acceptodds
Under review as a conference paper at ICLR 2027

SLARC: Shrinkage-Controlled Low-Rank Alignment and Regularized Classification for Exemplar-Free Class-Incremental Learning

Abstract

Continued adaptation of pre-trained representations can make stored class statistics stale. In exemplar-free class-incremental learning, old images are unavailable for recomputing these statistics. While existing methods compensate for representation drift and exploit class covariances for classification, two questions remain challenging without old images: how much of an estimated correction should be applied to old-class means, and how much class-specific covariance structure should the classifier retain? Shrinkage-Controlled Low-Rank Alignment and Regularized Classification (SLARC) addresses these questions through constrained mean alignment and a regularized Gaussian head. The alignment fits a low-rank affine residual and reduces its strength when disagreement between image-split fits is large relative to the proposed movement at old-class means. The head uses maintained means and combines retained class covariances with pooled and spherical estimates. Scores recorded at class introduction select the regularization parameters, allowing earlier classes to inform the choice without retaining their images or individual features. Across four ten-task benchmarks, SLARC improves E-LoRA's final accuracy by – percentage points.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual learning, Class-incremental learning, Exemplar-free, Semantic shift estimation

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_4L61OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an exemplar-free class-incremental learning framework designed to maintain reliable class statistics as pretrained representations evolve during continual adaptation. The method combines shrinkage-controlled low-rank mean alignment, which estimates representation drift from paired before/after features and reduces unreliable corrections using split-fit disagreement, with a regularized Gaussian classifier that combines class-specific, pooled and spherical class covariance estimates. Experiments on four ten-task benchmarks show improvements when integrated with E2-LoRA, SLCA, and TUNA.

Critical strengths
  1. The method motivation is clear. The paper identifies a practical problem in exemplar-free class-incremental learning with evolving representations: stored old-class statistics become stale as the feature space changes, while old samples are unavailable for recomputation. The method design closely follows the two motivating questions: i) split-fit shrinkage determines how much an estimated mean correction should be trusted, and ii) regularized Gaussian classification determines how much historical class-specific covariance structure should be retained for prediction.
  2. The experimental evaluation is thorough. The paper reports results on four class-incremental benchmarks and demonstrates integration with multiple methods, including E2-LoRA, SLCA, and TUNA. The evaluation further includes matched host comparisons, component and mean-update ablations, analyses of the shrinkage and covariance designs, and experiments with longer task sequences.
Critical weaknesses
  1. The experimental role and applicability of SLARC are not sufficiently clear. SLARC is presented as a plug-and-play framework, but it actually operates on top of a host continual-learning method with evolving representations. Table 2 reports SLARC component variants under E2-LoRA, while Table 1 applies SLARC only to SLCA, TUNA, and E2-LoRA. The paper should clearly state what properties a host method must satisfy for SLARC to be applicable, and explain why it is not integrated with other strong baselines such as SSIAT or RanPAC. Clarifying the compatibility conditions would make the scope of the plug-and-play claim much clearer.
  2. The plug-and-play claim could be better supported. SLARC is evaluated with SLCA, TUNA, and E2-LoRA, but the closest drift-estimation method, SLDC, is not reproduced as a complete matched baseline. Although SLDC is discussed in the main text and its sample-count shrinkage rule is evaluated in Appendix S6.6, the appendix explicitly states that this does not reproduce the full SLDC pipeline. A matched comparison with SLDC under the same setting would better clarify the advantage of the proposed low-rank alignment and split-fit shrinkage.

Reference: [1] Xuan Rao, Simian Xu, Zheng Li, Bo Zhao, Derong Liu, Mingming Ha, and Cesare Alippi. Compensating distribution drifts in continual learning with pre-trained vision transformers. In AAAI, 2026.

Questions for authors

Questions asked in the weakness section.

1 comment

Dear Reviewer 4L61,

Thank you again for your encouraging assessment and constructive suggestions. We have expanded our experiments and clarified the applicability of SLARC to different continual-learning methods, along with its relationship to SLDC.

We hope these additions have addressed the points you raised. Whenever convenient, we would appreciate any further feedback you may have. If there is anything else we can clarify, we would be very happy to do so.

Thank you for your consideration.

The Authors

Reviewer_VTqROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SLARC to address outdated class statistics caused by representation drift in exemplar-free class-incremental learning. The method uses paired current-task features extracted before and after training to fit an affine map within a low-rank subspace that preserves the mean displacement. It adjusts the strength of old-class mean updates based on split-fit disagreement. For classification, it retains the original class covariances and constructs a regularized Gaussian classification head by mixing class-specific and pooled covariances and applying spherical shrinkage. The head’s hyperparameters are selected using scores archived when each class is first introduced. The authors evaluate SLARC on four main benchmarks, across multiple host methods and task sequences of varying lengths. The results show improvements in classification accuracy in most settings.

Critical strengths

(1) The motivation is clear, and the method is well tailored to the setting. Under the constraint that neither old-class images nor per-example features can be retained, the authors focus on maintaining old-class means and using stored statistics. The split-fit procedure assesses the map’s sensitivity to sample selection at old-class means, where the map is actually applied, making this a reasonable design choice.

(2) The experimental evaluation is broad and provides substantial empirical support for the method’s effectiveness. The authors evaluate SLARC on four main benchmarks, across multiple host methods and task sequences of varying lengths. Ablation studies examine the roles of mean maintenance, the low-rank subspace, the shrinkage rule, and hyperparameter selection using archived scores. The paper also reports performance degradation in some settings and cases where shrinkage does not reduce class-mean estimation error, helping characterize the method’s scope of applicability.

Critical weaknesses

(1) SLARC’s contributions appear primarily incremental relative to SLDC. Both methods compensate for statistical drift through feature mapping and shrinkage. A direct comparison using the same host method and classification interface would help isolate the contributions of the low-rank constraint and adaptive shrinkage and substantiate the method’s novelty. (2) The connection between split-fit stability and the accuracy of old-class mean corrections remains insufficiently established. The two fitted maps may share a transfer bias, so agreement between their outputs does not guarantee accurate corrections. (3) The applicability of hyperparameter selection using archived scores under continually evolving representations requires further evaluation. These scores originate from different representations and class sets at different stages, and the relevance of early scores to current hyperparameter selection may change as training progresses. Further experiments comparing alternatives such as recent-task windows or time-decay weighting would help assess the influence of older scores on hyperparameter selection. (4) Figure readability and formatting consistency need improvement. Figures 1 and 2 contain densely arranged symbols and labels, making them harder to read and interpret. Caption placement is inconsistent between Tables 2 and 3.

Questions for authors

(1) Could the authors provide a direct comparison with the full SLDC method using the same pretrained model, task splits, and training budget to clarify SLARC’s additional contributions over its closest prior work? (2) Could the authors analyze the correlation between split-fit disagreement and the actual error in old-class mean updates, and examine whether larger disagreement is associated with greater benefits from shrinkage, to validate the effectiveness of this indicator? (3) The paper fixes the rank at 64 and the ridge regularization coefficient at 0.01. Could the authors report sensitivity analyses for these two hyperparameters on representative datasets to assess whether performance depends on these specific choices?

Reviewer_PsKrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses representation drift in exemplar-free class-incremental learning (EFCIL) by proposing SLARC (Shrinkage-Controlled Low-Rank Alignment and Regularized Classification). The framework introduces an identity-centered low-rank affine mapping to align historical class means, with a disagreement-based shrinkage mechanism to adaptively control correction magnitude. It further employs a regularized Gaussian classifier that combines class-specific, pooled, and spherical covariance estimates, along with an archived birth-score strategy for hyperparameter selection without retaining historical data. Experiments on CIFAR-100, ImageNet-R, ImageNet-A, and Cars-196 demonstrate consistent improvements over E2-LoRA, with final accuracy gains ranging from 0.70 to 4.08 percentage points. Additional ablation studies and evaluations across different task sequences further examine the effectiveness of the proposed components.

Critical strengths

The paper addresses an important challenge in EFCIL, namely the mismatch between stored class statistics and evolving feature representations. The proposed framework provides a practical solution for maintaining historical class means without retaining old training data. In particular, the disagreement-based shrinkage mechanism evaluates the reliability of alignment mappings at stored old-class means, offering a reasonable way to control correction magnitude. The integration of mean alignment and covariance-regularized classification is technically coherent and well aligned with the identified problem.

The paper evaluates SLARC across four benchmarks and multiple incremental learning frameworks, demonstrating consistent improvements over E2-LoRA, particularly on Cars-196 with a 4.08 percentage-point gain in final accuracy. The experimental analysis includes component ablations, comparisons of alignment and shrinkage strategies, and evaluations of classifier hyperparameter selection. Notably, the fixed-trajectory comparisons help isolate the effects of different mean-update strategies, providing stronger empirical evidence for the effectiveness of the proposed components

Critical weaknesses

The methodological novelty of SLARC remains insufficiently distinguished from closely related approaches, particularly SLDC, which already employs ridge-based feature alignment and identity shrinkage. Although SLARC introduces displacement-aware low-rank alignment, split-fit shrinkage, and archived hyperparameter selection, their individual advantages over existing techniques are not adequately established. The comparison with SLDC's shrinkage strategy retains SLARC's alignment formulation rather than evaluating the complete SLDC pipeline under matched conditions. Furthermore, integrating SLARC with TUNA modifies its original inference mechanism, and the resulting performance is inferior to the native-inference-preserving transport variant. These limitations weaken the evidence supporting the distinct contributions of the proposed framework.

The robustness and independent contribution of the regularized Gaussian classifier are not sufficiently validated. The archived birth-score strategy selects hyperparameters using historical validation scores, but its stability under evolving representations and different validation partitions remains unclear. This concern is particularly relevant to ImageNet-A, where numerous singleton and empty fitting subsets require fallback estimates, yet their impact on hyperparameter selection is not adequately analyzed. Moreover, the benefits of count-dependent covariance regularization are not clearly isolated. On CIFAR-100, where all classes have equal sample counts, this mechanism effectively introduces additional flexibility in covariance mixing rather than demonstrating the advantages of class-dependent weighting. These issues leave the reliability and distinct effectiveness of the proposed classification component insufficiently supported.

Questions for authors

Table 3 reports improvements from introducing the count-dependent parameter (τ\tau). However, on CIFAR-100, all classes have equal sample counts, meaning that varying (τ\tau) effectively expands the available global covariance mixing weights rather than introducing genuine class-dependent regularization. Without comparisons against class-independent covariance mixing under matched effective hyperparameter search spaces, particularly on datasets with heterogeneous class sizes, it remains unclear whether the improvements arise from sample-count adaptation or increased hyperparameter flexibility.

The proposed strategy selects classifier hyperparameters using validation scores collected under historical feature representations and class sets, but its robustness to representation changes and different validation partitions remains insufficiently investigated. In particular, Appendix S4.1 reports 50 singleton and 12 empty fitting halves on ImageNet-A that require special fallback estimates. The potential effects of these cases on the ranking of the 66 candidate configurations, hyperparameter selection, and final classification performance are not adequately analyzed.

Although Appendix S5 provides theoretical complexity analysis, SLARC introduces additional computational and storage costs through full class covariance matrices, multiple alignment mappings at task boundaries, and the evaluation of 66 Gaussian-head configurations. The paper lacks comprehensive empirical comparisons with the original E2-LoRA framework in terms of training time, task-boundary processing time, peak GPU memory usage, storage requirements, and inference latency. Given the relatively modest accuracy improvements on three of the four main benchmarks, the practical accuracy–efficiency trade-off remains insufficiently established.

Appendix S6.6 evaluates SLDC's sample-count shrinkage rule within SLARC's alignment framework rather than comparing against the complete SLDC pipeline, including its mean and covariance transport mechanisms. This makes it difficult to establish the advantages of SLARC over closely related methods. Furthermore, the contributions of the proposed alignment strategy and the replacement of the classification head are not sufficiently disentangled, leaving their respective impacts on performance unclear.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.