2: Lack of experimental comparisons with closely related baselines DUKAE and CKAA.

We agree that DUKAE and CKAA are directly relevant baselines. Our original experiments cover two published task-split protocols and a range of LoRA-based and other PECL methods, but do not include direct evaluations of these two closely related approaches. DUKAE trains its all-class heads using approximated distributions for missing old-class in later subspaces, though their outputs are excluded from voting. CKAA explicitly learns to classify features projected into incorrect subspaces and uses task-confidence estimates to aggregate adapters. RIDER instead accumulates sufficient statistics of actual observations in each fixed branch and solves the corresponding ridge objective. With fixed feature maps and regularization, its accumulated-statistic solution equals batch ridge fitting on those observed features (Appendix C.2), without pseudo-feature statistics or explicit cross-subspace feature alignment. Its separately fitted global reference and calibrated owner prototypes provide complementary evidence beyond native-head voting.

We address this concern with comparisons: complete methods under matched experimental configurations in the below Table 2 (Table 2 in the revised manuscript). RIDER follows the respective DUKAE and CKAA experimental configurations and matches the comparator’s seeds 1993, 1994, and 1995. AS shown in Table 2, RIDER achieves the highest mean Final and Avg.\ accuracy. These results demonstrate that RIDER maintains strong performance across different recognition benchmarks.

Table 2: Comparison with DUKAE and CKAA on five benchmarks. Results are reported as mean ± std. over three seeds. Bold denotes the best results.

MethodCIFAR-100 FinalCIFAR-100 Avg.ImageNet-R FinalImageNet-R Avg.ImageNet-A FinalImageNet-A Avg.CUB-200 FinalCUB-200 Avg.Cars-196 FinalCars-196 Avg.
DUKAE92.39 ± 0.1395.21 ± 0.0781.42 ± 0.3086.03 ± 0.38----89.16 ± 0.2092.15 ± 0.7584.59 ± 0.3388.34 ± 0.54
CKAA92.26 ± 0.3995.18 ± 0.8682.41 ± 0.1387.20 ± 0.3563.53 ± 0.6671.11 ± 1.03--------
RIDER92.45 ± 0.1795.35 ± 0.1782.50 ± 0.1887.30 ± 0.0564.98 ± 1.0373.79 ± 1.1389.89 ± 0.1793.08 ± 0.5884.63 ± 0.4589.36 ± 0.27
AuthorsOpenReview ↗

The authors look forward to your further review

Dear Reviewer,

Thank you very much for your thoughtful review and valuable feedback.

We have carefully addressed your concerns, provided detailed responses and additional experimental results, and revised both the main manuscript and appendix accordingly. We hope these updates have adequately addressed your questions and further clarified the contributions and applicability of our method.

As the rebuttal discussion deadline is fast approaching, we would greatly appreciate it if you could review our responses and share any further feedback at your earliest convenience.

If any concerns remain or further clarification is needed, we would be more than happy to address them promptly before the discussion period concludes.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_GA5cOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RIDER, a rehearsal-free class-incremental learning framework designed to improve cross-task prediction without requiring task identities at inference. Motivated by the observation that cross-task errors dominate in several evaluated settings, RIDER introduces a Support-Aware Analytical Evidence Construction mechanism that combines calibrated task-specific prototype scores, a global reference ridge classifier, and branch-native ridge classifiers restricted to directly observed classes. Additionally, the framework incorporates importance-constrained shared adaptation, direction–magnitude decoupling, and progressive rank reduction for parameter-efficient continual learning. Experiments across five benchmarks demonstrate consistent improvements in final and average incremental accuracy over existing methods, supported by component ablations and paired evaluations.

Critical strengths

Clear research motivation and well-targeted method design. The paper decomposes final classification errors into cross-task errors and within-task classification errors, revealing that cross-task errors dominate in several experimental settings. Based on this observation, the authors propose a prediction mechanism that integrates calibrated prototypes, an all-class reference classifier, and support-valid branch evidence. The overall research logic is clear, and the proposed method is well aligned with the identified challenges.

Strong and consistent experimental results. The paper achieves competitive performance across five class-incremental learning benchmarks under two task-split protocols. Furthermore, component ablation studies demonstrate the complementary contributions of different evidence sources.

Critical weaknesses

Limited methodological novelty and insufficient distinction from existing approaches. The cross-task prediction errors and asymmetric class support across branches addressed in this paper have been extensively studied in the class-incremental learning literature. In particular, the proposed Support-Aware Branch-Native Evidence is highly similar to the Adaptive Expertise Ensemble (AEE) in DUKAE in terms of valid branch selection and prediction aggregation. Furthermore, similar designs for ridge regression classifiers, importance-weighted regularization, and direction–magnitude decoupling can be found in existing works such as RanPAC, EWC-LoRA, and SD-LoRA. Therefore, the contribution of this paper appears to lie primarily in combining existing techniques, while the independent novelty of its core mechanism remains insufficiently demonstrated.

Insufficient comparisons with closely related methods and limited computational efficiency analysis. Although the paper provides relatively comprehensive component ablations, it lacks direct comparisons with related support-aware aggregation methods, such as DUKAE, under unified experimental settings. This makes it difficult to determine whether the performance improvements arise from the proposed evidence construction mechanism or simply from ensembling multiple analytical classifiers. Additionally, RIDER requires maintaining multiple task-specific branches and high-dimensional ridge regression statistics. However, the paper mainly reports the number of trainable parameters, without systematically evaluating total model storage, inference latency, or computational complexity, making it difficult to assess the practical scalability of the proposed method.

Questions for authors

Regarding the novelty of Support-Aware Evidence Construction: The proposed Support-Valid Branch-Native Evidence in RIDER appears highly similar to the Adaptive Expertise Ensemble (AEE) in DUKAE in terms of valid branch selection and prediction aggregation. What are the fundamental methodological differences between these two approaches? Could the authors provide a direct comparison with DUKAE-style aggregation using identical learned branches and classifier settings to demonstrate the independent contribution of the proposed mechanism?

Regarding computational efficiency and scalability: RIDER requires maintaining multiple frozen task-specific branches, along with high-dimensional random features and ridge regression statistics. As the number of tasks increases, these components may introduce substantial storage and inference overhead. Could the authors report the total model storage requirements, inference latency, and statistical update costs, and compare them with those of the main baselines?

5 comments

W1: The independent novelty of RIDRE's core mechanism.

RIDER’s contribution concerns evidence construction under asymmetric branch–class observation support. Unlike a single fixed-space analytical classifier, successive task-adapted branches observe different class subsets. Applying ridge regression to these branches alone leaves the treatment of unsupported class outputs and their combination with all-class evidence unspecified. RIDER standardizes native scores over each branch’s valid class support before averaging (Eq. 4), then combines them with calibrated owner-prototype scores and a separately fitted all-class reference (Eq. 5). This construction retains evidence from task-adapted spaces while providing a common reference covering every class, without completing unobserved branch–class statistics. The contribution lies in this specific construction and prediction rule. The low-rank learning components provide the representations on which this rule operates.

To assess the prediction-stage contribution independently of the adopted training components, we compare against RIDER+AEE while keeping the learned branches, fitted classifiers, and other inference components fixed, replacing only native aggregation with DUKAE-style AEE. Table 1 below reports higher mean final and average accuracy for RIDER in every setting. The ten-seed final-accuracy gains include 1.13 percentage points on CIFAR-100 with 20 tasks, 1.45 on ImageNet-R with 40 tasks, and 1.63 on ImageNet-A with 10 tasks. These differences arise without changing representation learning or classifier fitting.

Complementary ablations in Table 4 use identical learned branches: full evidence fusion improves final accuracy over prototype-only prediction by 6.14 and 5.63 percentage points on CIFAR-100 and ImageNet-R, respectively, and removing calibration, native evidence, or the global reference reduces both accuracy metrics. Together, the methodological distinctions and controlled comparisons support the contribution of our evidence construction beyond the shared eligibility rule and inherited training components.

W2, Q1 & Q2: Direct empirical comparison with DUKAE-style aggregation and computational/storage overhead.

Comparison with DUKAE-style aggregation. The core difference between RIDER’s native aggregation and AEE is that RIDER performs support-restricted normalization before branch averaging.

  1. Direct aggregation (AEE formulation) 1∣Bc∣∑b∈BcQx,cb\frac{1}{|\mathcal B_c|}\sum_{b\in\mathcal B_c}Q^b_{x,c} Averages raw branch scores for class cc directly across all supporting branches, with no per-branch normalization applied before aggregation.

  2. Normalization-then-aggregation (RIDER formulation) 1∣Bc∣∑b∈BcQx,cb−mbdb\frac{1}{|\mathcal B_c|}\sum_{b\in\mathcal B_c}\frac{Q^b_{x,c}-m_b}{d_b} For RIDER, Qb=SbQ^b=S^b corresponds to the ridge score matrix, and the normalized aggregate output corresponds to SbnS^{\mathrm{bn}}. Each branch’s scores are normalized before being averaged across branches.

  3. Branch-wise normalization statistics mb=mean⁡(Q:,Vbb),db=std⁡(Q:,Vbb)+εz,εz=10−8m_b=\operatorname{mean}(Q^b_{:,V_b}),\quad d_b=\operatorname{std}(Q^b_{:,V_b})+\varepsilon_z,\quad \varepsilon_z=10^{-8} The mean and sample standard deviation are scalar moments computed over all unlabeled evaluation inputs and supported class columns VbV_b. Unsupported class columns are excluded before estimating these moments.

  4. Support-dependent class offset term 1τ(c)∑b=1τ(c)mbdb⏟support-dependent offset\underbrace{\frac{1}{\tau(c)}\sum_{b=1}^{\tau(c)}\frac{m_b}{d_b}}_{\text{support-dependent offset}} The centering term of each individual branch is constant across its own supported classes, but the set of branches included in the average is determined by class cc. This offset term is therefore class-dependent and cannot be absorbed into one common global shift. Accordingly, RIDER’s native aggregation is invariant to such branch-wise constant offsets. Its key role is to perform centering and scaling within the effective support before branch offsets are transformed into class-dependent offsets via different sets of eligible branches. Thus, normalizing each supported branch before aggregation (RIDER) is intrinsically distinct from aggregating first and normalizing afterward (AEE), and can change cross-task decision outcomes.

Moreover, DUKAE fits subspace classifiers using Gaussian-sampled features and approximates missing old-class distributions in later subspaces. RIDER instead accumulates sufficient statistics of actual observations in each fixed branch and solves the corresponding ridge objective. With fixed feature maps and regularization, its accumulated-statistic solution equals batch ridge fitting on those observed features, without completing unsupported branch–class statistics. Its separately fitted global reference and calibrated owner prototypes provide complementary evidence beyond native-head voting.

RIDER+AEE. We replace only the branch-native evidence aggregation of RIDER with DUKAE-style AEE, while keeping the learned branches, classifiers, and other inference components fixed, and use matching seed sets. Results are in Table 1 below (Table 5 in the revised manuscript). RIDER achieves higher mean accuracy on both metrics in all dataset–task settings, with final-accuracy gains of 0.18–1.68 percentage points. Thus, when AEE is supplied with RIDER’s learned representations and fitted ridge classifiers, RIDER’s evidence aggregation achieves higher mean accuracy. This demonstrates that our evidence construction and fusion mechanism provides an independent decision-stage benefit under fixed representations and fitted classifiers.

Table 1: Comparison between RIDER and DUKAE-style AEE aggregation across benchmark--task settings. Results are reported as mean ±\pm std.

DatasetTSeedsRIDER FinalRIDER Avg.RIDER+AEE FinalRIDER+AEE Avg.
CIFAR-100201091.94 ± 0.2194.89 ± 0.3090.81 ± 0.4294.18 ± 0.51
ImageNet-R401080.14 ± 0.3285.30 ± 0.5578.69 ± 0.4384.40 ± 0.54
ImageNet-A101064.85 ± 0.7172.81 ± 1.5263.22 ± 0.8072.02 ± 1.51
VTAB51095.62 ± 0.2595.80 ± 0.2595.44 ± 0.1095.74 ± 0.11
CIFAR-10010592.42 ± 0.1595.31 ± 0.1491.47 ± 0.2494.75 ± 0.25
ImageNet-R10582.55 ± 0.1987.00 ± 0.3581.85 ± 0.3486.56 ± 0.48
ImageNet-A10564.81 ± 1.0072.92 ± 1.7063.13 ± 0.8071.80 ± 1.73
DomainNet5575.55 ± 0.0281.09 ± 0.0174.35 ± 0.0680.65 ± 0.04

The total model storage requirements. We compare RIDER with EWC-LoRA, CL-LoRA, and SD-LoRA across seven settings: CIFAR-100 with 5/10/20 tasks and ImageNet-R with 5/10/20/40 tasks. For RIDER, total storage includes the pretrained backbone, adapters and historical snapshots, classifiers, random projections, Gram and cross-statistic matrices, direction caches, and EWC state. Profiling uses RTX 3090 forward computation and CPU-side evidence computation with 224×224 inputs. Batch sizes are 64 for CIFAR-100 and 32 for ImageNet-R; average per-image inference time (ms/image) covers both computation stages.

As shown in Table 2, on ImageNet-R with 40 tasks, RIDER requires 6243.53 MB versus CL-LoRA’s 5975.23 MB, and takes 117.23 versus 299.35 ms/image, a 60.84% reduction. EWC-LoRA offers faster inference and lower computation: 11.87 ms/image and 41.13 GFLOPs/image, compared with RIDER’s 117.23 ms/image and 1598.19 GFLOPs/image. Regarding scalability, increasing ImageNet-R from 10 to 40 tasks increases RIDER’s storage, average per-image inference time, and computation by factors of 2.49, 3.74, and 4.14, respectively. These measurements quantify the cost of retaining additional branches.

Table 2: Comparison of storage overhead, inference latency, and computational cost. SO denotes Storage Overhead, Apit denotes Average per-image Inference Time, and CC denotes Computational Cost.

MethodDatasetTasksSO (MB)Apit (ms/image)CC (GFLOPs/image)
RIDERcifar10051940.1417.07191.45
EWC-LoRAcifar10052015.3910.4541.13
CL-LoRAcifar10051890.6346.97176.36
SD-LoRAcifar10051978.326.8835.85
RIDERinr51934.0220.76194.36
EWC-LoRAinr52023.2214.141.13
CL-LoRAinr51898.9544.52176.36
SD-LoRAinr51986.1426.3335.85
RIDERcifar100102494.0829.74382.65
EWC-LoRAcifar100102444.4710.9941.13
CL-LoRAcifar100102314.5783.94352.71
SD-LoRAcifar100102452.1246.3436.58
RIDERinr102504.3531.31385.68
EWC-LoRAinr102468.0311.9641.13
CL-LoRAinr102340.92101.9352.71
SD-LoRAinr102475.6849.6736.58
RIDERcifar100203639.1959.41770.73
EWC-LoRAcifar100203542.177.8541.13
CL-LoRAcifar100203513.52156.18705.43
SD-LoRAcifar100203558.4762.1338.03
RIDERinr203680.0459.81778.86
EWC-LoRAinr203492.168.8841.13
CL-LoRAinr203368.74163.75705.43
SD-LoRAinr203503.678938.03
RIDERinr406243.53117.231598.19
EWC-LoRAinr406057.3711.8741.13
CL-LoRAinr405975.23299.351410.86
SD-LoRAinr406179.18173.2640.94
AuthorsOpenReview ↗

The authors look forward to your further review

Dear Reviewer,

Thank you very much for your thoughtful review and valuable feedback.

We have carefully addressed your concerns, provided detailed responses and additional experimental results, and revised both the main manuscript and appendix accordingly. We hope these updates have adequately addressed your questions and further clarified the contributions and applicability of our method.

As the rebuttal discussion deadline is fast approaching, we would greatly appreciate it if you could review our responses and share any further feedback at your earliest convenience.

If any concerns remain or further clarification is needed, we would be more than happy to address them promptly before the discussion period concludes.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_GJ15OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies parameter-efficient continual learning methods for rehearsal-free class-incremental learning (CIL). The authors point out that fixed task-specific branches cannot accumulate statistics for earlier classes, resulting in asymmetric support across branches. To address this, they propose RIDER, whose core is support-aware analytical evidence construction. It fuses calibrated prototype scores from the owner branch, a global ridge classifier reference in the fixed first-task branch that covers all seen classes, and branch-native ridge classifier evidence that is valid only within the support each branch has directly observed. In addition, the training incorporates importance-constrained shared adaptation and direction-magnitude task-specific adaptation, together with a progressive rank reduction schedule to limit parameter growth. The authors report evaluations on five benchmarks under two protocols, showing consistent gains across all 55 paired runs, and provide multiple ablations and theoretical properties.

Critical strengths

1.The innovations of paper are of substantive value. The authors present a qualitative observation that the model exhibits many cross-task errors, provide a precise error decomposition, and further experimentally verify the improvement brought by calibrating task identity. In addition, extensive experiments demonstrate that the proposed evidence construction yields consistent gains across all datasets. This enhances the credibility and validity of contributions. 2.The proposed RIDER explicitly addresses the asymmetric support problem without forcing cross-branch feature alignment or hallucinating missing class prototypes. Using random-feature ridge classifiers updated via accumulated sufficient statistics to form support-valid native evidence and a global reference, this design is both mathematically sound and practically effective.

Critical weaknesses

1.The gaps compared with the strongest baselines such as MiN and CaRE are small in some settings. For example, the gain on CIFAR-100 under Protocol A is only 0.21, and more notably, compared with the substantial improvements on ImageNet-R and ImageNet-A, the model's gain on CIFAR-100 under Protocol B is quite small. 2.The main experiments rely on transductive normalization. The prediction rule depends on the moments of the entire unlabeled evaluation set, so a single sample's prediction depends on other samples in the same batch. Samples typically arrive in a streaming manner in real-world CIL. Yet the authors report in Appendix B.6 per-sample and fixed training-reference normalization alternatives with statistically insignificant differences, which somewhat alleviates this concern. Write

Questions for authors

1.If all eight entries in Table 1 were re-reported under fully inductive normalization, whether RIDER still maintain its performance advantage over MiN or CaRE, particularly for the CIFAR-100 Protocol B setting where the gap is small? Or do you think support-aware evidence construction remains beneficial in an online CIL setting with batch size 1? 2.The paper use fixed weights (0.70, 0.15, 0.15) across all datasets. Have you considered making the fusion weights adaptive based on the estimated reliability of each branch?

4 comments

W1, W2, Q1: Performance under fully inductive normalization and batch size 1.

For clarity, in the original manuscript’s Protocol A table, CIFAR-100 Final accuracy and Avg accuracy are 91.94 ± 0.21/ 94.89 ± 0.30 for RIDER and 90.93 ± 0.36 / 94.33 ± 0.32 for CaRE: the Final gain is 1.01%, the Avg. gain is 0.56%.

As shown in Table 1, the fully inductive normalization results below are mean ± standard deviation over ten seeds for Protocol A and five for Protocol B. Final / Avg. are accuracies (%); gains are percentage points over the stronger baseline for each metric. Under fully inductive normalization, RIDER retains higher mean final and average incremental accuracy than both MiN and CaRE in all eight entries. Specifically, on CIFAR-100 Protocol B, RIDER achieves 92.45 ± 0.16 / 95.30 ± 0.15 (Final/Avg.), compared with 92.31 ± 0.59 / 94.95 ± 0.31 for CaRE and 91.92 ± 0.19 / 95.00 ± 0.16 for MiN. Thus, the gains are +0.14/+0.35 over CaRE and +0.53/+0.3 over MiN. Therefore, RIDER maintains its performance advantage over MiN and CaRE under fully inductive normalization.

Table1 (Table 15 in the revised manuscript): Comparison of default inference and fully inductive inference across MiN, CaRE, and RIDER. Results are reported as mean ± std.

Protocol A

MethodCIFAR-100 FinalCIFAR-100 Avg.ImageNet-R FinalImageNet-R Avg.ImageNet-A FinalImageNet-A Avg.VTAB FinalVTAB Avg.
MiN90.88 ± 0.6393.85 ± 0.4677.51 ± 1.1883.61 ± 1.0860.21 ± 1.1969.84 ± 1.5693.16 ± 0.3994.16 ± 0.34
CaRE90.93 ± 0.3694.33 ± 0.3276.32 ± 0.8482.20 ± 0.8660.74 ± 1.2769.38 ± 1.9294.24 ± 0.1194.34 ± 0.17
RIDER (Default)91.94 ± 0.2194.89 ± 0.3080.14 ± 0.3285.30 ± 0.5564.85 ± 0.7172.81 ± 1.5295.62 ± 0.2595.80 ± 0.25
RIDER (Fully Inductive)91.98 ± 0.2094.91 ± 0.3180.15 ± 0.3585.29 ± 0.5364.78 ± 0.7972.83 ± 1.5395.49 ± 0.3195.78 ± 0.25

Protocol B

MethodCIFAR-100 FinalCIFAR-100 Avg.ImageNet-R FinalImageNet-R Avg.ImageNet-A FinalImageNet-A Avg.DomainNet FinalDomainNet Avg.
MiN91.92 ± 0.1995.00 ± 0.1680.13 ± 0.1085.32 ± 0.6360.65 ± 0.7269.90 ± 1.3573.40 ± 0.0679.88 ± 0.05
CaRE92.31 ± 0.5994.95 ± 0.3180.40 ± 0.1785.56 ± 0.4960.47 ± 0.7369.15 ± 0.9973.32 ± 0.1279.75 ± 0.08
RIDER (Default)92.42 ± 0.1595.31 ± 0.1482.55 ± 0.1987.00 ± 0.3564.81 ± 1.0072.92 ± 1.7075.55 ± 0.0281.09 ± 0.01
RIDER (Fully Inductive)92.45 ± 0.1695.30 ± 0.1582.49 ± 0.1486.95 ± 0.3764.79 ± 1.1272.95 ± 1.6875.49 ± 0.0281.05 ± 0.02

For the online CIL setting with batch size 1, these predictions use fixed training-reference statistics at both normalization levels without pooling other test samples. The completed CIFAR-100 experiments achieve 92.08/94.95 at Protocol A and 92.47/95.28 at Protocol B, exceeding both baselines on both metrics, providing direct evidence that RIDER remains effective for single-sample inference in both CIFAR-100 settings.So, support-aware evidence construction remains beneficial in the online CIL setting with a batch size 1.

Table2 (Table 13 in the revised manuscript): Comparison of default inference and batch-size-one inference across MiN, CaRE, and RIDER. Results are reported as mean ± std. Protocol A uses ten seeds, and Protocol B uses five seeds. BS=1 denotes batch-size-one inference.

Protocol A

MethodCIFAR-100 FinalCIFAR-100 Avg.ImageNet-R FinalImageNet-R Avg.ImageNet-A FinalImageNet-A Avg.VTAB FinalVTAB Avg.
MiN90.88 ± 0.6393.85 ± 0.4677.51 ± 1.1883.61 ± 1.0860.21 ± 1.1969.84 ± 1.5693.16 ± 0.3994.16 ± 0.34
MiN (BS=1)90.92 ± 0.5993.87 ± 0.4377.47 ± 1.1983.49 ± 1.2560.42 ± 1.1669.86 ± 1.5593.17 ± 0.3894.15 ± 0.34
CaRE90.93 ± 0.3694.33 ± 0.3276.32 ± 0.8482.20 ± 0.8660.74 ± 1.2769.38 ± 1.9294.24 ± 0.1194.34 ± 0.17
CaRE (BS=1)91.00 ± 0.3494.36 ± 0.3576.35 ± 0.8082.21 ± 0.8460.75 ± 1.2769.38 ± 1.9194.26 ± 0.1394.38 ± 0.16
RIDER91.94 ± 0.2194.89 ± 0.3080.14 ± 0.3285.30 ± 0.5564.85 ± 0.7172.81 ± 1.5295.62 ± 0.2595.80 ± 0.25
RIDER (BS=1)92.08 ± 0.2694.95 ± 0.3280.08 ± 0.2885.16 ± 0.5864.67 ± 0.8372.69 ± 1.5795.69 ± 0.1495.91 ± 0.16

Protocol B

MethodCIFAR-100 FinalCIFAR-100 Avg.ImageNet-R FinalImageNet-R Avg.ImageNet-A FinalImageNet-A Avg.DomainNet FinalDomainNet Avg.
MiN91.92 ± 0.1995.00 ± 0.1680.13 ± 0.1085.32 ± 0.6360.65 ± 0.7269.90 ± 1.3573.40 ± 0.0679.88 ± 0.05
MiN (BS=1)91.83 ± 0.2194.89 ± 0.2280.17 ± 0.1285.36 ± 0.6160.71 ± 0.7070.03 ± 1.3473.45 ± 0.0979.92 ± 0.06
CaRE92.31 ± 0.5994.95 ± 0.3180.40 ± 0.1785.56 ± 0.4960.47 ± 0.7369.15 ± 0.9973.32 ± 0.1279.75 ± 0.08
CaRE (BS=1)92.31 ± 0.5994.82 ± 0.3780.29 ± 0.2385.43 ± 0.4560.49 ± 0.7569.32 ± 0.9573.37 ± 0.1779.80 ± 0.09
RIDER92.42 ± 0.1595.31 ± 0.1482.55 ± 0.1987.00 ± 0.3564.81 ± 1.0072.92 ± 1.7075.55 ± 0.0281.09 ± 0.01
RIDER (BS=1)92.47 ± 0.1295.28 ± 0.1482.31 ± 0.2186.80 ± 0.3464.79 ± 1.0972.76 ± 1.7875.61 ± 0.0481.16 ± 0.03

Q2: Feasibility of reliability-adaptive fusion weights.

Thank you for the suggestion. Our fixed-weight sensitivity study (Appendix B.4) shows changes of at most 0.28 and 0.23 percentage points in mean final and average accuracy relative to the default on the two evaluated benchmarks, indicating limited sensitivity within the tested range. We leave reliability-adaptive fusion as a direction for future work.

I thank the authors for their response, which has addressed my concerns. My overall assessment remains unchanged, and I maintain my original rating.

We sincerely thank the reviewer for taking the time to carefully consider our rebuttal and for providing thoughtful and constructive feedback, which have helped us further clarify our contributions and improve the presentation of our work. We are grateful for the reviewer's time and effort throughout the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.