Pith. sign in

REVIEW 7 major objections 7 minor 75 references

Frugal Incremental Generative Modeling using Variational Autoencoders

T0 review · 7 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A replay-free VAE beats per-class generative replay, with 24x less memory.

desk verdict A useful combination of fixed-point priors and null-space projection for memory-frugal continual VAE, but the core projection is under-specified and the forgetting results are hard to verify as written. read the letter →

arxiv 2505.22408 v1 pith:7R24WGSG submitted 2025-05-28 cs.CV

Victor Enescu , Hichem Sahbi This is my paper
classification cs.CV
keywords continuallearningincrementalvariationalautoencoderscatastrophicforgettinggenerativereplaynull-spaceoptimizationfixed-pointiterationclass-incremental
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 tries to show that a single, replay-free conditional variational autoencoder can serve as the memory of a continual learner: it learns new classes from current-task data only, yet keeps generating reliable synthetic features for earlier classes. The authors identify per-class generative replay, which stores one generator per class, as the main scalability bottleneck, and replace it with one VAE whose latent space contains one well-separated gaussian per class and whose decoder updates are forced into the null space of previous tasks. If true, incremental learners would no longer need growing memory buffers or growing generator banks; the memory cost would stay nearly constant while accuracy matches or beats replay-based methods. On Cars196 the best variant reaches 90.61% final average accuracy, 1.3 points above the per-class replay baseline, using 6.39M versus 158.26M scalars of replay memory.

What carries the argument

The two load-bearing mechanisms are the fixed-point iteration of Eq. 5, which places each class's latent gaussian mean at a data centroid plus a repulsion term that pushes it away from all other means, and the null-space projector $\pi_l$ of Eqs. 7-9, built from the eigenvectors of the smallest-eigenvalue block of each decoder layer's feature covariance. The fixed-point priors make conditional generation label-reliable; the projector makes decoder updates invisible to previous tasks.

What would settle it

Track what data feeds the covariance update in Eq. 9: if replacing the claimed exact $\Sigma_l$ with one estimated from the decoder's own synthetic samples changes the reported proportion ratio $R$ or final accuracy substantially, then the method is not replay-free in the sense claimed. Conversely, if a purely incremental closed-form update of $\Sigma_l$ from current batch statistics, without stored samples or generations, reproduces the exact matrix, the claim stands.

Watch

Extended reading notes

Core claim

The paper's central claim is that catastrophic forgetting in a generative continual learner can be prevented without replay by combining two mechanisms. First, class-conditional gaussian priors in the VAE latent space are updated by a fixed-point iteration with a Kullback-Leibler repulsion term, so class means separate and labels stay reliable. Second, decoder gradients for each new task are projected onto the null space of the layer-wise covariance of previous-task features, so updates cannot disturb earlier class distributions. The paper reports that this design yields negligible forgetting across CUB200, Cars196, ImageNet-R, and CIFAR100, and that its dynamic variant (task-specific last layer and biases) sets a new state of the art on Cars196.

Load-bearing premise

The anti-forgetting guarantee rests on being able to maintain the layer-wise covariance matrix of previous-task data exactly while storing none of that data; the paper asserts exactness but does not specify how the matrix is updated, so if the update is actually done with synthetic samples the argument would be circular.

Editorial extensions

If this is right

  • On Cars196, the best variant reaches 90.61% final average accuracy versus 89.27% for the per-class generative replay baseline, while replay memory drops from 158.26M to 6.39M scalars.
  • The proportion ratio $R$, a direct indicator of forgetting under null-space optimization, stays below 0.05 on all benchmarks and near $10^{-8}$ on fine-grained datasets.
  • The static variant already outperforms an upper bound that stores all historical data on CUB200 and ImageNet-R while using less memory, indicating that generated diversity can beat raw storage.
  • Without a pretrained CLIP, the method beats a gaussian-prototype baseline by up to 3.5 accuracy points in the hardest single-class-per-task setting.
  • Per-new-class cost is 3.3k parameters versus 788k for the per-class replay method, a 238x reduction.

Reading between the lines

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

  • If the covariance matrix for previous tasks can truly be updated exactly without storing or replaying data, the same null-space projector could be applied to the decoders of other generative models, such as latent diffusion, without replay.
  • The KLD-repulsion fixed-point priors may transfer to any conditional latent-variable model as a label-noise reduction device; a direct test would replace the learned means with normally or uniformly sampled means and measure generation-label accuracy.
  • The reported memory savings assume the covariance matrix is maintained exactly without storing samples; if it is instead approximated from the model's own synthetic samples, the memory comparison would need to be recomputed.
  • The dynamic last-layer variant suggests a general recipe: project most layers into the null space of previous tasks while letting a tiny task-specific layer absorb new directions, delaying null-space saturation.
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

7 major / 7 minor

Summary. The paper proposes a replay-free continual learning method based on a single conditional VAE. The VAE is trained only on the current task's data, but its class-conditional Gaussian priors are learned through a fixed-point iteration (FPI) with a KLD repulsion term, and the decoder's gradients are projected into the null space of previous tasks' layer-wise covariance matrices. Two additional variants add task-specific decoder layers and conditional embeddings. The method is evaluated on CUB200, Cars196, ImageNet-R, and CIFAR100 with a frozen CLIP ViT-L/14 backbone and a learned prompt classifier; the authors report state-of-the-art accuracy on Cars196 (90.61 vs 89.27 for CGIL) with a much smaller generative-model memory footprint, and near-zero forgetting as measured by the null-space proportion ratio R. The paper also includes ablations, comparisons to replay and non-replay baselines, and additional experiments without CLIP (ResNet18 on CIFAR100).

Significance. If the central mechanism is correctly specified, the contribution is significant: it shows that a single conditional VAE can be trained incrementally without storing real images, with strong classification accuracy and a large reduction in memory compared to per-class generative replay. The paper is also thorough in its experimental protocol: three seeds, four benchmarks, ablations separating the FPI conditioning and null-space optimization, a dynamic-architecture variant, and an evaluation without CLIP features. These strengths make the empirical claims worth taking seriously. However, the null-space projection, which is the load-bearing anti-forgetting device, is not specified precisely enough to be implemented from the paper, and the source of the covariance matrices used for projection is left ambiguous. The significance is therefore conditional on resolving those specification issues.

major comments (7)
  1. [Section 4.2, Eqs. (7)-(9)] The projector is dimensionally inconsistent as written. For a fully connected layer with weight matrix W_l of size d_out x d_in, the gradient g_l has d_out*d_in components, while the matrix U_B_l defined in Eq. (9) is d_in x |B| if Sigma_l is the covariance of the layer's input activations. The operation pi_l(g_l) = U_B_l U_B_l^T g_l therefore cannot act on the flattened weight gradient unless d_out = 1. The correct null-space update for the weight matrix is Delta W_l = G_l P_l with P_l = U_B_l U_B_l^T (equivalently vec(Delta W_l) = (P_l tensor I_{d_out}) vec(G_l)). If the authors instead intend to apply P_l separately to each output neuron, that must be stated explicitly. As written, Eqs. (7)-(9) do not define a reproducible update rule, and the reported no-forgetting results cannot be traced to the stated mechanism.
  2. [Section 4.2 and Algorithm 1] The paper says the covariance matrix Sigma_l is 'incrementally estimated' and 'exact' (Section 4.2) but never states which data are used to form it. Since the method is described as replay-free, it is unclear how previous-task activations for Y^{1:t-1} are available at task t. If Sigma_l is computed from the model's own synthetic samples, the anti-forgetting guarantee is circular and the near-zero R values in Table 4 become uninterpretable. If Sigma_l is maintained as a running sum of outer products of real previous-task activations, that storage is a memory cost that must be reported in Tables 2 and 3, where only generative-model parameters are counted. The paper needs to state the exact data source, the exact incremental update formula, and the resulting memory overhead.
  3. [Appendix K, Eqs. (20)-(22) and Section 5.2] The derivation of the FPI update is not consistent with the implementation described in the main text. In Appendix K, the optimality condition contains a normalization by |Y^{1:t}| * (|Y^{1:t}| - 1), which is then 'absorbed' into lambda. The main text, however, fixes lambda = 900 for all experiments and all tasks. If the implementation uses Eq. (5) literally without that normalizer, the update is not a stationarity condition of Eq. (4); if it includes the normalizer, the effective repulsion strength changes as the number of classes grows, contradicting the statement that lambda is fixed. The exact update used in the experiments must be specified, and the question of whether the fixed-point iteration converges for the chosen lambda should be addressed.
  4. [Section 5.3, Tables 2 and 3] The memory comparisons are reported only in terms of generative-model parameters, but the proposed method also needs to store, at minimum, the eigenvectors or covariance matrices used for null-space projection, and the dynamic variant stores one decoder head per task. The numbers in Tables 2 and 3 labeled 'Replay memory size' do not include these items. If the covariance/eigenvector storage is negligible, that should be demonstrated by an explicit calculation; otherwise the advertised 24x memory reduction is not directly comparable to CGIL, which stores only the per-class VAEs.
  5. [Section 4.2 and Table 4] The reported proportion ratio R on CUB200 and Cars196 is on the order of 1e-8 (e.g., 3.0e-8 in Table 4 and 5.9e-8 in Table 5). These values are near the precision limit of standard floating-point eigendecompositions, so the claim of 'infinitesimal' forgetting may reflect numerical artifacts of the eigen-solver rather than a meaningful property of the learned representation. The paper should report the numerical precision used and discuss whether R values below, say, 1e-6 are distinguishable from zero in that precision.
  6. [Section 4.1.1, Eq. (5)] The FPI update uses the previous means mu_y^(tau-1) on the right-hand side, including mu_y itself in the sum over y' != y. The derivation of the KLD term in Appendix K (Eqs. (15)-(17)) sums derivatives over pairs, but the treatment of the self-term and the symmetry of the KLD for identity covariances should be made explicit; the current notation makes it easy to misread the update as including a self-repulsion term. A concrete statement of the exact sum computed in Eq. (5) would remove this ambiguity.
  7. [Section 5.3, Table 4 vs Table 5] Table 4 reports VAE-FO with classwise normalization as having R = 3.0e-8 on CUB200, while Table 5 reports the same configuration as 5.9e-8. The two tables should use the same configuration or explain the difference; otherwise the reader cannot determine which R value is the claim.
minor comments (7)
  1. [Abstract and Introduction] The paper uses 'replay-free' in the abstract and introduction, but Algorithm 1 and Section 4.1 generate a synthetic dataset from the decoder and use it to adapt the CLIP prompts. This is a form of generative replay for the classifier. The terminology should be clarified (e.g., 'replay-free for the VAE training') to avoid a contradiction with the later description.
  2. [Section 4.1.2, Eq. (6)] The notation D_KL(N(mu_phi,x,y, sigma_phi,x,y) || N(mu_y, I)) is used without explicitly stating that mu_y is the FPI-computed mean. Since the covariance of the prior is identity, the KLD term has a closed form; stating it would help reproducibility.
  3. [Section 4.3] The sentence 'the non-incremental training of the encoder q_phi_t enables a more optimal minimization of the prior matching loss' is vague; it is not clear why training the encoder on all tasks without the null-space constraint is 'more optimal' or how that claim is verified.
  4. [Section 5.2] The learning rate is reported as '5.0-4' and '5.0-5', which should be written as 5e-4 and 5e-5 to avoid ambiguity.
  5. [Appendix E, Eq. (10)] The definition of R uses Lambda_B and Lambda, but the text says 'the eigenvalues Lambda_B associated to U_B_l, and the eigenvalues Lambda associated to U_l.' The set of eigenvalues for U_B_l is not the full set of eigenvalues of Sigma_l, so the denominator in Eq. (10) should be clarified as the sum of all eigenvalues of Sigma_l, not just those associated with U_A_l or U_B_l.
  6. [Appendix K, Eq. (15)] The derivative expression is missing a denominator factor for the second equality; the text proceeds with a proportionality relation, but the exact constant factor (2 or 4) is not tracked. Since Eq. (5) in the main text uses a specific form, the derivation should state the exact constant that remains after the proportionality step.
  7. [References] Some references are incomplete (e.g., [53] lacks publication venue and year, [62] gives 'arxiv 2022' inconsistently, and [55] is an arXiv preprint that is later published). These should be standardized.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the FPI update is derived in-appendix from the paper's own loss, the null-space projector is an external construction, and the headline accuracy numbers are evaluated on real benchmark data; the only self-citation is non-load-bearing.

full rationale

The paper's central derivation chain is not circular. The fixed-point iteration in Eq. 5 is derived from the paper's own loss Eq. 4 in Appendix K, including the likelihood and KLD optimality steps, so it does not rely on an unstated external result; the same authors' prior work [63] is cited as the source of the formulation, but the appendix re-derives it, making that self-citation non-load-bearing. The null-space projector (Eqs. 7-9) is adapted from the external, independently published method [27], and the orthogonality idea is not claimed as a new first-principles theorem. The reported state-of-the-art claim on Cars196 is an external benchmark comparison: CLIP prompts are adapted on synthetic features but final accuracy is measured on real test data, so the headline result is not fitted into existence. The main reviewer concern is the underspecified covariance matrix Sigma_l in Section 4.2: the paper asserts it is 'incrementally estimated' and 'exact' without giving the update, and the projection in Eq. 8 is dimensionally inconsistent for flattened layer gradients. These are reproducibility and correctness defects, and the forgetting ratio R would indeed become self-referential if Sigma_l were computed from the model's own synthetic samples; however, the manuscript nowhere states that Sigma_l is computed from generated data, so that circularity cannot be exhibited and should not be assumed. Overall, the derivation chain contains no demonstrated step where a prediction reduces by construction to its inputs, so the appropriate finding is a minor, non-load-bearing self-citation rather than circularity.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim depends on three hand-set scalars (lambda, a, sample count) and on several unstated assumptions about the feature space and the covariance computation. The most fragile is the assumption that Sigma_l can be incrementally estimated exactly without replay; no procedure is given.

free parameters (3)
  • lambda (KLD repulsion weight) = 900
    Controls the repulsion between Gaussian means in the fixed-point iteration, Eq. 5. Set to 900 for all experiments with no ablation or sensitivity analysis.
  • a (eigenvalue threshold multiplier) = 100
    Selects the eigenvectors retained for null-space projection (eigenvalues smaller than a * lambda_min). Set to 100 for all experiments.
  • Number of synthetic samples per class = 500
    Used for CLIP adaptation after training the decoder; mentioned in Table 9 note. Not analyzed for sensitivity.
assumptions (6)
  • standard math VAE ELBO decomposition (Eq. 1) and reparameterization (Eq. 2)
    Background from Kingma and Welling, used throughout.
  • domain assumption CLIP image and text encoders are frozen and provide a fixed, informative feature space
    The method operates on 768-dim CLIP features, and the whole pipeline assumes these features are sufficient for classification and generative modeling.
  • domain assumption Class-conditional features are approximately Gaussian in the VAE latent space
    The prior p(z|y) is set to N(mu_y, I), and the encoder outputs Gaussian parameters. The FPI separation assumes this is a good model.
  • domain assumption The null-space projection preserves previous task functions when Sigma_l is accurate
    Adopted from Wang et al. [27]; the paper assumes the covariance of previous task activations can be computed exactly (Section 4.2) and that projecting gradients into the null space prevents interference.
  • ad hoc to paper Fixed-point iteration in Eq. 5 converges to a useful configuration of means
    Convergence is not proven; the paper states it is run once (till convergence) and observes separability in Figure 3, but no guarantee is given.
  • ad hoc to paper Previous task Gaussian means remain fixed and are sufficient statistics of previous classes
    The method stores only class means for previous tasks; the paper assumes these means plus the null-space-protected decoder are enough to generate useful synthetic features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Frugal Incremental Generative Modeling using Variational Autoencoders." pith.science (2026). https://pith.science/paper/7R24WGSG

@misc{pith2026250522408,
  author       = {Pith},
  title        = {Pith review of: Frugal Incremental Generative Modeling using Variational Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7R24WGSG}},
  note         = {Machine review of arXiv:2505.22408}
}
read the original abstract

Continual or incremental learning holds tremendous potential in deep learning with different challenges including catastrophic forgetting. The advent of powerful foundation and generative models has propelled this paradigm even further, making it one of the most viable solution to train these models. However, one of the persisting issues lies in the increasing volume of data particularly with replay-based methods. This growth introduces challenges with scalability since continuously expanding data becomes increasingly demanding as the number of tasks grows. In this paper, we attenuate this issue by devising a novel replay-free incremental learning model based on Variational Autoencoders (VAEs). The main contribution of this work includes (i) a novel incremental generative modelling, built upon a well designed multi-modal latent space, and also (ii) an orthogonality criterion that mitigates catastrophic forgetting of the learned VAEs. The proposed method considers two variants of these VAEs: static and dynamic with no (or at most a controlled) growth in the number of parameters. Extensive experiments show that our method is (at least) an order of magnitude more ``memory-frugal'' compared to the closely related works while achieving SOTA accuracy scores.

Figures

Figures reproduced from arXiv: 2505.22408 by the authors.

Figure 1
Figure 1. Visualization of the incremental learning of gaussians using the fixed-point iteration. The [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. This figure shows the proposed architecture for incremental learning based on conditional [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. TSNE projection of classwise noise sampled from gaussians from task 1 to task 6 on [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Accuracy and standard deviation on CUB200 as tasks evolve for VAE-FO. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Accuracy and standard deviation on Cars196 as tasks evolve for VAE-FO. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Accuracy and standard deviation on ImageNet-R as tasks evolve for VAE-FO. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Accuracy and standard deviation on CIFAR100 as tasks evolve for VAE-FO. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 43 canonical work pages

  1. [1]

    Catastrophic interference in connectionist networks: The sequential learning problem

    Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, volume 24, pages 109–165. Elsevier, 1989

    1989

  2. [2]

    Lifelong robot learning

    Sebastian Thrun and Tom M Mitchell. Lifelong robot learning. Robotics and autonomous systems , 15(1-2):25–46, 1995

    work page 1995

  3. [3]

    Child: A first step towards continual learning

    Mark B Ring. Child: A first step towards continual learning. Machine Learning, 28(1):77–104, 1997

    work page 1997

  4. [4]

    Continual lifelong learning with neural networks: A review

    German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54–71, 2019

    2019

  5. [5]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017

    2017

  6. [6]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021

    2021

  7. [7]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR, 2021

    2021

  8. [8]

    Scaling up visual and vision-language representation learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904–4916. PMLR, 2021

    2021

Show all 75 references
  1. [9]

    Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l on imagenet

    Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Shuyang Gu, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l on imagenet. arXiv preprint arXiv:2212.06138, 2022

    2022 arXiv

  2. [10]

    Improving clip fine-tuning performance

    Yixuan Wei, Han Hu, Zhenda Xie, Ze Liu, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Improving clip fine-tuning performance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5439–5449, 2023

    2023

  3. [11]

    Preventing zero-shot transfer degradation in continual learning of vision-language models

    Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xiangyu Yue, and Yang You. Preventing zero-shot transfer degradation in continual learning of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19125–19136, 2023

    2023

  4. [12]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022

    2022

  5. [13]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Ges- mundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790–2799. PMLR, 2019

    2019

  6. [14]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337–2348, 2022

    2022

  7. [15]

    Conditional prompt learning for vision- language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision- language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816–16825, 2022

    2022

  8. [16]

    Attriclip: A non-incremental learner for incremental knowledge learning

    Runqi Wang, Xiaoyue Duan, Guoliang Kang, Jianzhuang Liu, Shaohui Lin, Songcen Xu, Jinhu Lü, and Baochang Zhang. Attriclip: A non-incremental learner for incremental knowledge learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages ...

    2023

  9. [17]

    Clip with generative latent replay: a strong baseline for incremental learning

    Emanuele Frascaroli, Aniello Panariello, Pietro Buzzega, Lorenzo Bonicelli, Angelo Porrello, and Simone Calderara. Clip with generative latent replay: a strong baseline for incremental learning. In 35th British Machine Vision Conference, 2024

    2024

  10. [18]

    Semantic residual prompts for continual learning

    Martin Menabue, Emanuele Frascaroli, Matteo Boschini, Enver Sangineto, Lorenzo Bonicelli, Angelo Porrello, and Simone Calderara. Semantic residual prompts for continual learning. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2024

    2024

  11. [19]

    Fetril: Feature translation for exemplar-free class-incremental learning

    Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3911–3920, 2023. 10

    2023

  12. [20]

    Class-incremental learning via dual augmentation

    Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306–14318, 2021

    2021

  13. [21]

    Prototype augmentation and self- supervision for incremental learning

    Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self- supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871–5880, 2021

    2021

  14. [22]

    Diffusion model meets non-exemplar class- incremental learning and beyond

    Jichuan Zhang, Yali Li, Xin Liu, and Shengjin Wang. Diffusion model meets non-exemplar class- incremental learning and beyond. arXiv preprint arXiv:2408.02983, 2024

    2024 arXiv

  15. [23]

    Generative feature replay for class-incremental learning

    Xialei Liu, Chenshen Wu, Mikel Menta, Luis Herranz, Bogdan Raducanu, Andrew D Bagdanov, Shangling Jui, and Joost van de Weijer. Generative feature replay for class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Worksh...

    2020

  16. [24]

    Brain-inspired replay for continual learning with artificial neural networks

    Gido M Van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):4069, 2020

    2020

  17. [25]

    Task-agnostic continual learning with hybrid probabilistic models

    Polina Kirichenko, Mehrdad Farajtabar, Dushyant Rao, Balaji Lakshminarayanan, Nir Levine, Ang Li, Huiyi Hu, Andrew Gordon Wilson, and Razvan Pascanu. Task-agnostic continual learning with hybrid probabilistic models. arXiv preprint arXiv:2106.12772, 2021

    2021 arXiv

  18. [26]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  19. [27]

    Training networks in null space of feature covariance for continual learning

    Shipeng Wang, Xiaorong Li, Jian Sun, and Zongben Xu. Training networks in null space of feature covariance for continual learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 184–193, 2021

    2021

  20. [28]

    Packnet: Adding multiple tasks to a single network by iterative pruning

    Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages 7765–7773, 2018

    2018

  21. [29]

    Overcoming catastrophic forgetting with hard attention to the task

    Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International conference on machine learning, pages 4548–4557. PMLR, 2018

    2018

  22. [30]

    Lifelong learning with dynamically expandable networks

    Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017

    2017 arXiv

  23. [31]

    Progressive neural networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Ko- ray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016

    2016 arXiv

  24. [32]

    Continual learning through synaptic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987–3995. PMLR, 2017

    2017

  25. [33]

    Overcoming catastrophic forgetting in neural networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of scien...

    2017

  26. [34]

    Memory aware synapses: Learning what (not) to forget

    Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139–154, 2018

    2018

  27. [35]

    Overcoming catas- trophic forgetting by incremental moment matching

    Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catas- trophic forgetting by incremental moment matching. Advances in neural information processing systems, 30, 2017

    2017

  28. [36]

    Continual learning of context-dependent processing in neural networks

    Guanxiong Zeng, Yang Chen, Bo Cui, and Shan Yu. Continual learning of context-dependent processing in neural networks. Nature Machine Intelligence, 1(8):364–372, 2019

    2019

  29. [37]

    Gradient projection memory for continual learning

    Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762, 2021

    2021 arXiv

  30. [38]

    Ffnb: Forgetting-free neural blocks for deep continual visual learning

    Hichem Sahbi and Haoming Zhan. Ffnb: Forgetting-free neural blocks for deep continual visual learning. arXiv preprint arXiv:2111.11366, 2021

    2021 arXiv

  31. [39]

    Gradient based sample selection for online continual learning

    Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 11

    2019

  32. [40]

    Selective experience replay for lifelong learning

    David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018

    2018

  33. [41]

    Riemannian walk for incremental learning: Understanding forgetting and intransigence

    Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532–547, 2018

    2018

  34. [42]

    Using hindsight to anchor past knowledge in continual learning

    Arslan Chaudhry, Albert Gordo, Puneet Dokania, Philip Torr, and David Lopez-Paz. Using hindsight to anchor past knowledge in continual learning. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 6993–7001, 2021

    2021

  35. [43]

    Lifelong gan: Continual learning for conditional image generation

    Mengyao Zhai, Lei Chen, Frederick Tung, Jiawei He, Megha Nawhal, and Greg Mori. Lifelong gan: Continual learning for conditional image generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2759–2768, 2019

    2019

  36. [44]

    Variational continual learning

    Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. Variational continual learning. arXiv preprint arXiv:1710.10628, 2017

    2017 arXiv

  37. [45]

    Continual learning with deep generative replay

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017

    2017

  38. [46]

    Generative adversarial networks.Communications of the ACM, 63(11):139– 144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139– 144, 2020

    2020

  39. [47]

    Generative replay with feedback connections as a general strategy for continual learning

    GM Van de Ven and AS Tolias. Generative replay with feedback connections as a general strategy for continual learning. arxiv. arXiv preprint arXiv:1809.10635, 2018

    2018 arXiv

  40. [48]

    Normalizing flows: An introduction and review of current methods

    Ivan Kobyzev, Simon JD Prince, and Marcus A Brubaker. Normalizing flows: An introduction and review of current methods. IEEE transactions on pattern analysis and machine intelligence, 43(11):3964–3979, 2020

    2020

  41. [49]

    Class-prototype conditional diffusion model for continual learning with generative replay

    Khanh Doan, Quyen Tran, Tuan Nguyen, Dinh Phung, and Trung Le. Class-prototype conditional diffusion model for continual learning with generative replay. arXiv preprint arXiv:2312.06710, 2023

    2023 arXiv

  42. [50]

    Ddgr: Continual learning with deep diffusion-based generative replay

    Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. In International Conference on Machine Learning, pages 10744–10763. PMLR, 2023

    2023

  43. [51]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256–2265. PMLR, 2015

    2015

  44. [52]

    Unrolled generative adversarial networks

    Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial networks. arXiv preprint arXiv:1611.02163, 2016

    2016 arXiv

  45. [53]

    Normalizing Flows for Probabilistic Modeling and Inference

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshmi- narayanan. Normalizing Flows for Probabilistic Modeling and Inference

  46. [54]

    One-for-more: Continual diffusion model for anomaly detection

    Xiaofan Li, Xin Tan, Zhuo Chen, Zhizhong Zhang, Ruixin Zhang, Rizen Guo, Guanna Jiang, Yulong Chen, Yanyun Qu, Lizhuang Ma, et al. One-for-more: Continual diffusion model for anomaly detection. arXiv preprint arXiv:2502.19848, 2025

    2025 arXiv

  47. [55]

    Incremental learning of structured memory via closed-loop transcription

    Shengbang Tong, Xili Dai, Ziyang Wu, Mingyang Li, Brent Yi, and Yi Ma. Incremental learning of structured memory via closed-loop transcription. arXiv preprint arXiv:2202.05411, 2022

    2022 arXiv

  48. [56]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631–648. Springer, 2022

    2022

  49. [57]

    Coda-prompt: Continual decomposed attention- based prompting for rehearsal-free continual learning

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention- based prompting for rehearsal-free continual learning. In Proceedings of the IE...

    2023

  50. [58]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision , pages 709–727. Springer, 2022. 12

    2022

  51. [59]

    Visual prompt tuning in null space for continual learning

    Yue Lu, Shizhou Zhang, De Cheng, Yinghui Xing, Nannan Wang, Peng Wang, and Yanning Zhang. Visual prompt tuning in null space for continual learning. arXiv preprint arXiv:2406.05658, 2024

    2024 arXiv

  52. [60]

    Prompt gradient projection for continual learning

    Jingyang Qiao, Xin Tan, Chengwei Chen, Yanyun Qu, Yong Peng, Yuan Xie, et al. Prompt gradient projection for continual learning. In The Twelfth International Conference on Learning Representations, 2023

    2023

  53. [61]

    Consistent prompting for rehearsal-free continual learning

    Zhanxin Gao, Jun Cen, and Xiaobin Chang. Consistent prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28463– 28473, 2024

    2024

  54. [62]

    Understanding diffusion models: A unified perspective

    Calvin Luo. Understanding diffusion models: A unified perspective. arxiv 2022. arXiv preprint arXiv:2208.11970

    2022 arXiv

  55. [63]

    Learning conditionally untangled latent spaces using fixed point iteration

    Victor Enescu and Hichem Sahbi. Learning conditionally untangled latent spaces using fixed point iteration. In 35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25-28, 2024 . BMV A, 2024

    2024

  56. [64]

    Carl D. Meyer. Matrix analysis and applied linear algebra. Society for Industrial and Applied Mathematics, USA, 2000

    2000

  57. [65]

    juill 2011

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie.The Caltech-UCSD Birds-200-2011 Dataset. juill 2011

    2011

  58. [66]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013

    2013

  59. [67]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international c...

    2021

  60. [68]

    Learning multiple layers of features from tiny images.(2009), 2009

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.(2009), 2009

    2009

  61. [69]

    Slca: Slow learner with classifier alignment for continual learning on a pre-trained model

    Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19148–19158, 2023

    2023

  62. [70]

    Slca++: Unleash the power of sequential fine-tuning for continual learning with pre-training

    Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca++: Unleash the power of sequential fine-tuning for continual learning with pre-training. arXiv preprint arXiv:2408.08295, 2024

    2024 arXiv

  63. [71]

    Boosting continual learning of vision-language models via mixture-of-experts adapters

    Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219–23230, 2024

    2024

  64. [72]

    A comprehensive survey of continual learning: theory, method and application

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  65. [73]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  66. [74]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017

    2001

  67. [75]

    proportion

    Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, and De-Chuan Zhan. Pycil: a python toolbox for class- incremental learning. SCIENCE CHINA Information Sciences, 66(9):197101, 2023. 13 The numbering of sections/lines/references in this Appendix follows the submitted paper. A Content of th...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.