acceptodds
Under review as a conference paper at ICLR 2027

On Prototype Composition for Few-Shot Class-Incremental Learning: Breaking Statistical Bias via Attentive and Expansive Calibration

Abstract

Few-shot class-incremental learning (FSCIL) faces an inherent statistical imbalance: base classes are learned from abundant data, whereas each incremental class has only a few support examples. This imbalance produces two different vulnerabilities in prototype composition: base prototypes are trained without future classes, while centroids estimated from a few incremental supports may miss valid query directions. Their joint use can distort base–novel decisions in either direction. We refer to this decision-level asymmetry as statistical bias in prototype composition. To mitigate it, we propose BASIL (Base-Prototype Alignment with Support-Informed Local Expansion), an attentive and expansive prototype-calibration framework. First, class-weighted alignment constrains the displacement of base prototypes from their empirical centroids without discarding base-class discrimination. Second, real supports and surrogate tangent references expand the incremental-class reference set while the mean-prototype score is retained. BASIL requires no training during incremental learning and achieves strong results on widely used FSCIL benchmarks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Domain Adaptation, Continual Learning, Transfer Laerning, FSCIL

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_GHq2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies few-shot class-incremental learning (FSCIL), where base classes are learned from abundant data and novel classes are subsequently introduced with only a few labeled examples, with evaluation over all seen classes. The authors propose BASIL, which addresses the joint use of learned base-class prototypes and support-estimated novel-class prototypes. Following base-session VPT adaptation, the method freezes the representation and applies concentration-weighted attentive calibration to base prototypes and expansive calibration to novel-class scores using real support features and tangent references. Experiments on CUB-200, CIFAR-100, and mini-ImageNet show competitive performance, with additional evaluations on FGVC-Aircraft and iNF200.

Critical strengths
  1. Reasonable methodological motivation. The proposed method explicitly accounts for the different estimation conditions of base and novel prototypes. Applying centroid alignment to learned base prototypes and local-reference scoring to few-shot novel classes is a coherent design that reflects the statistical imbalance in FSCIL.
  2. Evaluation across multiple benchmarks. The method is evaluated on five FSCIL benchmarks, including CUB-200, CIFAR-100, mini-ImageNet, FGVC-Aircraft, and iNF200. The additional evaluations beyond the three main benchmarks broaden the empirical validation and help demonstrate the applicability of the approach across different datasets.
Critical weaknesses
  1. Limited methodological novelty. The proposed method combines concentration-weighted centroid alignment with interpolation between mean-prototype and local-reference scores. While the integration is reasonable and well motivated, the overall methodological novelty appears limited and mainly lies in combining these calibration mechanisms.
  2. Limited comparison with simpler calibration baselines. Although the paper includes component ablations, the additional benefit of the proposed calibration designs over simpler baselines remains somewhat unclear. Comparisons with simple weight–centroid interpolation and local scoring using only the mean prototype and real support features would strengthen the evaluation.
Questions for authors
  1. Base-prototype calibration. How does the proposed alignment compare with simpler alternatives, such as interpolation between learned base-class weights and empirical centroids?
  2. Contribution of tangent references. Could the authors compare against a local-reference set containing only the mean prototype and real support features, without tangent references?
  3. Concentration-dependent weighting. Does concentration-weighted alignment still outperform uniform alignment when the regularization strength is tuned separately for each variant?
Reviewer_twTZOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BASIL, a simple calibration method for few-shot class-incremental learning. It addresses two issues arising from the asymmetry between base and novel classes: learned base-class weights can deviate from their empirical centroids, while few-shot novel-class prototypes may poorly represent within-class variation. BASIL aligns base-class weights with their centroids and supplements novel-class scoring with support features and sampled tangent references, requiring no parameter updates during incremental sessions.

Critical strengths
  1. The motivation is clear and reasonable.
  2. The method is straightforward and requires no parameter updates or image replay during incremental sessions.
  3. Experiments include five datasets, with detailed ablations and analyses showing how each calibration component affects predictions.
Critical weaknesses
  1. The novelty is limited. The mismatch between base and new classes and the unreliability of few-shot prototypes are well-known issues. The proposed base-class solution is essentially a weighted centroid-alignment regularizer.

  2. New-class calibration provides modest additional gains across the three main datasets.

  3. Improvements over CASP are only 0.22% on CUB-200 and 0.37% on CIFAR-100.

Questions for authors
  1. Could the authors report base-class accuracy, novel-class accuracy, and their harmonic mean? These would make the effect on the base–novel balance clearer.

  2. CASP is the only 2026 baseline in the main results. Could the authors include more recent vision-only methods for fair comparisons?

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

The paper addresses statistical bias in prototype composition between base and incremental classes in Few-shot class-incremental learning (FSCIL) and proposes BASIL (Base-Prototype Alignment with Support-Informed Local Expansion), an attentive and expansive prototype-calibration framework. The framework consists of two core components. First, Attentive Calibration employs concentration-weighted alignment to constrain the displacement of base prototypes from their empirical centroids. Second, Expansive Calibration expands the incremental-class reference set using real support features and surrogate tangent references, while combining mean-prototype scores with local-reference scores to improve classification performance. The paper evaluates BASIL against existing methods on five FSCIL benchmarks, including CUB-200, CIFAR-100, mini-ImageNet, FGVC-Aircraft, and iNF200. It also conducts ablation studies to validate the effectiveness of the two calibration components, along with parameter sensitivity analyses of the number of tangent references, interpolation coefficient, and tangent perturbation magnitude. The experimental results demonstrate that BASIL achieves competitive classification performance across multiple FSCIL benchmarks, supporting the effectiveness of the proposed method.

Critical strengths
  1. The research problem is meaningful. The paper addresses distinct statistical biases affecting base and incremental classes in Few-shot class-incremental learning (FSCIL). The proposed method provides targeted solutions with clear research motivation.
  2. The method design is reasonable. Attentive Calibration balances empirical-centroid constraints with base-class discrimination through concentration-weighted alignment. Expansive Calibration enriches incremental-class classification references through tangent-reference expansion and interpolated similarity scoring while retaining the mean-prototype score.
  3. The experimental validation is comprehensive. The paper conducts performance comparisons and ablation studies on multiple public FSCIL benchmarks. The overall experimental results demonstrate competitive performance and provide empirical support for the effectiveness of BASIL.
Critical weaknesses
  1. The research motivation and method description remain insufficient. The Abstract and Introduction provide insufficient research background and justification for the method design. The distinctions between the proposed method and prior work remain unclear. In addition, Figure 2 does not fully illustrate the key computational processes or the interaction between Attentive Calibration and Expansive Calibration.
  2. Some ablation results are inconsistent, and the interaction between the two calibration components lacks sufficient explanation. Some ablation results show unexplained discrepancies across tables, weakening experimental rigor and reproducibility. In addition, Expansive Calibration alone does not demonstrate consistent performance gains. Its independent effectiveness and interaction with Attentive Calibration lack sufficient analysis.
  3. The evaluation of experimental stability and computational overhead remains insufficient. BASIL achieves relatively small performance gains on some datasets compared with existing methods. The paper lacks complete repeated experiments with multiple random seeds. In addition, Expansive Calibration increases feature storage requirements and similarity computation costs. The trade-off between performance gains and practical efficiency lacks sufficient analysis.
Questions for authors

no

Reviewer_3MQ2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies prototype composition in few-shot class-incremental learning using a pretrained ViT. To address the score competition between base classifiers and novel centroids, it introduces attentive base alignment weighted by feature concentration and expansive novel scoring using support vectors and tangent directions.

Critical strengths

Incremental sessions require no parameter updates or image replay.

Critical weaknesses
  1. The ablation study contains several internal numerical conflicts on CUB-200.

Table 5(b) reports 85.50% incremental average and 84.47% final accuracy for real supports with M = 0. In contrast, Table 6 reports 86.19% incremental average and 85.30% final accuracy for M = 0 under the same settings.

Table 4(b) lists 84.47% final accuracy for attentive only calibration, but Table 5(b) and Table 6 report 83.97% for the same mean prototype baseline.

The overall average of 85.33% in Table 5(b) row 1 is inconsistent with the reported base accuracy of 89.28% and incremental average of 85.43%, which imply an overall average of 85.78% across eleven sessions.

Because the claimed gains from individual components are modest, such as 0.85 percentage points from tangent references in Table 5(b), discrepancies of this magnitude make it difficult to verify the ablation findings.

  1. The empirical evaluation leaves important baseline combinations unexamined.

Table 4(a) shows that using empirical centroids on base classes increases incremental average accuracy from 79.02% to 84.82%, which accounts for most of the gain observed in attentive calibration (85.43%). While attentive calibration preserves base session accuracy at 89.28% and Table 12 demonstrates that concentration weighting outperforms uniform alignment, the paper never evaluates pairing empirical base centroids directly with expansive novel scoring. Without this baseline, it is difficult to determine how much additional benefit attentive base head fitting provides when expansive novel scoring is already applied.

The paper does not compare against simple score adjustments, such as a group-specific additive logit bias or separate temperature scaling for novel class scores.

ConCM is cited in related work as addressing prototype bias, but is omitted from benchmark tables. Given differences in backbones and replay assumptions, a discussion of ConCM's methodological and experimental distinctions would help position the contribution, while a controlled comparison under matched conditions would provide stronger evidence where feasible.

Questions for authors
  1. Can you explain the origin of the numerical discrepancies across Tables 4(b), 5(b), and 6, and provide the corrected numbers?
  2. What accuracy is achieved if empirical base centroids are paired directly with expansive novel scoring?
  3. How does BASIL compare against simple score adjustments, such as tuning an additive logit bias or temperature parameter specifically for novel classes using validation data permitted by the FSCIL protocol?
  4. Could you provide a discussion or matched comparison with ConCM, accounting for differences in architecture and replay protocols?
Reviewer_nVC7OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes BASIL for few-shot class-incremental learning. It calibrates base classifier rows through centroid alignment weighted by feature concentration, then scores incremental classes by interpolating their mean-prototype score with the maximum over real supports and random tangent references. The representation remains frozen after base training, and incremental sessions require no parameter optimization. Experiments on five benchmarks report competitive accuracy, supported by component ablations and geometric analyses.

Critical strengths

The incremental procedure is straightforward and avoids parameter updates or image replay. The manuscript explicitly distinguishes stored support features and synthetic references from additional independent observations, making the operational assumptions clear.

Critical weaknesses
  1. In my opinion, this is not a particularly polished piece of work; there are clear signs of heavy AI involvement—for instance, all the figures in the main text require revision. While the author addresses a very interesting question and has conducted an extensive set of experiments, the final writing appears quite rushed, and the AI-generated text is confusing in many places.

  2. There are numerous instances of data inconsistency. Please carefully correct them. These include, but are not limited to:

(1) The central ablations are internally inconsistent. Without tangent references, Table 5(b) reports 85.50% incremental average and 84.47% final accuracy, whereas Table 6 reports 86.19% and 85.30%. The attentive-only final accuracy also differs between Tables 4 and 5(b).

(2) The prediction-change counts do not reconcile with the accuracy tables. Adding expansive calibration reportedly corrects 54 novel predictions while harming 13 novel and three base predictions. This gives 38 additional correct predictions, or approximately 0.66 percentage points over 5,794 test images. Table 4 instead reports an increase from 84.47% to 85.66%, or 1.19 points.

  1. The tangent references may primarily prevent local scores from falling below the prototype score. They are isotropic perturbations around the mean, and Appendix C.3 reports that real-support scores exceed the mean for only 28.67% of queries.
Questions for authors

Please refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.