acceptodds
Under review as a conference paper at ICLR 2027

Informed Plastic Attractors for Adaptive Few-Shot Learning

Abstract

Hebbian plasticity enables gradient-free adaptation at deployment, making it attractive for few-shot and continual learning, but designing plasticity rules typically requires black-box optimization because their weight dynamics are not well understood. We introduce (IPA) that analytically define Hebbian plasticity matrices () in correspondence with the Information Bottleneck principle, resulting in unique local weight attractors. From labeled shots, these attractors adapt the head on a stream of single unlabeled data without gradients, replay or a query batch. We formalise a unique, locally contractive fixed point of the expected centroid mode and sample-wise dissipation of pre-activation energy. Under matched unlabeled-data access, IPA matches or improves nearest-class-mean classifiers, with gains where prototypes are weak, while batch-transductive methods win only with large, balanced query pools; the same plastic head meta-trains bodies that transfer across domains and heads; and in few-shot class-incremental learning with a frozen ResNet-12 it outperforms NCM-r, TENT, TIM, TEEN, ER, and the published NC-FSCIL result on CIFAR-100 without replay or gradient updates, is never below the prototype baseline, and with a frozen DINO backbone ends at final-session accuracy on MiniImageNet against for ER.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Hebbian Adaptation, Information Bottleneck, Few-shot learning, Test-time adaptivity

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

The paper introduces IPA, a hebbian adaptation mechanism. The motivation is clear: Hebbian plasticity allows NNs to adapt without backprop, but designing hebbian rules requires black-box optimizations of the parameters. Then, the question is if we can analytically construct such Hebbian rules from a small labeled support set and be competitive with few-shot and continual learning methods. The paper is difficult to follow as it condenses a lot of information, metrics and values into plain text. The benchmarks evaluated are broad but the results are relatively small compared to the baselines.

Critical strengths
  1. Motivation is clear and the method is interpretable. The central idea of analytically constructing Hebbian plasticity rules rather than relying on black-box optimization is interesting and has the potential to impact the literature. The method follows naturally from its motivation.

  2. Theoretical contribution is meaningful and limitation is acknowledged. For instance, Proposition 3.1 is useful and well explained.

  3. There is considerable experimental breadth. The paper evaluates the method across few-shot classification, meta-learning, and few-shot class-incremental learning, using different architectures, baselines, and ablations. While, I still have concerns about some of the comparisons and their interpretation, this breadth is a positive aspect, suggesting that the method is not exclusively desgined to just fit one experiment.

  4. The method has practical advantages at deployment. IPA operates on individual unlabeled samples without requiring gradient updates, replay buffers, or access to an entire query batch. The paper also provides measurements of computational and memory requirements.

Critical weaknesses

Major

  1. The experiments do not fully isolate the contribution of Hebbian plasticity. The method proposed, IPA-Approx (Sec. 4.3), combines the prototype-based initialization of the classifier, the hebbian updates and the confidence-weight of the class prototypes, while NCM-r uses hard nearest-centroid assignments but such comparison changes both the classifier adaptation and the prototype refinement strategy. Then, we see in Table 1 that NCM-r underperforms NCM on CIFAR-100. This seems like it inflates the benefit of IPA-Approx. I'd recommend introducing an ablation performing the same soft-prototype as IPA-Approx without the hebbian updates.

  2. The meta-learning comparison does not control for training conditions. Figure 3 results on cross-domain and cross-body are interesting. It shows an improvements over ANILr in 23/32 comparisons and on the transferability of IPA-trained backbones for 11/16 comparisons. However, these findings are not sufficient evidence to interpret it as if the Hebbian head encourages the transferable representations as stated in Lines 353-355. In particular, Section 4.2 mentions that Meta-IPA uses additional unlabeled episode data during meta-training, whereas MAML and ANILr adapt using only labeled support sets. If that is the case, could the authors clarify why the improvements are solely attributed to the hebbian adaptation mechanism? An ablation where Meta-IPA does not use additional unlabeled data would help establish a stronger claim about the representation-transfer advantage.

Minor

  1. The paper is hard to parse. There is a lot of information, results, math notation and citations embedded in the text. See 4.1 as an example. I understand the necessity of explaining and citing it all. However, lots of parts, as they stand right now, are extremely difficult to parse and understand the main claim.

  2. The experimental presentation is dense. Consider breaking down Fig. 3 into smaller figures or a different representation to improve readability.

  3. The practical impact and relevance of results is not sufficiently discussed.

Questions for authors
  1. What improvement shown in the results is uniquely attributed to the Hebbian update? I understand it is important to completely isolate the contribution in order to evaluate such claims.

  2. Could the meta-learning results be reproduced under comparable training budgets? (See Major Weakness 2) Also, how much of the cross-domain improvement of Fig. 3 is due to the learned representations?

  3. Meta-learning learning rate (Section 4.2, p. 6; Table 6, p. 22). Could the authors clarify which value was used for the results in Figure 3 and Table 7? Section 4.2 reports a meta-learning rate of (10^{-2}) for Meta-IPA, whereas Table 6 specifies (10^{-3}). This paragraph of Sec 4.2 is also a good example for the Minor Weakness 1.

  4. Omniglot results (Table 7, p. 24). I'm curious to understand what might have caused Meta-IPA to achieve 98.98% accuracy in the Omniglot 20-way 1-shot setting but just 95.71% with 5 shots. Could the authors explain why that could have happened?

3 comments
AuthorsOpenReview ↗

Review response LPjD

Thank you for the detailed review and the presentation suggestions. We will try to improve the text and considered your feedback helpful in improving the manuscript overall.

W1 / Q1. The soft-NCM-r ablation is presented in the common response. Here we show, that uniquely attributable to the Hebbian update are: head isolation +0.21 (CIFAR-100/ResNet-12), +2.37 (CIFAR-100/DINO), +0.91 (MiniImageNet/DINO) and +0.28 (MiniImageNet/ResNet-12); Robustness at test-time adaptation under skewed streams IPA-Approx (DINO) 100-fold class imbalance 67.45, against 64.37 for the control and 66.04 for NCM.

W2. Meta-IPA adapts on unlabeled episode images during meta-training; MAML and ANILr adapt on the labeled support only. We ran the ablation you suggest on MiniImageNet 5-way 5-shot. We meta-trained Meta-IPA twice with identical seeds (N=10 per condition), tasks and budget (1,000 meta-steps of 32 tasks), once with the inner loop on unlabeled episode images and once on the support embeddings only, and evaluated both on 600 held-out episodes per run (mean ± std over 5 seeds).

Inner loop during meta-trainingHeld-out accuracy (%)
Unlabeled episode images78.57 ± 1.25
Support embeddings only78.75 ± 1.08
Paired difference−0.18 ± 1.56 (p = 0.81)

We would like to point out that the lower number of meta-updates reduce the effect of possible data imbalance; and the gradient graph is not retained over time (i.e., unlabeled data never enter a loss, and no gradient flows through their features). Unlabeled data only drives head dynamics defined by the Hebbian rules conditioned on the support, so the two effects cannot be separated. Support-only meta-training is as accurate, so the extra unlabeled data is not what drives Meta-IPA's results. At evaluation Meta-IPA adapts on the support only, like MAML and ANILr.

Q2. Meta-IPA runs under a lower meta-budget (1000 meta-steps of 32 tasks) with 100 unlabeled Hebbian inner steps; MAML and ANILr run 10000 meta-steps with 5 gradient inner steps, their standard protocol. Meta-IPA thus receives a tenth of the body updates. (The "100 meta-steps" of the submitted §4.2 and Table 6 was an error for 1,000.) On how much of the cross-domain gain is due to the learned representation, the cross-body swap shows that the bodies carry part of it.

Q3. We follow the best performing sweep setup, our apologies for the confusion: Meta-IPA meta_lr 1e-2 in Figure 3 and Table 7. For the meta benchmarks we follow the original meta_lr 1e-3 for MAML and ANILr. The 1e-3 for Meta-IPA in Table 6 is a typo.

Q4. Omniglot is the only dataset with 20 images per character. At 1-shot, four unlabeled images per character remain after the support and the fifteen queries, and the inner loop adapts on those. At 5-shot none remain, so it adapts on the support embeddings without labels. The Hebbian loop then sees only the images that already define the prototypes. We hypothesise that this weakens the meta-learning signal for the body. The inversion does not occur on CUB200 or MiniImageNet, where a separate unlabeled set exists at both shot counts.

M1–M3. We will restructure §4.1 into a setup paragraph and a results paragraph with the full grid in the appendix, redesign Figure 3 into a cross-domain figure in the main text and a cross-body figure in the appendix, and add with a paragraph on practical relevance, adaptation one sample at a time in a single pass, 250 KiB of state, no replay buffer and no gradients.

Thank you to the authors for your rebuttal. My concerns have only been minimally addressed. I still have concerns about the viability of contribution to Hebbian plasticity as the isolation has not been clarified. I do not see the revision pdf and cannot make judgements on the improvement of readability and presentation either.

I will be keeping my Weak Rejection (2) rating. I hope the authors take the feedback time to improve the manuscript and to provide more concrete evidence.

Thank you

Thank you for the follow-up. We have uploaded a preliminary revision of the PDF. Changes are marked in red (we did not highlight every change). It restructures §4.1 into a setup and a results paragraph, splits Figure 3 into a cross-domain figure in the main text and a cross-body figure in the appendix, and states the practical relevance in the deployment-cost paragraph of §4.3 (one sample at a time, a single pass, 250 KiB of state, no replay buffer and no gradients).

On the effect size/outperformance issue. We are mostly presenting a novel method for test-time adaptation. The Hebbian head is what lets IPA adapt one sample at a time with no gradients, replay or query batch. It never falls below plain NCM, and it adds clearly where the refinement fails (self-supervised features, imbalanced or contaminated streams).

On the isolation of the Hebbian update. The control soft-NCM-r differs from IPA-Approx in one component only. It uses the same prototype construction, the same confidence-weighted refinement, temperature, running counts, sample order and budget, and removes the plastic head, so that the assignments come from the current prototypes instead of the Hebbian weights (Algorithm 2). The difference between the two is therefore due to the Hebbian update alone. It is positive on ResNet-12 (+0.21 on CIFAR-100, +0.28 on MiniImageNet, p < 1e-3) and in all nine combinations of three self-supervised backbones and three datasets (+0.53 to +6.04, p < 1e-4). Within the head, setting β = 0 removes the BC term and keeps the rest of the update, which removes most of the gain on DINO features (69.53 against 67.54 on CIFAR-100, last column of Table 2, and 74.18 against 69.81 on CUB200, Table 21). Table 2 in §4.3 now reports this decomposition step by step, from NCM over the prototype construction and the refinement to the head and the BC term, and Table 20 repeats it for the further backbones.

Where your comments are addressed in the preliminary revision (line numbers of the PDF).

  • W1 / Q1, soft-NCM-r as the matched control (lines 418–420, Algorithm 2 at line 1280), contribution of the head (lines 456–461, Table 2 at lines 432–441), the BC term and β (lines 463–473, Table 21 at line 1757), stream conditions and further backbones (Section C.19 from line 1693, Tables 18–20).
  • W2, support-only meta-training and the claim limited to what was measured (lines 356–360, Appendix C.5 at lines 1166–1170).
  • Q2, training budgets (lines 304–305 and 1162–1164).
  • Q3, lr_meta = 1e-2 for Meta-IPA (line 303, line 1161 and Table 7 at line 1244).
  • Q4, Omniglot 1- versus 5-shot (lines 1150–1153 and from line 1398).
  • M1, §4.1 split into a setup paragraph (line 257) and a results paragraph (line 268).
  • M2, cross-domain Figure 3 in the main text (lines 340–342) and cross-body Figure 12 in the appendix (line 1395).
  • M3, practical relevance (lines 475–480).
Reviewer_QqVHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IPA, a support-conditioned plasticity rule for few-shot adaptation. Class prototypes determine an analytical ABCD update, allowing a classifier head to adapt to unlabeled samples without deployment-time backpropagation or a replay buffer. The paper analyzes centroid-mode stability and evaluates the approach in few-shot classification, meta-learning, and few-shot class-incremental learning.

Critical strengths
  1. Simple and practical adaptation mechanism. The analytical rule supports sample-wise updates with modest state requirements and avoids deployment-time gradient computation.

  2. Useful analysis of the update dynamics. The centroid-mode stability result and first-order activity dissipation provide a concrete explanation of the feedback mechanism.

  3. Extensive empirical investigation. The paper includes multiple feature extractors, initialization and feedback ablations, robustness tests, and a single-pass experiment. The authors also acknowledge several limitations and settings with small or absent gains.

Critical weaknesses
  1. The contribution of the plastic head is not sufficiently isolated. Algorithm 2 uses soft confidence-weighted assignments for IPA-Approx but hard nearest-centroid assignments for NCM-r. Their comparison therefore changes both the head and the prototype-refinement mechanism. This matters because prototype initialization is already strong, and removing the BC term changes FSCIL AA by only 0.12 points in the reported ablation.
  2. The meta-learning evaluation is insufficiently specified. I could not locate explicit meta-train/validation/test class splits or an unambiguous definition of the 30 runs/episodes. These details are necessary to interpret the reported novel-class generalization and uncertainty. The text and Table 6 also report different meta learning rates. The disclosed differences in unlabeled-data access and training budgets further limit attribution of the gains to the plastic rule alone.
  3. Some theoretical statements need tighter conditions. In Appendix A.4, transverse symmetry does not by itself cancel the term proportional to 1E[h(az)]\mathbf{1}\mathbb{E}[h(az)], so the full-weight fixed-point extension appears to require an additional condition or projection. This does not invalidate the scalar centroid-mode result. Appendix B should also reconcile its heuristic interpretation with the stronger statements about minimizing conditional entropy and an MI upper bound.
Questions for authors
  1. Could you add a soft-NCM refinement baseline with matched preprocessing, initialization, temperature, running counts, sample order, and update budget? How much improvement remains attributable specifically to the Hebbian head?

  2. Please specify the meta-training, validation, and test class splits, independent training seeds, test episodes per seed, and statistical units used for uncertainty estimates. Which meta learning rate was used?

1 comment
AuthorsOpenReview ↗

Review response QqVH

Thank you for the precise reading of Algorithm 2 and Appendix A.4. The requested control is in our common response; the points specific to your review follow.

W1 / Q1. With preprocessing, initialization, temperature, running counts, sample order and budget matched, the gain attributable to the Hebbian head is +0.21 (CIFAR-100/ResNet-12), +2.37 (CIFAR-100/DINO), +0.91 (MiniImageNet/DINO) and +0.28 (MiniImageNet/ResNet-12). It is true that the effect of the compression term on the performance is not consistent BC ablation showed on ResNet-12 (0.11 points), while on DINO BC adds 2.0 (full table in the common response).

W2 / Q2. The protocol is as follows.

  • Runs and statistical unit. 30 independent meta-training runs per method (seeds 0–29, shared across methods), 10 evaluation tests per run, Figure 3 uses two-sample t-tests over runs.
  • Class splits. Each seed draws a random subset of classes for meta-training; evaluation classes are disjoint from it. MNIST, 5 digits / the other 5; Omniglot, 4,780 of the 4,800 training classes (1,200 characters × 4 rotations) / the 1,692 classes of the 423 test characters; CUB200, 180 / 20; MiniImageNet, 59 of the 64 training classes / the 16 validation classes.
  • The hyperparameters of the Figure 10 sweep were selected on the accuracy of one held-out episode from the evaluation classes over 3 repetitions per sweep setting (seeds 42-44). There is no early stopping, and every run is assessed on its final checkpoint.
  • Test episodes. Same-domain table, 10 episodes per run with 15 queries per class. Cross-domain (Figure 3), 10 episodes of the target dataset per body with 64 test samples per episode, mean over the 30 bodies. Cross-body (right panel of Figure 3), the same.
  • Learning rates. lr_meta = 1e-2 for Meta-IPA, 1e-3 for MAML and ANILr. The 1e-3 for Meta-IPA in Table 6 was a typo.
  • Budget. Meta-IPA runs 1,000 meta-steps of 32 tasks with 100 Hebbian inner steps per task; MAML and ANILr run 10,000 meta-steps with 5 gradient inner steps. The "100 meta-steps" of the submitted §4.2 and Table 6 was an error for 1,000. Meta-IPA thus receives a tenth of the meta-gradient updates to the body.
  • Omniglot inner loop. With 20 images per character, the 1-shot inner loop adapts on the four images left after the support and the fifteen queries. At 5-shot none remain, and it adapts on the support embeddings without labels.

W3. The expected BC update is −ηβ,a,E[h(az),(1+I)⊤]-\eta\beta\\, a\\, \mathbb{E}\big[h(az)\\,(\mathbf{1}+I)^\top\big], and its component along 1\mathbf{1}, −ηβ,a,E[h(az)],1⊤-\eta\beta\\, a\\, \mathbb{E}[h(az)]\\,\mathbf{1}^\top, is not removed by transverse symmetry. Under the proposition's assumption 1⊤I=0\mathbf{1}^\top I = 0 that component is orthogonal to μi\mu_i and inert on every pre-activation, so the fixed point a⋆a^\star holds for the expected dynamics modulo span⁡1\operatorname{span}\\{\mathbf{1}\\}. We state the full-weight extension with this projection and restrict the symmetry condition to the component orthogonal to μi\mu_i and 1\mathbf{1}. In Appendix B we replace "directly minimizing H(Y∣O)H(Y\mid O)" and "minimizes the upper bound on I(X;O)\mathcal{I}(X;O)".

Reviewer_ThchOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method to derive Hebbian plasticity rules in closed form instead of relying on evolutionary search. The authors use class centroids from a few labeled examples to define the plasticity rules. The resulting head adapts to unlabeled data one sample at a time, without gradients. The authors motivate the design using the Information Bottleneck principle and provide theoretical results on local stability. The method is mainly evaluated in few-shot learning and meta-learning settings, including few-shot class-incremental learning.

Critical strengths
  • Deriving plasticity rules in closed form instead of searching for them is an interesting direction for the plastic-network literature. The authors do a good job of highlighting this gap in the existing literature.
  • The gradient-free adaptation, without requiring a buffer or query batch, is attractive for deployment.
  • The empirical comparisons with other adaptation methods suggest that the proposed method is reasonably effective.
  • The paper is generally well written and covers an interesting topic with several potential applications.
Critical weaknesses
  • In my opinion, the main issue with the paper is its current framing. As I understand it, the method combines a contrastive prototype classifier with an Oja-style normalization term and, in IPA-Approx, there is additionally the soft prototype refinement. The framing around the Information Bottleneck suggests a stronger theoretical connection than what is established. The authors also acknowledge this at the end of section 3.1 and mention that IPA does not optimize an IB objective and that accuracy depends very little on beta.
  • I think presenting IPA as a closed-form Hebbian rule that extends prototype classifiers, with the IB perspective as a broader motivation, would make the contribution clearer. The analysis could then focus more on what the online plasticity dynamics add beyond a prototype classifier.
  • Empirically, the online adaptation adds only about 0.6 percentage points over the prototype initialization on CIFAR-100 with ResNet-12. This raises questions about how much of the performance comes from the proposed plasticity dynamics or the initial prototype construction.
  • I am not confident in some of the gradient-based baseline results in the FSCIL experiments. With a frozen DINO backbone on MiniImageNet, Finetune, EWC, and ER all reach 99.95% in the base session, while with ResNet-12 they start well below the prototype classifier on the same features.
Questions for authors
  • Where do the gains over NCM come from: per-class rescaling, dynamics away from the centroid direction, or prototype refinement in IPA-Approx?
  • Appendix C.16 compares against soft-k-means in the meta-learning setting. How would the results look in the FSCIL setting?
  • Could the authors clarify the gradient-based baseline results, mainly the near-perfect base-session accuracy with DINO?
3 comments
AuthorsOpenReview ↗

Review response Thch

Thank you for the careful reading. The control is in our common response; the decomposition you ask for follows under Q1.

W1, framing. We agree on the framing. An IB-based learning rule needs a surrogate, because the IB defines no loss on the weights and mutual information is intractable; the HSIC bottleneck (Ma et al., 2020) and its three-factor Hebbian form (Pogodin and Latham, 2020) take the same route. In those works the rule is the gradient of the surrogate. In IPA, AD is the exact gradient of a contrastive alignment surrogate for I(O;Y), and BC is a homeostatic surrogate for the compression term I(X;O), i.e., not the exact gradient of a scalar objective (as shown in Appendix B). We will position this more clearly in the revision. On β the picture is more nuanced. With self-supervised features the compression term carries most of the head's gain (67.54, 69.53 and 71.62 at β = 0, 1 and 8 on CIFAR-100/DINO; part 2 of the common response). As you suggest, the analysis now centres on what the dynamics add beyond a prototype classifier.

W2. Correct. On CIFAR-100/ResNet-12, everything after the prototype initialization adds 0.62 points (67.63 → 68.25), 0.41 from the refinement and 0.21 from the head. Most of the gain over NCM on these features comes from the construction. On self-supervised features the head adds more (+2.37 on CIFAR-100/DINO).

Q1. On CIFAR-100 (ResNet-12 / DINO), in the terms of your question.

  • Rescaling, +0.79 / +1.58.
  • Prototype refinement without the head, +0.41 / −0.47.
  • Head on top of the refinement, +0.21 / +2.37.
  • Dynamics away from the centroid direction come only from the BC term, since at β = 0 the head stays a scaled prototype classifier; they account for 0.11 / 2.0 of the head's gain.

Q2. soft-NCM-r is the FSCIL counterpart of soft-k-means refinement (soft assignments from centroid similarities, running means, no head). It is a column of the table in the common response, for all four configurations. The refinement alone gains +0.2 to +1.2 over NCM, and IPA-Approx is above it in all four configurations (+0.21 to +2.37; common response).

W3 / Q3. Thank you for noticing. We became aware of a contamination of the MiniImageNet evaluation, one image pool served the base session, the stream and the evaluation, and re-ran all methods on a disjoint 500/100 per-class split with retrained backbones (S0 / S8 / AA, %).

DINOResNet-12
NCM85.45 / 71.95 / 77.7383.33 / 61.34 / 70.87
IPA86.55 / 74.34 / 79.6483.70 / 62.80 / 72.22
IPA-Approx86.66 / 73.23 / 78.7983.64 / 63.20 / 72.32
ER91.27 / 69.00 / 78.5784.61 / 61.56 / 71.54
Finetune, EWC91.27 / 15.3 / 20.584.61 / 4.2 / 14.9
Joint (oracle)91.27 / 86.92 / 88.8784.61 / 70.31 / 75.53

The CE head now starts above NCM (84.61 vs 83.33 on ResNet-12, 91.27 vs 85.45 on DINO), the ordering you expected. Note that protocol of Table 1 and this table differ slighty (at most .2).

Thank you for the thorough rebuttal and the additional controls, which address my main concerns. I will be raising my score accordingly. I hope the revised version discusses the beta sweep in the main text and whether the gain of the BC term on DINO features reflects compression of class-irrelevant information or a normalisation of the features. This would strengthen the scientific story of this work.

Thank you for the constructive discussion. We will move the β sweep into the main text.

To separate the BC's compression vs normalisation effects, we repeated the β comparison of IPA-Approx on DINO features where we normalise the feature to test if compression still adds gain. The table reports the average accuracy over sessions (AA, %), with the gain over β = 0 in parentheses (30 repetitions for raw features).

DINO featuresβ = 0β = 1β = 8
CIFAR-100, raw67.5469.53 (+1.98)71.62 (+4.07)
CIFAR-100, normalised67.5568.07 (+0.52)70.34 (+2.79)
CUB200, raw69.8174.18 (+4.37)77.82 (+8.00)
CUB200, normalised69.4270.99 (+1.57)75.27 (+5.85)

Feature normalisation alone does not change accuracy at β = 0 (67.55 against 67.54 on CIFAR-100, 69.42 against 69.81 on CUB200). At β = 1 the gain of the BC term shrinks, because of tge reduced input norm (from about 63 to about 20) and the strength of the BC term scales with the input norm. At β = 8 most of the gain returns (+2.79 and +5.85, about 70% of the gain on raw features), and β = 1 already improves over β = 0 in every repetition (p < 1e-6 on both datasets). If the BC term only normalised the features, normalisation would remove its effect at every β.

We therefore read the gain as a reduction of co-activation between classes based on the same features. This is consistent with the compression we measure in the information plane of the toy experiments (Appendix C), although on DINO we measure accuracy rather than information.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response to all reviewers

We would like to thank all reviewers and are encouraged by their recognition of the novelty of the method and of its application to test-time adaptation. Reviewer-specific points follow in the individual replies.

Framing the contribution. IPA is a mechanism for test-time adaptation based on the IB. Its novelty lies in a closed-form, gradient-free Hebbian head that adapts one unlabeled sample at a time, without replay or a query batch. We will more clearly position our work as such and emphasize its the novelty and efficacy for test-time adaptation, without overstating outperformance.

Effectiveness of the plastic head. All reviews ask for the source of the gain of IPA-Approx, i.e., how much derives from the plastic head rather than from the prototype construction and the confidence-weighted refinement. Here we found that the head accounts for +0.2 to +0.3 AA on ResNet-12 features and +0.9 to +2.4 on self-supervised DINO features (+6.0 on CUB200).

The requested control. soft-NCM-r is IPA-Approx without the Hebbian head, i.e., the same confidence-weighted prototype refinement, with assignments from the current de-meaned unit prototypes and the same preprocessing, initialization, τ = 0.25, running counts, sample order and budget. AA (%), 30 repetitions shared by all columns, paired tests.

NCMsoft-NCM-r (no head)IPA-Approxhead
CIFAR-100 / ResNet-1266.8468.0468.25+0.21 (p = 1.5e-4, 30/30)
CIFAR-100 / DINO66.0567.1669.53+2.37 (30/30)
MiniImageNet / DINO77.5677.7878.69+0.91 (30/30)
MiniImageNet / ResNet-1270.8171.9372.21+0.28 (p = 6e-4, 30/30)
CUB200 / DINO72.0768.1474.18+6.04 (30/30)

The rest of the gain over NCM comes from the construction (de-meaned, unit-normalised centroids) and from the refinement; the full decomposition is in the reply to reviewer Thch.

Further backbones and datasets. We added two frozen backbones, DINOv2 ViT-S/14 and CLIP ViT-L/14, on CIFAR-100 and MiniImageNet, and the CUB200 dataset with all three backbones in an FSCIL protocol of 100 base classes and ten 10-way 5-shot sessions (last row above). The head's gain over the matched soft-NCM-r control is positive in all nine backbone–dataset combinations (paired p < 10⁻⁴, range +0.53 to +6.04). Where the refinement alone falls below plain NCM (68.14 vs 72.07 on CUB200 with DINO, 70.91 vs 72.60 on CIFAR-100 with CLIP), the head restores the accuracy.

β, the weight of the BC term. We further analysed the effectiveness of the BC term by sweeping β (AA %, IPA-Approx, 30 repetitions).

β00.512
CIFAR-100 / ResNet-1268.1468.2268.2568.23
CIFAR-100 / DINO67.5468.6469.5370.66
CUB200 / DINO69.8172.6674.1875.89

On DINO features accuracy keeps rising up to β = 8 (71.62 on CIFAR-100, 77.82 on CUB200). At β = 0 the head stays a scaled prototype classifier (§3.2), so on these features most of the head's gain comes from the BC term, which reduces the weights of features that co-activate with a row's output. The effect of β thus depends on the features; β = 1, the value of every reported result, was fixed before these experiments, and the sweep is reported as an analysis.

MiniImageNet FSCIL, correction. Our MiniImageNet numbers were inflated by data contamination (one image pool served the base session, the stream and the evaluation). All methods were re-run on a disjoint 500/100 per-class split with retrained ResNet-12 backbones (MiniImageNet rows above; full numbers in the reply to reviewer Thch). The ordering of the methods is unchanged, and CIFAR-100 is unaffected.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.