Pith. sign in

REVIEW 4 major objections 5 minor 27 references

MCIGLE: Multimodal Exemplar-Free Class-Incremental Graph Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MCIGLE claims that an exemplar-free, multimodal graph learner can add new classes without storing old data and still match joint training on accuracy and forgetting.

desk verdict An integration of existing modules with fresh benchmark numbers, but the central non-forgetting claim rests on an invalid RLS update and unverifiable experiments. read the letter →

arxiv 2509.06219 v1 pith:KN6ED57S submitted 2025-09-07 cs.LG cs.MM

Haochen You , Baojing Liu This is my paper
classification cs.LGcs.MM
keywords class-incrementallearningexemplar-freemultimodalgraphneuralnetworksrecursiveleastsquarescatastrophicforgettingoptimaltransportcontinual
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes MCIGLE, a method for class-incremental learning on multimodal graphs that does not store any old training samples. It claims that by aligning text and visual features and updating a linear classifier through recursive least squares, the model can add new classes while keeping accuracy on old classes high. Experiments on four public datasets report higher accuracy and lower forgetting than several baselines. The authors argue the approach approaches the performance of training on all data at once.

What carries the argument

Concatenated Recursive Least Squares (C-RLS): a recursive update rule for a linear classifier that maintains an inverse autocorrelation matrix via the Sherman-Morrison formula, adjusting weights with a gain term as new phases arrive. Because it only needs current-phase features and labels, it avoids storing old data. The residual fitting module adds a nonlinear correction stream trained on the current phase's prediction error.

What would settle it

Train MCIGLE on a sequence of classes, then retrain the same architecture on all classes jointly from scratch; if joint training clearly beats MCIGLE on old-class accuracy, the claimed joint-training equivalence fails. Also, run a variant with the feature extractor unfrozen after phase one; if forgetting spikes, the method's non-forgetting property depends on freezing.

Watch

Extended reading notes

Core claim

The central claim is that exemplar-free class-incremental graph learning can be made practical by combining four modules: a multimodal feature processor that aligns visual and textual node features via optimal transport, a Fourier-based periodic feature extractor, a non-forgetting linear head updated with Concatenated Recursive Least Squares (C-RLS), and a residual fitting network that compensates for the linear head's underfitting. The C-RLS module maintains recursive autocorrelation and cross-correlation matrices with a forgetting factor and uses the Sherman-Morrison formula to avoid storing historical data. The paper reports that MCIGLE consistently outperforms five baselines across four

Load-bearing premise

The recursive least squares update is assumed to preserve old-class knowledge even though it is only fitted to current-phase data with a forgetting factor that down-weights old information and no old labels to correct it.

Editorial extensions

If this is right

  • If the claim holds, exemplar-free CIL can match joint training in graph node classification without memory overhead.
  • Multimodal graph data (text and visual) can be handled in one incremental framework without replay.
  • The method gives a practical recipe for privacy-sensitive continual learning where old data cannot be kept.
  • The Fourier extractor reduces parameter count compared to CNNs or MLPs while capturing periodic patterns.
  • The residual stream can correct linear-head underfitting, improving performance on complex samples.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The forgetting factor beta is a key hyperparameter; its optimal value likely depends on task similarity, and sweeping it may reveal a capacity-forgetting tradeoff the paper does not analyze.
  • The C-RLS update's 'joint-training equivalence' is asserted, not derived; a derivation or counterexample would clarify whether the method is truly non-forgetting or just a well-tuned online linear head.
  • The framework could be extended to other modalities or node-level tasks like link prediction, provided the alignment step generalizes.
  • If the feature extractor is allowed to update after the first phase, the C-RLS guarantee likely breaks; freezing the backbone may be essential to the method.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MCIGLE, a framework for exemplar-free class-incremental graph learning on multimodal data. It combines a multimodal feature processing module based on optimal transport, a Fourier-based periodic feature extraction module, a 'Non-Forgetting Mainstream Module' using Concatenated Recursive Least Squares (C-RLS), and a residual fitting enhancement module. The framework is evaluated on four public datasets (COCO-QA, VoxCeleb, SNLI-VE, AudioSet-MI) against five baselines, with tables and figures reporting accuracy and forgetting metrics. The central claim is that C-RLS recursively updates weights without storing historical data and achieves performance comparable to joint training, thereby enabling exemplar-free class-incremental learning.

Significance. The problem addressed—exemplar-free class-incremental learning on multimodal graph-structured data—is timely and relevant, and the paper targets a real gap in the literature. The proposed architecture combines several plausible components, and the experimental section attempts a broad comparison. However, the significance is conditional on the validity of the C-RLS non-forgetting mechanism, which is the paper's core contribution. Since the mathematical formulation of C-RLS is internally inconsistent and the empirical results are not reproducible from the information given, the paper does not currently establish its central claim. No code, proofs, or detailed experimental setups are provided, so the reported gains cannot be independently verified.

major comments (4)
  1. [§2.3, Eqs. (8)–(10)] The C-RLS update is dimensionally invalid for the stated setting. Eq. (10) writes the gain k_k with a scalar denominator β + X_{M,k}Φ_{k-1}^{-1}X_{M,k}^T, but for a batch of N_k samples this term is an N_k×N_k matrix and the gain should involve its inverse; the given expression is only valid for N_k=1. The appeal to Sherman–Morrison is also not justified for batch updates. More importantly, the claimed non-forgetting property is not derived: since β is introduced as a forgetting factor, β<1 exponentially down-weights earlier data, and no old labels or features are available to correct the update. The assertion that the recursion 'achiev[es] performance comparable to joint training' is therefore unsupported and in tension with the equations as written. This is the load-bearing mechanism of the paper.
  2. [§2.1, Eq. (4)] The optimal transport alignment objective is not a valid OT program. In Eq. (4), the cost term (1−λ1)L(x_i, y_j, x_i′, y_j′)P_{i′j′} depends on P twice, and the summation over four index sets i,i′,j,j′ is not a standard transport cost. No marginal constraints on P are specified, the regularization term εW(P) is undefined, and no algorithm for computing P* is given. As a result, the 'multimodal feature processing' module is not concretely specified and cannot be implemented or evaluated from the text.
  3. [§3, Table 1] The empirical evaluation lacks essential reproducibility information. Table 1 reports no error bars, number of runs, or statistical significance tests. The task schedule (number of phases, class order, number of classes per phase), dataset splits, graph construction, and hyperparameters (λ1, λ2, γ, β, learning rates, number of layers) are not provided. Without this information, the claimed improvements over baselines cannot be distinguished from noise, and the ablation figures (Figures 2 and 3) do not have labeled axes or numeric values. A rigorous comparison is necessary to support the paper's central empirical claim.
  4. [§2.4, Eqs. (11)–(13)] The residual fitting module's 'phase-wise label exclusivity' is asserted without justification. Zero-padding the residual with 0_{N0:k−1×dy,k−1} presupposes a consistent label dimension across phases, but in class-incremental learning new classes are added and the label space changes. The claim that this construction 'prevents error propagation from earlier phases' is not derived, and the mechanism by which the compensation stream improves old-class retention is unclear. This weakens the completeness of the framework description.
minor comments (5)
  1. [Abstract] The acronym MCIGLE is never expanded in the text.
  2. [Figure 1] The framework figure contains garbled mathematical expressions and unlabeled components; it is not useful for understanding the architecture.
  3. [Figure 3] The captions mention 'TA, UCL, DGR' while the text describes 'knowledge distillation, parameter regularization, memory replay'; these should be aligned.
  4. [§3] The statement that 'the only exception is Acc on AudioSet-MI, where CavRL performs better, which is expected given its specialization' is not a scientific justification and should be supported by analysis or removed.
  5. [References] Some references lack venues or are clearly preprints; please ensure all citations are complete and accurate.

Circularity Check

0 steps flagged · score 2.0 of 10

No equation-level circularity; minor non-load-bearing self-citations and a naming overclaim in the C-RLS 'non-forgetting' assertion.

full rationale

I walked the paper's derivation chain and found no step that reduces by construction to its own inputs. The empirical comparison in Section 3 (Table 1) is measured against external baselines; the reported accuracy and forgetting values are not derived from fitted parameters renamed as predictions. Section 2.3 asserts that the C-RLS module achieves 'performance comparable to joint training' without deriving this from Eq. 9-10, and the equations as written are dimensionally questionable for a batch forgetting-factor RLS update (Eq. 10 uses a scalar denominator for a gain that should involve an N_k x N_k inverse). This is a mathematical correctness risk, but it is not circularity: the paper does not use the joint-training equivalence as a fitted input or as a definitional premise. The two self-citations ([9] and [23], both co-authored by H. You) are not load-bearing: they support generic claims about activation diversity and underfitting of linear models. The unusual '/uni000...' runs in the manuscript appear to be OCR artifacts and carry no assertions. The main non-circular concern is that the module is named 'Non-Forgetting' and the conclusion repeats the non-forgetting claim as a result, but this is a labeling/overclaim issue rather than an equation-level reduction. Score 2 reflects the minor non-load-bearing self-citations and the naming overclaim; no stronger circularity is demonstrated.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method's success depends on assumptions not justified in the paper: RLS preserves old tasks without exemplars, frozen Fourier features generalize to new classes, phase-wise zero-padding avoids error propagation, and four non-graph datasets can be meaningfully cast as graphs. Hyperparameters for OT alignment, ridge regularization, RLS forgetting, and stream mixing are tuned but not reported.

free parameters (4)
  • lambda_1 (OT balance) = not reported
    Eq (4) balances transport cost and node similarity; its value controls the alignment and is not reported.
  • lambda_2 (stream mixing) = not reported
    Eq (14) mixes the mainstream and compensation predictions; the value is tuned but not reported.
  • gamma (ridge regularization) = not reported
    Eq (8) regularizes the main stream weights; the value affects the closed-form solution and is not reported.
  • beta (RLS forgetting factor) = not reported
    Eq (10) includes the forgetting factor; its value directly controls how much old data is discounted, and it is not reported.
assumptions (5)
  • standard math Sherman-Morrison update preserves the closed-form ridge regression solution.
    Invoked in Section 2.3 to maintain Phi^{-1} without matrix inversion; standard but unproved in the paper.
  • ad hoc to paper Features frozen after the first phase remain discriminative for future classes.
    Section 2.2 says all parameters are frozen after training; if future classes need new features, the method cannot learn them.
  • ad hoc to paper Zero-padding with phase-wise label exclusivity prevents error propagation from earlier phases.
    Eqs (11) and (13) enforce this structure, but no derivation or analysis shows that it avoids cross-phase interference.
  • domain assumption COCO-QA, VoxCeleb, SNLI-VE, and AudioSet-MI can be meaningfully cast as multimodal graphs for node classification.
    Section 3 uses these datasets without describing graph construction; these are not standard graph benchmarks, so the assumption is nontrivial.
  • ad hoc to paper Recursive least squares with only current-phase data achieves the joint-training objective without exemplars.
    This is the central non-forgetting claim in Section 2.3; it is not derived and is questionable because beta < 1 discounts old data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MCIGLE: Multimodal Exemplar-Free Class-Incremental Graph Learning." pith.science (2026). https://pith.science/paper/KN6ED57S

@misc{pith2026250906219,
  author       = {Pith},
  title        = {Pith review of: MCIGLE: Multimodal Exemplar-Free Class-Incremental Graph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KN6ED57S}},
  note         = {Machine review of arXiv:2509.06219}
}
read the original abstract

Exemplar-free class-incremental learning enables models to learn new classes over time without storing data from old ones. As multimodal graph-structured data becomes increasingly prevalent, existing methods struggle with challenges like catastrophic forgetting, distribution bias, memory limits, and weak generalization. We propose MCIGLE, a novel framework that addresses these issues by extracting and aligning multimodal graph features and applying Concatenated Recursive Least Squares for effective knowledge retention. Through multi-channel processing, MCIGLE balances accuracy and memory preservation. Experiments on public datasets validate its effectiveness and generalizability.

Figures

Figures reproduced from arXiv: 2509.06219 by the authors.

Figure 1
Figure 1. Framework of MCIGLE. where C is cost matrix, L the node similarity, and W the regularization term. Finally, node labels are predicted by: yˆu = softmax  tanh  W h(l) u + b  , (5) with W and b being trainable parameters. 2.2 Periodic Feature Extraction Module After obtaining node-level embeddings, we flatten them into a global feature matrix and normalize the input. The data is then passed through neural layers b… view at source ↗
Figure 2
Figure 2. Changes in Prediction Accuracy of Various Models with Cumulative Class. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Ablation results on accuracy and forgetting across model variants. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 21 canonical work pages

  1. [1]

    Advances in neural information processing systems32 (2019)

    Ahn, H., Cha, S., Lee, D., Moon, T.: Uncertainty-based continual learning with adaptive regularization. Advances in neural information processing systems32 (2019)

    work page 2019

  2. [2]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Belouadah, E., Popescu, A.: Il2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583–592 (2019)

    work page 2019

  3. [3]

    In: Proceedings of the ACM Web Conference 2022

    Cai, J., Wang, X., Guan, C., Tang, Y., Xu, J., Zhong, B., Zhu, W.: Multimodal continual graph learning with neural architecture search. In: Proceedings of the ACM Web Conference 2022. pp. 1292–1300 (2022)

    work page 2022

  4. [4]

    arXiv preprint arXiv:2401.15275 (2024)

    Cai, Y., Rostami, M.: Dynamic transformer architecture for continual learning of multimodal tasks. arXiv preprint arXiv:2401.15275 (2024)

    arXiv 2024

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cao, X., Lu, H., Huang, L., Liu, X., Cheng, M.M.: Generative multi-modal models are good class incremental learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28706–28717 (2024)

    work page 2024

  6. [6]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    D’Alessandro, M., Alonso, A., Calabrés, E., Galar, M.: Multimodal parameter- efficient few-shot class incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3393–3403 (2023)

    work page 2023

  7. [7]

    arXiv preprint arXiv:2410.02675 (2024)

    Dong, Y., Li, G., Tao, Y., Jiang, X., Zhang, K., Li, J., Su, J., Zhang, J., Xu, J.: Fan: Fourier analysis networks. arXiv preprint arXiv:2410.02675 (2024)

    arXiv 2024

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gao, J., Hu, W., Lu, Y.: Recursive least-squares estimator-aided online learning for visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7386–7395 (2020)

    work page 2020

Show all 27 references
  1. [9]

    He, H., Xie, H., Shen, G., Fu, B., You, H., Sanchez Silva, V.: 4s-classifier: empower- ing conservation through semi-supervised learning for rare and endangered species (2025)

    2025

  2. [10]

    In: Proceedings of the Asian Conference on Computer Vision (2020)

    Lei, C.H., Chen, Y.H., Peng, W.H., Chiu, W.C.: Class-incremental learning with rectified feature-graph preservation. In: Proceedings of the Asian Conference on Computer Vision (2020)

    2020

  3. [11]

    arXiv preprint arXiv:2404.19756 (2024) 8 H

    Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljačić, M., Hou, T.Y., Tegmark, M.: Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756 (2024) 8 H. You and B. Liu

    2024 arXiv

  4. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(5), 5513–5533 (2022)

    Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Van De Weijer, J.: Class-incremental learning: survey and performance evaluation on image classi- fication. IEEE Transactions on Pattern Analysis and Machine Intelligence45(5), 5513–5533 (2022)

    2022

  5. [13]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Mittal, S., Galesso, S., Brox, T.: Essentials for class incremental learning. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 3513–3522 (2021)

    2021

  6. [14]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Petit, G., Popescu, A., Schindler, H., Picard, D., Delezoide, B.: Fetril: Feature translation for exemplar-free class-incremental learning. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3911–3920 (2023)

    2023

  7. [15]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Petit, G., Soumm, M., Feillet, E., Popescu, A., Delezoide, B., Picard, D., Hudelot, C.: An analysis of initial training strategies for exemplar-free class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1837–1847 (2024)

    2024

  8. [16]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Pian, W., Mo, S., Guo, Y., Tian, Y.: Audio-visual class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7799–7811 (2023)

    2023

  9. [17]

    arXiv preprint arXiv:2310.06743 (2023)

    Rußwurm, M., Klemmer, K., Rolf, E., Zbinden, R., Tuia, D.: Geographic location encoding with spherical harmonics and sinusoidal representation networks. arXiv preprint arXiv:2310.06743 (2023)

    2023 arXiv

  10. [18]

    Advances in neural information processing systems30(2017)

    Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. Advances in neural information processing systems30(2017)

    2017

  11. [19]

    In: IJCNN’99

    Silvescu, A.: Fourier neural networks. In: IJCNN’99. International Joint Conference on Neural Networks. Proceedings (Cat. No. 99CH36339). vol. 1, pp. 488–491. IEEE (1999)

    1999

  12. [20]

    Pattern Recognition140, 109561 (2023)

    Sun, W., Li, Q., Zhang, J., Wang, D., Wang, W., Geng, Y.a.: Exemplar-free class incremental learning via discriminative and comparable parallel one-class classi- fiers. Pattern Recognition140, 109561 (2023)

    2023

  13. [21]

    In: Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision

    Szatkowski, F., Pyla, M., Przewilikowski, M., Cygert, S., Twardowski, B., Trz- ciński, T.: Adapt your teacher: Improving knowledge distillation for exemplar-free continual learning. In: Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision. pp. 197...

    1977

  14. [22]

    IEEE Transactions on Circuits and Systems for Video Technology33(10), 5921–5931 (2023)

    Wang, S., Shi, W., Dong, S., Gao, X., Song, X., Gong, Y.: Semantic knowledge guided class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology33(10), 5921–5931 (2023)

    2023

  15. [23]

    Xu,S.,Ye,Y.,Li,M.,You,H.,Wang,K.,Zhang,W.:Drco:atoolkitforintelligently curbing illegal wildlife trade (2025)

    2025

  16. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Ye, Y., Xie, Y., Zhang, J., Chen, Z., Wu, Q., Xia, Y.: Continual self-supervised learning: Towards universal multi-modal medical data representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 11114–11124 (2024)

    2024

  17. [25]

    IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2024)

    Zhou,D.W.,Wang,Q.W.,Qi,Z.H.,Ye,H.J.,Zhan,D.C.,Liu,Z.:Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2024)

    2024

  18. [26]

    Knowledge- Based Systems304, 112513 (2024)

    Zhu, B., Wang, C., Xu, K., Feng, D., Zhou, Z., Zhu, X.: Learning incremental audio–visual representation for continual multimodal understanding. Knowledge- Based Systems304, 112513 (2024)

    2024

  19. [27]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhuang, H., He, R., Tong, K., Zeng, Z., Chen, C., Lin, Z.: Ds-al: A dual-stream analytic learning for exemplar-free class-incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 17237–17244 (2024)

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.