acceptodds
Withdrawn from ICLR 2027

Exemplar-Free Few-Shot Class-Incremental Learning with Frozen Statistics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large-scale pretrained visual encoders have substantially raised the representation ceiling for few-shot class-incremental learning (FSCIL) on images and point clouds, allowing new classes to be learned from a few samples while old ones are retained. Yet many existing methods still replay raw exemplars, raising privacy concerns, or add prompts, auxiliary modules, or per-session gradient updates that increase complexity and overwrite previously learned knowledge. We show that exemplar-free FSCIL needs none of these: once the pretrained encoder is frozen, compact class statistics suffice. We propose FROzen STatistics (FROST), which replaces exemplar replay with two compact, complementary memories updated analytically in each session. A correlation memory accumulates sufficient statistics over features for a closed-form ridge solution, which is discriminative and needs no backpropagation but is biased toward data-rich base classes. A prototype memory keeps one normalised mean per class, so every class counts equally. Fusing the two combines discrimination with class balance and restores parity between base and novel classes. Across four point-cloud and three image benchmarks, FROST matches its exemplar-based variant and outperforms prior methods, most notably by +14.7 points on CO3D and +14.9 points on FGVC-Aircraft in terms of last-session accuracy. These results indicate that a dual-memory design over frozen features suffices for competitive exemplar-free FSCIL.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Class-Incremental Learning, Few-Shot Learning, Pre-trained Models, Point Clouds

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9f2nOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

FROST is an exemplar-free FSCIL method that freezes a base-adapted pretrained encoder and incrementally updates two memories: a closed-form ridge-based correlation memory and a class-prototype memory. The method is evaluated on multiple 2D and 3D benchmarks, where it reports strong improvements over prior FSCIL baselines.

Critical strengths
  1. Strong empirical performance across diverse 2D/3D FSCIL and cross-domain benchmarks.
  2. Simple exemplar-free design with no session-wise gradient updates, offering good efficiency and practicality.
Critical weaknesses
  1. CorrMem closely follows prior analytical ridge-based continual learning (e.g., RanPAC), while ProtoMem is a standard prototype-based classifier. The main contribution is largely their simple fusion.

  2. The paper lacks strong direct baselines such as RanPAC adapted to the same FSCIL protocol, with and without prototype fusion, making it difficult to isolate where the reported gains come from.

Questions for authors
  1. How does FROST differ technically from RanPAC/ACIL beyond adding ProtoMem and fixed logit fusion? Please clarify the methodological novelty.

  2. Could you report a direct baseline of RanPAC under the identical FSCIL protocol, as well as RanPAC + prototype fusion, to isolate the contribution of FROST?

  3. FROST substantially outperforms the joint-training Oracle in several settings. Could you clarify why this occurs and whether the Oracle uses exactly the same backbone, adaptation, and training protocol?

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

FROST is an exemplar-free few-shot class-incremental learning method for images and point clouds. It trains a lightweight adapter during the base session and then freezes the encoder. A correlation memory (CorrMem) accumulates second-order feature statistics GG and feature–label statistics QQ in a fixed random feature space, yielding a closed-form ridge classifier. A prototype memory (ProtoMem) stores a normalized mean for each class. The final classifier uses a fixed linear combination of ridge logits and prototype cosine scores. Novel sessions require neither old samples nor gradient updates.

Critical strengths
  1. The incremental procedure is simple and genuinely exemplar-free. Sections 3.3–3.6 train the adapter once, freeze the feature map, and subsequently update CorrMem using Gt=Gt−1+∑hh⊤G_t=G_{t-1}+\sum hh^\top and Qt=Qt−1+∑hey⊤Q_t=Q_{t-1}+\sum he_y^\top. ProtoMem appends one normalized mean for each novel class. With a stationary encoder, these statistics have a clear additive interpretation. The method avoids historical-sample re-encoding and per-session gradient optimization, which is useful when replay is unavailable.

  2. The evaluation covers several settings of practical interest. In Table 1, exemplar-free FROST reaches final accuracies of 91.3% on ModelNet, 89.2% on ShapeNet, and 72.0% on CO3D. Table 2 evaluates synthetic-to-real point-cloud transfer, and Table 3 includes CUB-200, iNF200, and FGVC-Aircraft. The paper also reports closely matched exemplar-based and exemplar-free variants and compares the two memories separately in Table 4. These results make a good case for treating frozen-feature analytical methods as serious FSCIL baselines.

Critical weaknesses
  1. The algorithmic novelty is limited, and the decisive controls for the dual-memory claim are missing. Section 2.3 discusses RanPAC, whose fixed random-feature ridge formulation closely matches CorrMem's accumulation of G,QG,Q and closed-form solution. ProtoMem is normalized class-mean cosine classification, and Eq. (10) is a fixed logit interpolation with α=0.5\alpha=0.5. The main claimed advance is their complementarity under base/novel sample imbalance. Yet Table 4 reports 90.36% final accuracy for CorrMem alone and 91.27% for the fusion on ModelNet, a gain of only 0.91 points. The fusion also reduces novel-class accuracy from 80.00% for ProtoMem alone to 76.43%. This demonstrates a particular overall-accuracy trade-off, not a clear restoration of base–novel parity or a nontrivial advantage over straightforward calibration. Comparisons with original RanPAC, normalized nearest-class-mean, class-balanced ridge, temperature-calibrated prototypes, and a simple ridge/NCM mixture should be performed with the same frozen encoder, base adapter, training data, and validation budget. Without such controls, it is difficult to identify a substantive methodological contribution beyond combining established classifiers.

  2. The reported margins are not disentangled from pretraining and base adaptation, and some evaluation terminology is ambiguous. Appendix Table 6 specifies Uni3D-Base for point clouds and ViT-B/16 pretrained on ImageNet-21K for images, with base-session adapter training; the image adapters are trained for 100 or 120 epochs. The main comparisons do not establish that each baseline uses the same checkpoint, adaptation schedule, augmentation, and selection budget. On FGVC-Aircraft, for example, FROST begins at 49.86% base-session accuracy compared with 27.67% for LGSP-P (Table 3), so the final 38.74% versus 23.87% margin cannot be attributed entirely to the dual-memory update. The “Joint Oracle” has 80.5% final accuracy on ModelNet, below FROST's 91.3% (Table 1). This is not necessarily an error, but the term “oracle” needs a precise specification of its backbone, optimization budget, and intended upper-bound interpretation. Section 4.1 defines AbA_b as accuracy after the base session while also describing the harmonic mean as one of final base and novel accuracies. The paper should distinguish A0A_0, ATbaseA_T^{base}, and ATnovelA_T^{novel} and report the exact quantity used in Figure 2. Finally, the V=2000V=2000 Gram matrix and repeated ridge solves have nontrivial memory and compute costs that should be compared with an actual exemplar buffer rather than described solely as compact statistics.

Questions for authors
  1. Under identical Uni3D/ViT checkpoints, base-session adapters, and extracted features, how do original RanPAC, plain NCM, class-balanced ridge, calibrated NCM, and simple score fusion compare with FROST? How much of the reported advantage survives this control?
  2. Given that fusion lowers novel accuracy relative to ProtoMem in Table 4, what final-session base/novel accuracy and confusion analysis supports the claim of restored parity? Please specify exactly when AbA_b is measured and how HH is computed.
  3. What are the encoder, pretrained weights, data access, and optimization budget of the “Joint Oracle” in Tables 1–2, and why does it perform substantially below FROST? What are the actual memory footprint and solve time of CorrMem at V=2000V=2000 relative to exemplar storage and replay?
Reviewer_YLT2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FROST for exemplar-free few-shot class-incremental learning on images and point clouds. After training a lightweight adapter on the base session, the method freezes the representation and maintains two classifiers: an analytical ridge-regression classifier constructed from sufficient statistics of random features, and a cosine classifier based on normalized class means. Their scores are combined with a fixed fusion weight. The paper reports strong final-session accuracy across in-domain, cross-domain, and fine-grained recognition protocols, together with similar performance between the exemplar-free and replay-augmented variants in the reported point-cloud experiments.

Critical strengths
  1. The method is simple and compatible with the stated exemplar-free sequential setting. Accumulating sufficient statistics avoids backpropagation during incremental sessions, while the frozen representation makes previously computed prototypes reusable. This is an attractive practical design.
  2. The evaluation covers both images and point clouds, including synthetic-to-real transfer. The substantial reported gains on CO3D and FGVC-Aircraft make the approach worth investigating and reproducing.
  3. The single-memory ablations exhibit different overall-versus-novel accuracy trade-offs, and their fusion yields a useful operating point. This provides preliminary evidence of complementarity beyond simply reporting the complete system.
Critical weaknesses
  1. CorrMem closely follows RanPAC, yet Table 4 evaluates fusion only on ModelNet, with a 0.91-point final-accuracy gain. The added contribution therefore remains unclear. Please compare CorrMem, ProtoMem, and FROST using matched features on CO3D and FGVC-Aircraft, and isolate the effects of fixing the ridge penalty and adding ProtoMem.
  2. On FGVC-Aircraft, FROST’s advantage over LGSP-P is already 22.19 points at the base session, versus 14.87 points at the final session (Table 3). This obscures gains from incremental learning itself. Please control checkpoints, training samples, and base-training budgets to isolate these effects, and specify the Oracle configuration.
  3. The bias-correction mechanism is insufficiently established: predictions depend on both Gt and Qt, prototype normalization does not remove few-shot uncertainty, and equal fusion weights do not ensure comparable score influence. Please test class-balanced ridge regression that reweights both statistics and score calibration using only base validation data.
  4. Section 4.1 uses Ab for both initial and final base accuracy, while Table 4 leaves the scope of An unclear. Figure 2 covers only two cross-domain protocols. Please distinguish evaluation stages and report final base/novel accuracies, their harmonic mean, and micro/macro averages for the component and image evaluations.
  5. Updating Gt can change old-class weights; equivalence to batch ridge regression on identical features, regularization, and samples does not guarantee fixed decision boundaries. Replaying already-accounted-for exemplars under deterministic feature extraction mainly reweights historical data. Please revise the boundary-preservation claim, clarify EB’s accumulation, and avoid generalizing EF–EB parity to adaptive replay systems.
Questions for authors
  1. Under identical adapted features, random projections, data indices, and base-only ridge selection, what are the final accuracies of CorrMem, ProtoMem, and FROST on CO3D and Aircraft? What are the paired per-seed improvements over CorrMem?
  2. What class set and evaluation time does A_n refer to? Is A_T micro- or macro-averaged? Please provide final base and all-novel accuracies, their test-set sizes, and the corresponding harmonic mean.
  3. Does FROST retain a consistent advantage over ridge regression that class-balances both G and Q? Does that advantage persist after score-scale calibration using base validation data only?
  4. With exactly the same frozen features and ridge penalty, what is the numerical discrepancy between incremental CorrMem and batch ridge regression? Does EB add exemplar contributions already present in the historical statistics, and is stochastic augmentation involved?
  5. Please specify the adapter loss and architecture, exact pretrained checkpoints, selected ridge penalties, hold-out handling, per-session update and solver latency, and persistent state size under the default configuration.

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

Related papers

Discussion (0)

Sign in to comment.