acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Semantic Evolution for Semi-Supervised Few-Shot Incremental Learning

Abstract

Semi-Supervised Few-Shot Class Incremental Learning (SSFSCIL) aims to continuously acquire novel concepts from scarce annotations and abundant unlabeled data while preserving previously learned knowledge. However, existing approaches mainly rely on static pseudo-labeling and incremental adaptation, which struggle with three fundamental challenges: unreliable semantic discovery under limited supervision, unstable representation evolution across sessions, and insufficient decision boundary characterization with few-shot samples. In this paper, we introduce DSE-SSF, a novel framework that reformulates continuous SSFSCIL as a process of dynamic semantic evolution rather than sequential parameter optimization. Based on this insight, DSE-SSF establishes a closed-loop learning paradigm consisting of semantic discovery, knowledge stabilization, and boundary refinement, implemented by three novel mechanisms: (1) Dynamic Semantic Graph Evolution progressively uncovers reliable semantic structures from unlabeled data through evolving manifold propagation; (2) Adaptive Dual-Expert Stabilization leverages temporal consistency and adaptive uncertainty calibration to prevent incremental knowledge drift; and (3) Boundary-aware Feature Manifold Exploration explicitly investigates uncertain feature regions to construct robust decision boundaries under sparse supervision. By jointly modeling the evolution of semantic topology, representation stability, and boundary geometry, DSE-SSF enables reliable knowledge acquisition throughout continuous incremental learning. Extensive experiments on SSFSCIL benchmarks demonstrate that DSE-SSF consistently outperforms existing methods, validating the effectiveness of dynamic semantic evolution for long-term few-shot visual learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
semi-supervised learning, few-shot class-incremental learning, continual learning, pseudo-labeling

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_c4dyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DSE-SSF for semi-supervised few-shot class-incremental learning. It trains a separate LoRA expert for each session on a frozen ViT-B/16. DSGE refreshes a kNN graph during training and propagates labels to unlabeled data. ADES selects pseudo-labels using an EMA expert, a consistency loss and an adaptive threshold. BFME perturbs features toward the decision boundary and enforces prediction consistency. The experiments cover CIFAR-100, CUB-200 and ImageNet-100, with diagnostics of pseudo-label error and old-class drift and a theoretical analysis in the appendix.

Critical strengths
  1. All of them address the failure modes that were measured at the start in Figure 1.
  2. On CUB-200, DSE-SSF outperforms DiL in final accuracy (83.98 vs. 80.05). Removing any of the three parts on CIFAR-100 reduces the final accuracy by 3-5 points (Table 2) and exceeds the variation among seeds.
Critical weaknesses
  1. The novelty is limited. DSGE is similar to the iterative graph propagation of Iscen et al. (2019) and has a hard agreement filter. It is difficult to reconcile with the claim on line 324 that "existing graph methods construct a fixed graph once"; ADES is essentially Mean Teacher with a KL-based threshold, and BFME is feature-space VAT. The paper does not formalise "semantic evolution", and it also provides no ablations for the above canonical methods despite citing all of them, or comparisons with strong CIL baselines using pretrained backbones. Therefore, based on the above results, it cannot be determined whether the improvements are due to the new modules or the frozen ViT and separate LoRA experts for each session.
  2. Some of the empirical claims are not supported by the results. (1) Table 1 does not support the claim that DSE-SSF "consistently outperforms" the existing methods. Final accuracy on CIFAR-100 is 83.82±1.90 for DSE-SSF and 83.68±0.33 for DSGD; Δ\Delta is 10.01 for DSE-SSF, compared to 7.23 for DSGD and 9.88 for KeepLoRA++. Avg. accuracy on ImageNet-100 for DSE-SSF is 86.67; both DSGD (89.20) and KeepLoRA++ (87.88) are higher. (2) based on Δ=Acc0−AccT\Delta = \mathrm{Acc}_0 - \mathrm{Acc}_T, the estimated base-session accuracy on ImageNet-100 for DSE-SSF is approximately 87.2%, and that of DSGD, KeepLoRA++, and DTE are 93.3%, 92.7%, and 92.5%, respectively. DSE-SSF 5-6 points are lower in the base session. (3) DSGE has little impact on new-class accuracy. Removing it reduces average new-class accuracy by only 0.93 points, from 95.45 ± 1.56 to 94.52 ± 1.68. The above difference is within the given standard deviation and has little margin for error at this accuracy. The above results are inconsistent with the idea that DSGE "enhances novel-class acquisition" (L. 461-462).
  3. Errors exist in the formulation and theory. (1) Equation (11) is inconsistent with the behaviour mentioned in Section 3.3. Mean KL divergence U‾e\overline{U}^{e} is non-negative; thus, only increasing λselect\lambda_{\mathrm{select}} is possible. It can never be lowered as said. (2) Theorem 1 is not applicable as presented. The nbn_b explored points depend on the model trained on the nn labeled samples and their augmentations; they are not i.i.d. draws from the data distribution. Therefore, Rademacher's inequality in Eq. (47) is not applicable to the augmented sample of size n+nbn + n_b. Then, just kk-fold replication of the training set would reduce the complexity and confidence terms by k\sqrt{k} without introducing new information. (3) The second inequality in Eq. (50) is in the wrong direction. Substitution of Rnn/(n+nb)R_{n}\sqrt{n/(n+n_{b})} by Rn+nbR_{n+n_{b}} requires the inverse of Eq. (49). Replacement of R^u0noisy\hat{R}_{u_0}^{\text{noisy}} by R^u0\hat{R}_{u_0} also requires the reverse of Eq. (48).
Questions for authors
  1. How does DSGE differ from Iscen et al. (2019) or DTE beyond replacing entropy weighting with a hard agreement filter? Have you made direct comparisons?
  2. What happens when ADES is replaced with plain Mean Teacher and BFME with feature-space VAT?
  3. Is BFME used only for labelled data, or also on pseudo-labelled data in D^t\hat{D}^t? If it contains the latter, then how does an incorrect pseudo-label affect LBFMEL_{BFME}?
  4. Frozen ViT-B/16 with a prototype classifier or RanPAC how do they perform under the above protocol? DSE-SSF's performance without unlabeled data?
  5. How is the expert chosen at test-time when task IDs are unknown? Router is mentioned only once in Appendix C.3.
  6. Given that the shared projection and LayerNorm are frozen, and each session has its own LoRA (L. 415), is the 0% drift in Figure 3(b) guaranteed by design? Drift without ADES that you have observed?
  7. Is the EMA in Eq. (9) updated at the end of an epoch or after each iteration? m = 0.99 and 30 epochs; thus, an update per epoch would leave the teacher near its initialisation.
7 comments
AuthorsOpenReview ↗

Response to Reviewer c4dy

We thank the reviewer for the thorough, constructive critique and for recognizing our quantitative motivation diagnostics and three-seed ablations. Your detailed questions regarding canonical baselines, active-expert drift, EMA dynamics, and theoretical bounds have helped us substantially strengthen both our empirical evidence and mathematical presentation in the revised manuscript.

Q1. How Does DSGE Differ from Iscen et al. [1] and DTE [4]?

In DSE-SSF, "dynamic semantic evolution" refers to a concrete closed-loop feedback mechanism during incremental expert adaptation (Secs. 3.2–3.3): as the active LoRA expert updates its representation every η=5\eta=5 epochs, its kk-NN manifold topology is reconstructed; sample selection then requires joint agreement between graph propagation and the expert's local cosine classifier (Eq. 7), subject to margin constraints, per-class budgets, and an uncertainty threshold dynamically adjusted by the temporal expert's disagreement (Eq. 11).

We have revised Sec. 2 and Sec. 3.2 (Line 324) to explicitly credit Iscen et al. [1] for alternating representation learning and label propagation, and focused our distinction from both [1] and DTE [4] on our multi-criteria graph-model agreement gate and temporal uncertainty feedback. In direct controlled comparisons within the exact same multi-expert architecture, Full outperforms LP-adapted [1] by +0.83±0.71+0.83 \pm 0.71 Final and +0.91±0.35+0.91 \pm 0.35 Average points on ImageNet-R, and achieves +0.55±0.27+0.55 \pm 0.27 points higher final Fixed-Base retention on CUB-200 (Table R1, positive in all three seeds), complementing our 11–26-point11\text{--}26\text{-point} gains over DTE [4] in Table 1.

Q2. Replacing ADES with Mean Teacher [2] and BFME with Feature-Space VAT [3]

We have implemented and evaluated both canonical replacements within our exact expert architecture under identical initialization, support sets, and update budgets (1,2001{,}200 joint steps; see Response to Reviewer hQ1L, W1 and Table R1):

  • MT-adapted uses independent augmentations, unidirectional teacher-to-student KL (temperature 11), fixed-threshold selection (removing ADES's adaptive threshold feedback), and step-wise EMA (mstep=0.991/4≈0.99749m_{\mathrm{step}} = 0.99^{1/4} \approx 0.99749).
  • Feature-space VAT uses label-agnostic predictive KL divergence on normalized features (ℓ2\ell_2 radius 0.050.05, finite-difference scale 0.0010.001, 11 power iteration), whereas BFME uses label-guided cross-entropy for 33 ℓ∞\ell_\infty-projected ascent steps (radius 0.0050.005) followed by symmetric KL.

On CUB-200 (Table R1), Full outperforms MT-adapted by +0.14±0.12+0.14 \pm 0.12 paired Final points and Feature-space VAT by +0.22±0.08+0.22 \pm 0.08 paired Final points, with positive gains across all three seeds. On ImageNet-R, their overall accuracies are closely matched to Full (MT-adapted slightly higher, Feature-space VAT slightly lower). Because each single-component replacement retains the other two native DSE-SSF mechanisms in the loop, these experiments complement our component-removal ablations (Table 2, where removing ADES or BFME drops CIFAR-100 Final accuracy by 3.493.49 and 5.265.26 points) and our direct mechanism probes (Table R3).

Q3. Does BFME Use Pseudo-Labels, and What Happens When a Pseudo-Label Is Wrong?

Yes, BFME is applied to the combined training batch drawn from both the 55-shot labeled support set Dlt\mathcal{D}_l^t and the selected reliable pseudo-labeled set D^t\hat{\mathcal{D}}^t. Specifically, the sample's label yy (true support label or selected pseudo-label) guides the inner ascent cross-entropy in Eqs. (13)–(14) to locate low-confidence boundary regions, while the outer loss in Eq. (15) enforces label-free symmetric-KL consistency between the reference feature h′h' and the explored feature h′+ϵh'+\epsilon.

If an admitted pseudo-label is incorrect, it can steer the inner perturbation ϵ\epsilon toward the wrong class boundary, and the outer consistency loss does not by itself correct the label error. This is precisely why DSGE and ADES's selective filtering is placed upstream of BFME. In our three-seed Full runs, the pseudo-labels admitted into D^t\hat{\mathcal{D}}^t achieve an empirical precision of 98.57±0.3298.57 \pm 0.32\\% (at 54.66±0.4254.66 \pm 0.42\\% coverage) on CUB-200 and 98.26±0.4298.26 \pm 0.42\\% (at 33.62±0.2833.62 \pm 0.28\\% coverage) on ImageNet-R. Thus, $ > 98\%$ of the pseudo-labeled samples entering BFME carry true class labels, ensuring that boundary exploration predominantly targets genuine decision boundaries. We have clarified this sample scope and the distinct roles of labels in Eqs. (14) and (15) in Sec. 3.4 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer c4dy

Q4. Frozen ViT, RanPAC, and DSE-SSF Without Unlabeled Data (No-U)

To directly test whether our performance comes merely from the frozen ViT-B/16 and per-session LoRA isolation rather than unlabeled semantic evolution, we added a strict No-U (No Unlabeled Data) control on ImageNet-R (Sec. 4.2 & Table 6, Appendix C.3). No-U keeps the exact same frozen ViT-B/16, SHA-256-matched Base checkpoints, per-session LoRA experts, and 55 labeled supports per incremental class, while completely disabling all unlabeled training, pseudo-label calibration, and unlabeled routing records (verified to have zero unlabeled sample usage across all seeds).

Holding the backbone and expert architecture identical, exploiting unlabeled data via Full improves ImageNet-R Final accuracy by +1.42±0.79+1.42 \pm 0.79 points and Average accuracy by +0.82±0.14+0.82 \pm 0.14 points over No-U, with positive gains in every seed (driven by stronger novel-class acquisition). While training-free prototype classifiers and closed-form random-projection methods like RanPAC explore orthogonal classifier designs on frozen features, our matched No-U ablation and Table 2 ablations (+3.07+3.07 to +5.26+5.26 points on CIFAR-100) directly confirm that DSE-SSF's semi-supervised evolutionary training provides substantial gains beyond the frozen ViT and LoRA architecture, with minimal overhead (1.18M1.18\mathrm{M} trainable LoRA parameters per session; Table 8, Appendix C.4).

Q5. How Is the Expert Chosen at Test Time Without Task IDs?

No task ID is used at test time. Similar to prototype-based routing in continual LoRA methods (e.g., CL-LoRA), all experts are evaluated in parallel and combined via input-dependent fusion weights gt,j(x)g_{t,j}(x) computed from cached feature prototypes and calibrated output statistics (storing no raw images). We moved this complete formulation into Sec. 3.1 and detailed the feature-cache budget in Appendix C.1.1.

Q6. Is the 0% Drift in Fig. 3(b) Guaranteed by Freezing? What Is the Drift Without ADES?

You are spot-on regarding Fig. 3(b): because the Base expert and shared projection are frozen during incremental sessions, the Base feature path is invariant by construction. The actual target of ADES is to stabilize the trainable active expert Et\mathcal{E}_t (whose LoRA weights and cosine classifier are actively updated via gradient descent during the 3030 joint-training epochs).

To directly measure whether ADES prevents representation drift in the trainable parameters, we tracked the active expert's feature shift from the end of 55-shot support warm-up to the end of joint semi-supervised training within each session across ImageNet-R and CUB-200 (Table 5, Appendix C.3).

Table 5. Trainable active-expert representation stability on ImageNet-R and CUB-200. Session averages within each seed, reported as three-seed mean ±\pm sample SD over fixed class-balanced probes (≤512\le 512 samples/session). Cosine Drift is 1−cos⁡(fafter,fbefore)1 - \cos(f_{\mathrm{after}}, f_{\mathrm{before}}); kk-NN-1010 Churn is the fraction of changed neighbors.

ConfigurationImageNet-R Drift ↓\downarrowImageNet-R Churn (%) ↓\downarrowCUB-200 Drift ↓\downarrowCUB-200 Churn (%) ↓\downarrow
Full (with ADES)0.01529±0.000870.01529 \pm 0.0008714.63±0.7714.63 \pm 0.770.02726±0.002470.02726 \pm 0.0024713.74±0.2613.74 \pm 0.26
w/o ADES0.03361±0.003320.03361 \pm 0.0033221.05±1.5921.05 \pm 1.590.06464±0.000980.06464 \pm 0.0009819.41±0.2619.41 \pm 0.26
w/o companion consistency0.03267±0.003700.03267 \pm 0.0037020.82±1.6220.82 \pm 1.620.06037±0.001890.06037 \pm 0.0018918.58±0.3918.58 \pm 0.39

Removing ADES causes the active expert's feature drift to more than double, increasing by +119.8%+119.8\% on ImageNet-R (ADES reduces drift by 54.5%54.5\%) and by +137.1%+137.1\% on CUB-200 (ADES reduces drift by 57.8%57.8\%), while kk-NN-1010 churn jumps from 14.63%→21.05%14.63\% \to 21.05\% and 13.74%→19.41%13.74\% \to 19.41\%. We explicitly clarified Fig. 3(b)'s frozen-Base invariance in Sec. 4.3, Fig. 3's caption, and Appendix C.3, and reported these direct trainable active-expert measurements in Table 5 (alongside fixed-Base cohort retention in Table 4).

AuthorsOpenReview ↗

Response to Reviewer c4dy

Q7. Is EMA Updated per Epoch or per Step?

In Full, the EMA temporal expert is updated once per epoch (3030 updates over the 3030 joint-training epochs, at 44 optimization steps per epoch), matching the epoch index ee in Eq. (9). Why does m=0.99m=0.99 over 3030 epochs work well here? Because the temporal expert E~t\tilde{\mathcal{E}}_t is initialized after the 88 supervised warm-up epochs on the 55-shot support set Dlt\mathcal{D}_l^t, rather than from random or cold parameters. Thus, E~t\tilde{\mathcal{E}}_t starts as an already support-adapted model and acts as a conservative anchor that retains 0.9930≈0.7400.99^{30} \approx 0.740 weight on the clean support-warmed state while gradually incorporating (1−0.740)=26%(1 - 0.740) = 26\% of the unlabeled joint-training trajectory. (In MT-adapted in Table 3, we also tested step-wise EMA with matched terminal decay mstep=0.991/4≈0.99749m_{\mathrm{step}} = 0.99^{1/4} \approx 0.99749.) We made this post-warm-up initialization and update schedule explicit in Sec. 3.3 and Appendix C.1.2.

Response to Weaknesses W2 (Claims, Base Gap, New-Class Gain) & W3 (Eq. 11 & Theory)

  • Calibrating claims & explaining the ImageNet-100 Base gap (W2.1 & W2.2): We calibrated our performance claims in the Abstract, Sec. 1, Sec. 4.1, and Appendix C.1.3 to reflect exact dataset and metric results rather than blanket "consistently outperforms" phrasing. Regarding ImageNet-100 (6060 Base + 8×58\times 5 incremental classes): Table 1's Final (86.17±1.65%86.17 \pm 1.65\% vs. 85.90±0.31%85.90 \pm 0.31\%) and Δ\Delta (1.05±1.071.05 \pm 1.07 vs. 7.43±0.177.43 \pm 0.17) imply a Session-0 Base accuracy of 87.22%87.22\% for DSE-SSF vs. 93.33%93.33\% for DSGD (as baselines use their native Base training recipes while DSE-SSF adapts rank-88 LoRA, LayerNorm, a 512512-d projection, and a cosine head). Far from weakening our incremental contribution, this shows that DSE-SSF's Final accuracy on ImageNet-100 (86.17%86.17\%) is achieved despite starting 6.116.11 points lower in Base accuracy—dropping only 1.051.05 points across 88 incremental sessions compared to 7.437.43 points for DSGD (while the lower Session-0 starting point mechanically lowers the 9-session Avg. mean). Furthermore, all our new component and replacement experiments use SHA-256-identical Base checkpoints per seed to eliminate any starting-checkpoint differences.
  • Global new-class acquisition without the local-head ceiling (W2.3): In Table 2 (CIFAR-100), Avg. New was evaluated on the active expert's 55-way local head in isolation, where accuracy is already near the ceiling (∼95%\sim 95\%). When we evaluate the 100100 incremental classes on ImageNet-R under full 200200-class global competition (Sec. 4.2 & Table 6, Appendix C.3), adding DSGE improves novel-class recognition by +6.25±1.72+6.25 \pm 1.72 points (positive in every seed) and improves overall 200200-class Final accuracy by +1.16±0.73+1.16 \pm 0.73 points, directly substantiating our novel-class acquisition claim.
  • Fixing Eq. (11) and Theorem 1 / Eq. (52) (W3): Thank you for catching the typographical recurrence in Eq. (11) and the reversed substitutions in the old Theorem 1 proof. In our implementation, the selection threshold at epoch ee, expansion round jj is recomputed relative to the fixed base threshold λ0\lambda_0 rather than accumulated recursively:

λe,j=clip[0.30, 0.85](λ0+αUˉe,j),\lambda_{e,j}=\mathrm{clip}_{[0.30,\,0.85]}\left(\lambda_0+\alpha\bar{U}_{e,j}\right),

where Uˉe,j\bar{U}_{e,j} is the mean untempered KL(pEMA∥pstudent)\mathrm{KL}(p_{\mathrm{EMA}}\|p_{\mathrm{student}}) over the remaining unlabeled pool. Thus, as teacher-student disagreement Uˉe,j\bar{U}_{e,j} decreases during training, λe,j\lambda_{e,j} naturally decreases (loosens) toward λ0\lambda_0. We corrected Eq. (11) in Sec. 3.3 and updated the schematic expression in Fig. 2. Finally, as detailed in our response to Reviewer 9jbN (Q1), we retained the standard nn-sample bound in Eq. (47) and replaced the n+nbn+n_b independent-sample counting and one-sided substitutions in Theorem 1 with our tight, dependency-free pairwise margin stability bound (μ(h′+ϵ,y)≥μ(h′,y)−2LBFME\mu(\mathbf{h}'+\epsilon, y) \geq \mu(\mathbf{h}', y) - \sqrt{2\mathcal{L}_{\mathrm{BFME}}}) and corrected Eq. (52) in Appendix A.

Reviewer_hQ1LOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DSE-SSF for semi-supervised few-shot class-incremental learning. The framework uses session-specific LoRA experts on a frozen pretrained backbone and combines three components. Dynamic Semantic Graph Evolution updates a neighborhood graph and propagates labels to select reliable pseudo-labeled samples. Adaptive Dual-Expert Stabilization combines temporal consistency with uncertainty-aware pseudo-label selection. Boundary-aware Feature Manifold Exploration enforces prediction consistency under adversarial feature perturbations.

Critical strengths
  1. The paper is well written and well structured.

  2. The method connects dynamic graph-based pseudo-labeling, temporal consistency, and boundary refinement within a single learning loop. Updating the graph alongside the learned representations allows pseudo supervision to evolve during training. This is a useful design for learning new classes under sparse supervision.

Critical weaknesses
  1. The additional contribution over prior work needs clearer evidence. Iscen et al. [1] already combine iterative graph-based pseudo-labeling with Mean Teacher [2]. BFME shares the local adversarial-consistency principle of VAT [3], although its perturbation objective differs. DTE [4] also uses graph-based label propagation in SSFSCIL. These works are cited, but the benefits of the proposed modifications remain insufficiently isolated. The removal and static-graph ablations support the usefulness of the components. Comparisons with established alternatives under the same expert architecture would help assess the methodological advance.

  2. Global prediction and knowledge retention need clearer validation. The method uses session-specific experts and local classifiers. Appendix C.3 mentions router calibration with cached features, but the routing rule and retained information are not clearly specified. These details are important for understanding prediction across all seen classes. Moreover, Δ and RPD compare accuracies over changing class sets, as acknowledged in Appendix C.2. They do not directly measure old-class forgetting. A complete inference protocol and accuracy on a fixed set of old classes across sessions would help assess the claimed knowledge stabilization.

[1] Iscen et al. Label Propagation for Deep Semi-Supervised Learning (https://arxiv.org/abs/1904.04717). CVPR, 2019.

[2] Tarvainen and Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results (https://arxiv.org/abs/1703.01780). NeurIPS, 2017.

[3] Miyato et al. Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning (https://arxiv.org/abs/1704.03976). IEEE TPAMI, 2019.

[4] Han et al. Semi-supervised few-shot class-incremental learning based on dynamic topology evolution (https://doi.org/10.1016/j.engappai.2024.108528). Engineering Applications of Artificial Intelligence, 2024.

Questions for authors

Please see Critical Weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer hQ1L

We thank the reviewer for the positive assessment of our paper's structure and for recognizing the value of coupling dynamic graph propagation, temporal consistency, and boundary exploration. Below, we present the direct comparisons against canonical alternatives within the same architecture (W1) and the fixed-cohort old-class retention trajectories under global prediction (W2) that we have added to the revised manuscript.

W1. Methodological Advance and Direct Comparisons with Established Alternatives

While DSE-SSF draws inspiration from foundational semi-supervised principles, its core advance lies in how graph discovery, temporal uncertainty, and boundary exploration are coupled into a mutually reinforcing loop tailored for few-shot incremental experts (Secs. 3.2–3.4 of the paper):

  1. DSGE vs. Iscen et al. [1] and DTE [4]: While Iscen et al. [1] alternate representation learning and label propagation using entropy weights, DSGE introduces a selective admission gate that requires strict agreement between the propagated graph prediction and the active expert's local prediction (Eq. 7), combined with margin constraints, per-class budgets, and an uncertainty threshold dynamically modulated by the temporal expert. Unlike DTE [4], which propagates over Voronoi partitions and static Hebbian category graphs, DSGE continuously reconstructs the kk-NN topology as the active LoRA expert adapts.
  2. ADES vs. Mean Teacher [2]: Beyond standard EMA weight averaging, ADES uses the predictive KL divergence between the active and temporal experts over the unlabeled pool as an intrinsic session-level stability gauge to dynamically tighten or relax DSGE's sample-admission threshold (Eq. 11).
  3. BFME vs. VAT [3]: Whereas VAT [3] searches for label-agnostic adversarial directions by maximizing predictive divergence, BFME uses label-guided cross-entropy (over supports and high-precision pseudo-labels) to steer perturbations specifically toward true decision boundaries within an ℓ∞\ell_\infty ball, and then enforces symmetric-KL consistency (Eq. 15).

Direct replacements within the exact same architecture. Following your suggestion, we have implemented and added direct canonical replacements inside our exact multi-expert architecture on both ImageNet-R and CUB-200:

  • LP-adapted replaces DSGE's selective admission with entropy-weighted label propagation and class balancing [1];
  • MT-adapted replaces ADES with a standard Mean Teacher (independent views, unidirectional KL, fixed selection threshold, and step-wise EMA) [2];
  • Feature-space VAT replaces BFME with label-agnostic virtual adversarial perturbation in the normalized feature space [3].

Within each seed (4242, 10241024, 19931993), all variants start from the exact same SHA-256-verified Base checkpoint, use the same 55-shot supports, and run for the exact same 1,2001{,}200 joint optimization updates. Each variant replaces one module family while keeping the remaining two native DSE-SSF mechanisms active (note that LP-adapted and MT-adapted naturally disable the cross-module threshold feedback between ADES and DSGE).

AuthorsOpenReview ↗

Response to Reviewer hQ1L

W1. Direct-Comparison Results

Under the matched protocol described in Part 1, the controlled results are as follows.

On ImageNet-R (where rendition shifts make initial features noisy), Full outperforms LP-adapted by +0.83±0.71+0.83 \pm 0.71 Final and +0.91±0.35+0.91 \pm 0.35 Average points, while MT-adapted and Feature-space VAT achieve overall accuracies close to Full (with MT-adapted slightly higher and Feature-space VAT slightly lower). On CUB-200, Table R1 reports both overall cumulative accuracy and final fixed-Base retention at Session 10 to capture both novel-class acquisition and old-class preservation.

Table R1. Direct canonical replacements within the same expert architecture on CUB-200. Mean ±\pm sample SD over seeds 4242, 10241024, and 19931993 (100100 Base classes + ten 1010-class, 55-shot sessions; 1,2001{,}200 joint updates). Average includes Session 0.

ConfigurationFinal Accuracy (%)Average Accuracy (%)Final Fixed-Base Accuracy at Session 10 (%)
Full (DSE-SSF)84.23±0.2184.23 \pm 0.2185.18±0.0485.18 \pm 0.0486.03±0.2886.03 \pm 0.28
LP-adapted [1]84.36±0.2684.36 \pm 0.2685.17±0.0885.17 \pm 0.0885.49±0.2185.49 \pm 0.21
MT-adapted [2]84.09±0.1484.09 \pm 0.1485.16±0.0885.16 \pm 0.0885.98±0.1285.98 \pm 0.12
Feature-space VAT [3]84.01±0.2584.01 \pm 0.2585.17±0.0585.17 \pm 0.0585.94±0.1885.94 \pm 0.18

As shown in Table R1, when testing single-component substitutions while the other two native modules remain in place (e.g., graph-based selection still maintains high pseudo-label precision of 98.57±0.3298.57 \pm 0.32\\% for Full, 98.05±0.4598.05 \pm 0.45\\% for MT-adapted, and 98.55±0.3598.55 \pm 0.35\\% for Feature-space VAT), Full consistently surpasses both MT-adapted (+0.14±0.12+0.14 \pm 0.12 paired Final points) and Feature-space VAT (+0.22±0.08+0.22 \pm 0.08 paired Final points) in every seed. Comparing Full with LP-adapted reveals a clear plasticity-stability trade-off: while LP-adapted attains a slightly higher overall Final score on CUB-200 (+0.14±0.08+0.14 \pm 0.08 paired points), it causes greater degradation on old classes, whereas Full preserves +0.55±0.27+0.55 \pm 0.27 points higher Fixed-Base accuracy at Session 10 (positive in every seed) and outperforms LP-adapted by +0.83+0.83 Final / +0.91+0.91 Average points on the harder ImageNet-R benchmark. We have incorporated these comparisons and discussions into Sec. 2, Sec. 4.2, and Appendix C.

AuthorsOpenReview ↗

Response to Reviewer hQ1L

W2. Global Prediction Protocol and Fixed-Cohort Old-Class Retention

1. Complete global inference protocol.

As detailed in our response to Reviewer 9jbN (Q2), at test session tt, every retained expert j∈0,…,tj \in \\{0, \dots, t\\} produces local class probabilities qj(c∣x)q_j(c\mid x) for c∈Cjc \in \mathcal{C}_j, which are weighted by prototype- and feature-calibrated fusion weights gt,j(x)g_{t,j}(x) to form a unified probability distribution pt(c∣x)=gt,j(x)qj(c∣x)p_t(c\mid x) = g_{t,j}(x)q_j(c\mid x) over all seen classes C≤t\mathcal{C}_{\leq t}. Because the ViT-B/16 backbone is frozen, calibration only requires lightweight cached feature embeddings and class prototypes—no raw images are stored, expert parameters remain frozen during calibration, and neither hidden unlabeled class labels nor test labels are ever accessed.

2. Fixed-cohort old-class retention across sessions.

We fully agree that aggregate metrics (Δ\Delta and RPD) reflect accuracy changes over an expanding label set rather than pure old-class forgetting (Appendix C.2). To directly measure old-class knowledge retention under global competition, let Bk\mathcal{B}_k denote the fixed test set of classes introduced at session kk. At any subsequent session t≥kt \geq k, we evaluate its fixed-cohort accuracy while all seen classes C≤t\mathcal{C}_{\leq t} actively compete:

\sum_{(x,y)\in\mathcal{B}_k}\mathbf{1}\\{\hat{y}_t(x)=y\\}, \qquad t\geq k.$$ Table R2 reports the session-by-session fixed-Base retention trajectory ($A_{0,t}$) from our matched three-seed CUB-200 study. Both `Full` and `w/o ADES` start from the exact same Base checkpoint in each seed (where Base-only accuracy at Session 0 is $88.43 \pm 0.54\\%$). We track how accuracy on the exact same Base test cohort evolves from the first incremental session ($t_1$, competing against $110$ classes) to the final session ($t_{10}$, competing against all $200$ classes). **Table R2. Fixed-Base retention trajectory ($A_{0,t}$) on CUB-200 under global class competition.** Mean $\pm$ sample SD over three seeds. Decline is computed within each seed as $A_{0,1}-A_{0,10}$. | Configuration | Session 1 ($110$ classes, %) | Session 5 ($150$ classes, %) | Session 10 ($200$ classes, %) | Incremental Decline ($A_{0,1}-A_{0,10}$, points) $\downarrow$ | |---|---:|---:|---:|---:| | **`Full` (DSE-SSF)** | **$86.22 \pm 0.44$** | **$86.41 \pm 0.40$** | **$86.03 \pm 0.28$** | **$0.19 \pm 0.17$** | | `w/o ADES` | $86.28 \pm 0.40$ | $86.20 \pm 0.11$ | $85.93 \pm 0.22$ | $0.35 \pm 0.27$ | As Table R2 demonstrates, even as the competing prediction space expands from $110$ to $200$ classes across $10$ incremental sessions, **`Full` experiences a negligible fixed-Base decline of only $0.19 \pm 0.17$ points**, reducing incremental Base forgetting by **$0.16 \pm 0.11$ paired points** relative to `w/o ADES` (positive in all three seeds). Moreover, directly probing the trainable active expert (Table R3 in c4dy Q6) confirms that ADES reduces within-session feature drift by **$57.8\\%$ on CUB-200** and **$54.5\\%$ on ImageNet-R**. We have incorporated this formal fixed-cohort retention metric and trajectory analysis into Sec. 4.2 and Appendix C.2 of the revised manuscript. **References** [1] Iscen et al. Label Propagation for Deep Semi-Supervised Learning. CVPR, 2019. [2] Tarvainen and Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017. [3] Miyato et al. Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning. IEEE TPAMI, 2019. [4] Han et al. Semi-supervised few-shot class-incremental learning based on dynamic topology evolution. EAAI, 2024.
Reviewer_9jbNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DSE-SSF, a framework for Semi-Supervised Few-Shot Class-Incremental Learning (SSFSCIL). The proposed method combines three components: (1) Dynamic Semantic Graph Evolution (DSGE), which periodically reconstructs a semantic graph and performs label propagation to select reliable pseudo-labeled samples; (2) Adaptive Dual-Expert Stabilization (ADES), which employs an EMA-based temporal expert and uncertainty-aware pseudo-label selection; and (3) Boundary-aware Feature Manifold Exploration (BFME), which introduces adversarial feature perturbations and consistency regularization to improve decision boundary robustness.

The method uses a frozen pretrained ViT backbone with expandable LoRA experts. Experiments on CIFAR-100, CUB-200, and ImageNet-100 demonstrate competitive performance, and the component ablations indicate that all three modules contribute to the final results.

Although the framework is potentially useful, I have several concerns regarding its methodological novelty, theoretical correctness, experimental comparisons, and reproducibility.

Critical strengths

The paper identifies three relevant challenges in semi-supervised few-shot class-incremental learning, namely unreliable pseudo-labeling, representation instability, and fragile decision boundaries. The proposed framework addresses these challenges through complementary graph-based semantic discovery, temporal expert stabilization, and feature-space consistency regularization. This integrated perspective is potentially useful for improving incremental learning under limited supervision.The method achieves competitive performance on three benchmarks, with particularly promising results on CUB-200, where it obtains 83.98% final accuracy and 85.07% average accuracy. The ablation results in Table 2 further demonstrate that removing DSGE, ADES, or BFME consistently reduces performance. These findings provide empirical evidence for the practical value of the proposed components.

Critical weaknesses

The theoretical analysis in Appendix A contains several questionable derivations. In particular, Theorem 1 treats boundary-explored features generated from existing samples as additional independent observations when deriving the generalization bound. This assumption is not justified, and the Rademacher complexity inequality in Eq. (49) does not generally follow without additional conditions. Furthermore, Eq. (54) incorrectly equates the symmetric KL divergence with Jensen–Shannon divergence. Proposition 2 also does not adequately account for the projected feature perturbations used in the actual implementation. These issues are important because the paper explicitly claims theoretical guarantees for the effectiveness of BFME, whereas the current proofs do not sufficiently support these claims.The empirical improvements are not consistently substantial. In Table 1, DSE-SSF improves final accuracy over DSGD by only 0.14 percentage points on CIFAR-100 and 0.27 percentage points on ImageNet-100, with relatively large standard deviations. Moreover, its average accuracy on ImageNet-100 (86.67%) is lower than DSGD (89.20%). These observations weaken the claim of consistent superiority. More importantly, the framework maintains multiple LoRA experts, but the paper does not sufficiently explain how these experts are selected or combined during task-agnostic class-incremental inference. Without a clear description and controlled evaluation of this mechanism, it is difficult to isolate the benefits of the proposed three components from the expandable-expert architecture itself.

Questions for authors

Q1. Theoretical correctness. Could the authors clarify how Theorem 1 accounts for the statistical dependence between original training samples and their boundary-explored features? In particular, what assumptions justify using (n+n_b) as the effective sample size in the generalization bound? Additionally, could the authors correct the mathematical inconsistency in Eq. (54) and clarify whether Proposition 2 remains valid under the projected perturbation constraint used in the implementation? A rigorous correction would substantially affect my assessment of the technical soundness. Q2. Incremental inference mechanism. How are the multiple session-specific LoRA experts selected, combined, or calibrated during inference without access to task identity? Could the authors provide the complete inference procedure and clarify the role of the router mentioned in Appendix C.3? It would also be important to quantify how much performance improvement comes from expert expansion and routing rather than DSGE, ADES, and BFME. Q3. Significance and fairness of experimental improvements. Given the small final-accuracy improvements over DSGD on CIFAR-100 and ImageNet-100 and the relatively large standard deviations, can the authors provide paired statistical comparisons to substantiate the claimed superiority? Furthermore, can the authors clarify whether the compared methods use identical base checkpoints and explain the substantially different baseline accuracies reported in Table 1 and the DTE-related diagnostic in Table 4? These clarifications would help determine whether the empirical gains can be reliably attributed to the proposed method.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9jbN

We thank the reviewer for highlighting our motivating diagnostics, CUB-200 gains, and ablation studies. Below, we detail the dependency-free theoretical formulation for BFME, the task-agnostic prototype/feature routing mechanism, and the experimental comparisons that we have incorporated into the revised manuscript.

Q1. BFME Theory: Dependent Features, Consistency, and Projected Exploration

We appreciate the reviewer's careful reading of Appendix A. You are completely right that boundary-explored features depend on the training samples, that Eq. (54) conflated symmetric KL with Jensen-Shannon divergence (JSD), and that the finite ℓ∞\ell_\infty-projected ascent should be stated explicitly. Importantly, the theoretical guarantee for BFME does not require treating explored features as independent draws of size n+nbn+n_b—instead, it follows directly from a dependency-free pairwise margin stability bound built on Eq. (15) and Lemma 1 (Appendix A.1.3).

Specifically, for local classifier distributions p,qp, q (at temperature 11 with natural logarithms), let the symmetric KL loss and the prediction margin for class yy be defined as:

\qquad M_y(p)=p_y-\max_{c\ne y}p_c.$$ Because $p$ and $q$ are valid probability distributions, $\sum_c (p_c - q_c) = 0$, which implies that the sum of positive coordinate differences equals the sum of negative coordinate differences, both equaling the total variation distance $\operatorname{TV}(p,q)=\tfrac{1}{2}\\|p-q\\|_1$. Combining this exact identity with Pinsker's inequality yields a tight coordinate-wise bound: $$\\|p-q\\|_\infty \leq \frac{1}{2}\\|p-q\\|_1 = \operatorname{TV}(p,q) \leq \sqrt{\frac{S(p,q)}{2}}.$$ Consequently, the classification margin at the explored feature distribution $q$ is lower-bounded by: $$M_y(q) \geq M_y(p) - 2\operatorname{TV}(p,q) \geq M_y(p) - \sqrt{2S(p,q)}.$$ Therefore, $M_y(p) > \sqrt{2S(p,q)}$ is a sufficient condition to preserve the top-1 prediction between the reference and explored features. **This establishes that minimizing the BFME symmetric-KL objective directly bounds the maximum possible margin reduction across explored neighborhoods, while fully accommodating statistical dependence between reference and explored features.** In the revised manuscript, we have made the following updates to Appendix A: 1. We have retained the BFME objective in Eq. (15), the standard $n$-sample Rademacher bound on original labeled data in Eq. (47), and the tightened Lemma 1, while replacing the $n+n_b$ independent-sample counting argument in Theorem 1 and Eqs. (48)–(50) with this rigorous pairwise margin preservation guarantee. 2. We have corrected Eq. (54) to distinguish symmetric KL from JSD and adjusted the trajectory upper bound in Proposition 3 accordingly. 3. We have connected Proposition 2's idealized ascent analysis ($M^*$ steps) directly to our finite implementation (Sec. 4.1: $3$ projected steps within an $\ell_\infty$ ball of radius $r=0.005$ prior to feature normalization) by writing the explicit projected update: $$\delta^{k+1}=\Pi_{\\|\delta\\|_\infty\leq r} \left(\delta^k+\eta\\,\operatorname{sign}(\nabla_\delta\ell_k)\right), \qquad \\|\delta^{k+1}\\|_2\leq\sqrt{d}\\,r,$$ where $\ell_k$ is the label-guided exploration cross-entropy, and explicitly stated that reaching a target low-confidence region requires that it intersect the $\ell_\infty$ feasible set within the finite step budget. Empirically, alongside the **+5.26-point Final gain** on CIFAR-100 (Table 2), our three-seed ImageNet-R local-head attack probe at $r=0.005$ confirms that BFME improves local boundary stability by **$0.38 \pm 0.13$ paired points** (positive in all three seeds).
AuthorsOpenReview ↗

Response to Reviewer 9jbN

Q2. Task-Agnostic Inference and Isolating Module Contributions from Expert Expansion

1. Task-agnostic prototype/feature routing (no task IDs, no raw image replay).

At test time, no task ID is ever provided: all retained experts are evaluated in parallel, and all seen classes C≤t=⋃j=0tCj\mathcal{C}_{\leq t} = \bigcup_{j=0}^t \mathcal{C}_j compete in a single unified prediction space (Sec. 4.1). Similar to prototype-based expert routing in parameter-efficient continual learning (e.g., CL-LoRA), because the ViT-B/16 backbone is frozen, we store no raw historical images. Instead, each expert jj caches lightweight pre-extracted feature embeddings and class prototypes (Appendix C.3). Let qj(c∣x)q_j(c\mid x) denote expert jj's local class probability distribution over its disjoint class set Cj\mathcal{C}_j, and let gt,j(x)g_{t,j}(x) denote its normalized expert fusion weight. The global prediction over all seen classes is:

\qquad \hat{y}_t(x)=\arg\max_{c\in\mathcal{C}_{\leq t}}p_t(c\mid x).$$ Here, $g_{t,j}(x)$ is computed by combining test feature-to-prototype compatibility and calibrated feature/output confidence, scaled by the expert's assigned class count $|\mathcal{C}_j|$. During post-session calibration, all expert LoRA and classifier parameters remain strictly frozen, and calibration uses cached features without any backbone back-propagation. In our ImageNet-R study, the cached feature budget corresponds to $10$ calibration embeddings per Base class, $5$ support embeddings plus up to $10$ selected pseudo-labeled embeddings per incremental class, and up to $300$ session-level routing embeddings per expert (which may overlap). Routing calibration uses only session membership—hidden unlabeled class labels and test labels are strictly excluded. We have added this complete formulation and feature-cache budget to Sec. 3.1 and Appendix C.3 of the revised manuscript. **2. Isolating DSGE, ADES, and BFME from the multi-expert architecture.** To verify that our performance comes from the proposed training mechanisms rather than expert expansion or routing alone, we have provided three complementary controls where the expandable LoRA architecture is held strictly identical: * **Same-architecture ablations (Table 2):** Within the exact same multi-expert architecture and router, removing DSGE, ADES, or BFME drops CIFAR-100 Final accuracy by **$3.07$, $3.49$, and $5.26$ points**, respectively. * **Router-agnostic fixed-prior evaluation:** In our three-seed ImageNet-R study, we also evaluated trained experts by completely replacing the learned fusion weights $g_{t,j}(x)$ at inference with a static, input-independent class-count prior ($g_{t,j} \propto |\mathcal{C}_j|$), which leaves all expert tensors unchanged. Under this fixed prior, `Full` still consistently outperforms `w/o DSGE` in every seed by **$+0.28 \pm 0.08$ points**, and pairing DSGE's improved expert representations with prototype/feature-calibrated fusion amplifies this gain to **$+1.24 \pm 0.72$ points** (also positive in all seeds). * **Removing unlabeled supervision (`No-U`):** Keeping the frozen ViT, LoRA experts, and $5$-shot supports fixed while disabling unlabeled data reduces ImageNet-R Final accuracy by **$1.42 \pm 0.79$ points** (see c4dy Q4). Together, these experiments confirm that the proposed evolutionary training modules substantially improve expert representations beyond the architecture itself.
AuthorsOpenReview ↗

Response to Reviewer 9jbN

Q3. Comparison with DSGD, Base Checkpoints, and Consistency Across DTE Tables

  • Comparison with DSGD & the ImageNet-100 Base offset: In Table 1 (three-seed mean ±\pm SD), DSE-SSF achieves substantial gains over DSGD on CUB-200 (+6.71+6.71 Final / +5.67+5.67 Average points) and CIFAR-100 Average (+1.67+1.67 points), while achieving comparable Final accuracy on CIFAR-100 (83.82±1.9083.82 \pm 1.90\\% vs. 83.68±0.3383.68 \pm 0.33\\%) and ImageNet-100 (86.17±1.6586.17 \pm 1.65\\% vs. 85.90±0.3185.90 \pm 0.31\\%), whereas DSGD has a higher ImageNet-100 Average (89.20±0.1289.20 \pm 0.12\\% vs. 86.67±1.0986.67 \pm 1.09\\%). Why does DSGD have a higher Average on ImageNet-100 (6060 Base + 8×58\times 5 incremental classes)? From Table 1's Final and Base-to-Final drop (Δ\Delta) columns, the implied Session-0 Base accuracy (Final+ΔFinal + \Delta) is 87.2287.22\\% for DSE-SSF vs. 93.3393.33\\% for DSGD, because each baseline uses its native Base training procedure while DSE-SSF adapts only rank-88 LoRA, LayerNorm, a 512512-d projection, and a cosine head on the frozen ViT-B/16. Crucially, this reveals that DSE-SSF suffers only a 1.051.05-point drop across the 88 incremental sessions compared to a 7.437.43-point drop for DSGD, allowing DSE-SSF to overtake DSGD in Final accuracy (86.1786.17\\% vs. 85.9085.90\\%) despite starting 6.116.11 points lower (which mechanically pulls down the 9-session Avg. mean). We have revised Sec. 4.1 and Appendix C.1.3 to state these dataset- and metric-specific findings precisely.
  • Paired statistical comparisons from identical Base checkpoints: As noted in Appendix C.1, each seed (4242, 10241024, 19931993) has its own trained Base checkpoint. In all our component and replacement studies (CIFAR-100, CUB-200, and ImageNet-R), every compared variant within a given seed loads the exact same Base checkpoint file (verified by SHA-256) and shares identical pretraining, class order, and 55-shot supports. Under this strictly paired protocol on ImageNet-R, Full outperforms w/o DSGE by +1.16±0.73+1.16 \pm 0.73 Final and +0.49±0.16+0.49 \pm 0.16 Average points (positive in every seed).
  • Reconciling DTE in Table 1 and Table 4: The difference between Table 1 and Table 4 arose because they used different backbones and corresponding Base checkpoints for different purposes: Table 1 evaluates complete methods under the main benchmark setup, whereas the motivation diagnostic in Table 4 (Appendix C.4) used DTE's ResNet-18 setup with a shared diagnostic Base checkpoint to isolate the downstream impact of static vs. oracle-corrected pseudo-labels from a common starting point. In the revised manuscript, we have unified the backbone settings for the two DTE evaluations and explicitly specified the backbone, checkpoint source, seed unit, and averaging convention (Avg. includes Session 0; Avg. Global averages incremental sessions 1…T1\dots T) in every table caption.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response — Additional Revisions

  1. Explicit Prototype/Feature Routing & Implementation Clarifications (Revised Sec. 3.1, Sec. 3.3, Sec. 3.4, Sec. 4.1 & Appendix C.1):
  • We have moved the complete task-agnostic prototype-and-feature routing formulation into Sec. 3.1 and detailed the lightweight feature-cache budget (zero raw images stored, similar to prototype-based continual LoRA routing such as CL-LoRA) in Appendix C.1.1.
  • We have clarified the post-warm-up EMA initialization (Eq. 9, Appendix C.1.2), corrected the non-recursive threshold update expression (Eq. 11), specified the pseudo-label scope and >98%>98\% upstream precision in BFME (Sec. 3.4, Table 7), and clarified the ImageNet-100 Base offset (Appendix C.1.3) and DTE diagnostic table settings (Table 9, Appendix C.5).
  1. Rigorous, Dependency-Free Theoretical Formulation for BFME (Revised Appendix A):
  • We have revised Proposition 2 (Remark 2), Lemma 1, Theorem 1, and Proposition 3 in Appendix A—replacing the n+nbn+n_b independent-sample assumption with a tight, dependency-free pairwise margin stability bound (μ(h′+ϵ,y)≥μ(h′,y)−2LBFME\mu(\mathbf{h}'+\epsilon, y) \geq \mu(\mathbf{h}', y) - \sqrt{2\mathcal{L}_{\mathrm{BFME}}}), writing out the explicit ℓ∞\ell_\infty-projected ascent updates, and distinguishing symmetric KL from Jensen-Shannon divergence (JSD≤12LBFME\mathrm{JSD} \leq \frac{1}{2}\mathcal{L}_{\mathrm{BFME}} in Eq. (52)). Below, we address each reviewer's comments in detail. Unless otherwise noted, all new results report the mean and sample standard deviation over three seeds (4242, 10241024, and 19931993), with paired differences computed from unrounded values within each seed before aggregation.

References

[1] Iscen et al. Label Propagation for Deep Semi-Supervised Learning. CVPR, 2019.

[2] Tarvainen and Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017.

[3] Miyato et al. Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning. IEEE TPAMI, 2019.

[4] Han et al. Semi-supervised few-shot class-incremental learning based on dynamic topology evolution. EAAI, 2024.

AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback and for recognizing our paper's practical motivation, strong CUB-200 performance, and systematic three-seed ablations. The central contribution of DSE-SSF is to reframe semi-supervised few-shot class-incremental learning as a self-correcting semantic evolution loop, coupling dynamic graph discovery, temporal dual-expert stabilization, and boundary-aware exploration to jointly improve pseudo-supervision reliability and representation discriminability. Rather than treating semi-supervised incremental learning as static pseudo-labeling over a fixed representation, the proposed DSE-SSF implements the self-correcting semantic evolution process through three novel components: (1) DSGE refreshes the neighborhood graph as the expert adapts to filter reliable pseudo-labels; (2) ADES leverages temporal expert disagreement to stabilize feature evolution and dynamically calibrate selection thresholds; (3) DSE-SSF introduces the BFME that regularizes decision boundaries via label-guided feature exploration.

Summary of New Experiments and Manuscript Revisions

In response to the reviewers' feedback, we have conducted extensive new three-seed controlled experiments on CUB-200 and ImageNet-R and have already incorporated all corresponding revisions into the updated manuscript:

  1. Isolating Module Gains & Direct Canonical Replacements (Revised Sec. 4.2 & Appendix C.3, Tables 3 & 6):
    • Canonical Replacements & Methodological Progress (Table 3 & Table 6): Within the exact same frozen ViT-B/16 and per-session LoRA architecture (matched by SHA-256 Base checkpoints and 1,2001{,}200 joint updates), we replaced DSGE, ADES, and BFME with LP-adapted [1], MT-adapted [2], and Feature-space VAT [3] on CUB-200 and ImageNet-R. Full surpasses LP-adapted by +0.83±0.71+0.83 \pm 0.71 Final and +0.91±0.35+0.91 \pm 0.35 Average points on ImageNet-R and preserves +0.55±0.27+0.55 \pm 0.27 points higher final fixed-Base retention on CUB-200, while consistently outperforming MT-adapted (+0.14±0.12+0.14 \pm 0.12) and Feature-space VAT (+0.22±0.08+0.22 \pm 0.08) in CUB-200 Final accuracy across all seeds. Rather than relying on a single standalone SSL trick, DSE-SSF's methodological contribution lies in unifying graph-based pseudo-label discovery, dual-expert discrepancy-calibrated thresholding, and boundary-aware feature consistency inside a single continual learning loop, where each single-component replacement tests one swap in isolation while the full loop delivers consistent seed-level gains and stronger long-horizon retention.
    • Unlabeled Data & Router Controls (Table 6): Disabling unlabeled data within the exact same architecture (No-U) reduces ImageNet-R Final accuracy by 1.42±0.791.42 \pm 0.79 points; under full 200200-class competition, DSGE boosts recognition of the 100100 incremental classes by 6.25±1.726.25 \pm 1.72 points; and replacing learned routing with a static class-count prior still yields consistent gains in every seed (+0.28±0.08+0.28 \pm 0.08). Together with Table 2 (Appendix C.2, where removing DSGE, ADES, or BFME drops CIFAR-100 Final accuracy by 3.073.07, 3.493.49, and 5.265.26 points), these results confirm substantial training gains beyond the multi-expert architecture.
  2. Global Fixed-Cohort Retention & Active-Expert Drift Analysis (Revised Sec. 4.2, Sec. 4.3 & Appendix C.3, Tables 4 & 5):
    • Fixed-Cohort Retention (Table 4): Tracking a fixed cohort of Base test samples (A0,tA_{0,t}) under global competition shows that Full's Base accuracy declines by only 0.19±0.170.19 \pm 0.17 points from Session 1 (110110 classes) to Session 10 (200200 classes) on CUB-200, reducing incremental forgetting by 0.16±0.110.16 \pm 0.11 paired points over w/o ADES.
    • Trainable Active-Expert Stability (Table 5): We clarified in Sec. 4.3, Fig. 3's caption, and Appendix C.3 that the 0% Base-path shift in Fig. 3(b) reflects frozen parameters by construction, and added direct within-session drift measurements on the trainable active expert Et\mathcal{E}_t in Table 5, demonstrating that ADES reduces feature cosine drift by 54.5%54.5\% on ImageNet-R and 57.8%57.8\% on CUB-200.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.