Pith. sign in

REVIEW 6 major objections 5 minor 33 references

ViRN: Variational Inference and Distribution Trilateration for Long-Tailed Continual Representation Learning

T0 review · 6 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ViRN reconstructs sparse tail-class distributions to lift long-tailed continual learning accuracy by 10.24% over state-of-the-art.

desk verdict Plausible method, misreported headline: the claimed 10.24% average gain is an aggregation artifact, and missing baselines and error bars weaken the SOTA claim. read the letter →

arxiv 2507.17368 v1 pith:42UVKZBR submitted 2025-07-23 cs.LG

Hao Dai , Chong Tang , Jagmohan Chauhan This is my paper
classification cs.LG
keywords continuallearningclass-incrementallong-taileddistributionvariationalinferenceautoencoderWassersteindistancetrilaterationpre-trainedmodels
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 sets out to establish that long-tailed class-incremental learning can be solved without storing raw data by reconstructing the distributions of sparsely sampled tail classes from the geometry of a pre-trained embedding space. Its method, ViRN, fits each class as a Gaussian through a variational autoencoder, then corrects tail classes by blending the $K$ classes nearest in 2-Wasserstein distance with inverse-distance weights. Across six long-tailed acoustic and image benchmarks, ViRN is reported to beat state-of-the-art baselines by an average of 10.24% accuracy and to approach the balanced-data upper bound. If true, this would give a rehearsal-free route to continual learning for rare acoustic events, accents, and infrequent visual categories.

What carries the argument

The load-bearing identity is the closed-form 2-Wasserstein distance between two Gaussians, $W_2^2(i,j) = \|\mu_i-\mu_j\|^2 + \operatorname{Tr}(\Sigma_i + \Sigma_j - 2(\Sigma_j^{1/2}\Sigma_i\Sigma_j^{1/2})^{1/2})$, which measures both mean displacement and covariance mismatch. ViRN uses it to find the $K$ nearest classes for each tail class (Eq. 7), turns those distances into inverse-distance weights $\omega_j$ (Eq. 8), and fuses the neighbors' Gaussians with the VAE estimate through an adaptive weight $\alpha_i$ (Eqs. 9-11). The fusion equation for the covariance adds an extra term $\alpha_i(1-\alpha_i)\sum_j \omega_j(\mu_i-\mu_j)(\mu_i-\mu_j)^\top$ to preserve directional inter-class differences. This geometric reconstruction is what claims to make tail-class distributions trustworthy from a handful of samples.

What would settle it

Replace the Wasserstein-selected neighbors with $K$ random classes of matched sample-count and rerun the same six benchmarks; if tail-class accuracy does not drop materially, the geometric retrieval is not driving the gain. A complementary check is to print, for a handful of tail classes, the actual retrieved neighbor class labels and verify that they are semantically related to the target.

Watch

Extended reading notes

Core claim

The central claim, stated the way a fair reader would state it, is that distribution-level reasoning replaces data replay in long-tailed continual learning. ViRN models each class-conditional feature distribution as a Gaussian $\mathcal{N}(\mu_c, \Sigma_c)$; a VAE produces stable mean and covariance estimates even for tail classes with very few samples, and a 'trilateration' step then uses the $K$ nearest classes by 2-Wasserstein distance to correct those estimates. The corrected distribution is a convex blend of the VAE estimate and the inverse-distance-weighted neighbor mixture, with an adaptive coefficient $\alpha_i$ set by the tail class's sample count relative to its neighbors. This yields a generative classifier that classifies by Mahalanobis distance and can also sample synthetic features, without ever revisiting old data.

Load-bearing premise

The method stands on the assumption that the pre-trained embedding space is organized so that the classes nearest to a tail class by 2-Wasserstein distance are semantically relevant, and that a VAE estimate from a handful of samples is stable enough to make the geometric fusion meaningful; the paper asserts the second part in Section 3.2 without demonstrating it.

Editorial extensions

If this is right

  • Long-tailed class-incremental learning can be handled without replay buffers or stored raw data, using only a pre-trained embedding backbone and per-class Gaussian estimates.
  • The largest reported gains appear in acoustic tasks, indicating that distributional trilateration is especially useful when tail classes are extremely sparse and semantically clustered.
  • ViRN approaches the balanced-data upper bound (524.26 vs. 538.52 overall), showing that most of the remaining gap comes from the long-tailed split rather than from forgetting.
  • Because the corrected distributions are analytic, the reconstruction can support generative uses such as sampling synthetic tail-class features, not just classification.

Reading between the lines

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

  • A dedicated single-sample-per-class version of the six benchmarks would test how far the geometric prior can stretch; the paper's $\rho=0.01$ splits already contain near-singleton tail classes, but it does not isolate exactly one sample per class.
  • The value of the Wasserstein metric could be isolated by rerunning the neighbor retrieval with cosine or Euclidean distances while keeping the fusion equations unchanged.
  • The reported 10.24% gain is against the specific baseline set in Table 1; comparisons with replay-based or prompt-tuning methods outside that set would be needed to position ViRN against the broader state of the art.
  • Because the recipe needs only a pre-trained embedding and per-class Gaussian estimates, the same distributional reconstruction should transfer to other modalities, such as medical or sensor signals, whenever the embedding space shows semantic clustering.
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

6 major / 5 minor

Summary. The paper proposes ViRN, a rehearsal-free method for long-tailed class-incremental learning. It represents each class by a Gaussian in a pre-trained embedding space, uses a variational autoencoder to estimate the class-conditional parameters, and then refines tail-class distributions by weighted interpolation from k nearest neighbors under 2-Wasserstein distance (Eqs. 6-11). Experiments are reported on four audio and two image datasets, compared with iCaRL, DGR, SLCA, and LAE, plus ablations of the VAE and redistribution components. The abstract claims a 10.24% average accuracy gain over state-of-the-art methods.

Significance. If the results were reproduced as stated, ViRN would be a useful contribution to rehearsal-free long-tailed CIL, particularly for audio, and the idea of repairing tail-class distributions using geometric neighbours in embedding space is original and worth pursuing. The paper reports multi-dataset evaluation with shared backbones and ablates each of the two proposed components. However, the headline improvement is not supported by the paper's own table: the 'Overall' column sums per-dataset accuracies, so the 61.44% in Section 4.2 is a sum of accuracy points and the 10.24% is an average absolute difference, not a relative gain. Without error bars and with small gains on image benchmarks, the state-of-the-art claim is not established. The significance is therefore conditional on a corrected and statistically supported comparison.

major comments (6)
  1. [Abstract and §4.2, Table 1] The headline metrics do not follow from Table 1. The 'Overall' column is a sum of six per-dataset accuracies, and the 61.44% reported in §4.2 is the sum 524.26 - 462.82, not a percentage. The abstract's 10.24% is this sum divided by 6, i.e., an average absolute accuracy difference against LAE, not a relative gain over state-of-the-art methods. Moreover, SLCA's missing ESC-LT entry is implicitly treated as 0 in its Overall total, so the comparison is distorted. Computed per dataset against the best baseline, ViRN's average absolute gain is (2.06 + 1.02 + 4.80 + 18.89 + 1.58 + 1.31)/6 = 4.94 points, and its average relative gain is about 6.9%, not 10.24%. These numbers should be corrected and the aggregation convention stated explicitly.
  2. [§4.2, Table 1] No standard deviations or numbers of seeds are reported for any accuracy value. The two image-benchmark gains (C100-LT +1.58, TIN-LT +1.31) are small enough to be within run-to-run variation, especially for a generative classifier on a fixed pre-trained embedding. The claim of consistent superiority across all six benchmarks therefore lacks statistical support; at minimum, multiple seeds with mean and std should be reported.
  3. [§3.2, Eqs. 3-5] The VAE component is under-specified. The text says that µc and Σc are treated as latent variables z and that their posterior is approximated with a VAE, but Eq. 5 is a generic ELBO over data x and no mapping is given from the learned q(z|x) to the class-conditional (µc, Σc) used in Eq. 4. Neither the VAE architecture nor its training procedure (inputs, outputs, loss terms beyond L_ELBO, number of epochs) is described. Without this, the core variational-inference contribution cannot be reproduced, and the claim that the VAE yields stable tail-class estimates from few samples is not demonstrated.
  4. [§3.3, Eqs. 10-11] Eq. 10 defines the refined mean as µ'_i = α_i µ_i + (1-α_i) Σ_{j∈N_i} ω_j µ_j. If Eq. 11 is intended as the covariance of the corresponding mixture, the cross term should be α_i(1-α_i)(µ_i - Σ_j ω_j µ_j)(µ_i - Σ_j ω_j µ_j)^T, since the second component has mean Σ_j ω_j µ_j. The term written in Eq. 11, α_i(1-α_i) Σ_j ω_j (µ_i - µ_j)(µ_i - µ_j)^T, is not equal to this in general. If this is a deliberate heuristic rather than a mixture-covariance formula, that should be stated and its effect on the Mahalanobis classifier validated.
  5. [§3.3, §4.1] The redistribution mechanism depends on two untested choices: the number of neighbors k and the fusion weight λ, and on the assumption that W2-nearest classes in the embedding space are semantically similar to the tail class. Section 4.1 fixes λ=0.7 and k=3 for all datasets without sensitivity analysis, and no experiment inspects whether the retrieved neighbors are actually semantically related or whether the reconstructed distribution is closer to the true tail-class distribution than the VAE estimate. A sensitivity study over λ and k, and a qualitative check of retrieved neighbors, are needed to support the sample-efficiency claim.
  6. [§4.1, Related Work] The comparison omits the most closely related methods cited in the paper itself: FeCAM (cited as an exemplar-free generative-classifier CIL method) and DiffuLT (diffusion-based long-tail recognition). Since the paper claims state-of-the-art performance in the same setting, these baselines should be included or a justification should be given for their exclusion.
minor comments (5)
  1. [§4.1, Figure 2] The baseline abbreviated DGR is cited to 'Gradient reweighting' (He & Zhu, 2024), but DGR usually denotes Deep Generative Replay; the reference list contains no Deep Generative Replay entry to support the table label. Please clarify which method was actually run and correct the citation. Also, the legend in Figure 2 spells it 'DRG'.
  2. [Eq. 1] The exponent is typeset ambiguously as ρ^{c_i/|C|-1}; this should be ρ^{c_i/(|C|-1)} or similar.
  3. [Table 2] The metrics 'Forgetting' and 'Novelty' are not defined in the text; please define them and state the task-averaging convention.
  4. [§4.2] The phrases 'surpassing prior methods by 61.44%' and 'degrading overall accuracy by 1.27%' should refer to accuracy points, not percentages, once the aggregation is corrected.
  5. [Table 1] The header has 'C100-LTTIN-LT' missing a space; minor typographical issue.

Circularity Check

1 steps flagged · score 2.0 of 10

Derivation chain is self-contained: Eqs. 6-11 define, rather than assume, the tail-class fusion, and all evaluation is benchmark-based against external baselines. Only flagged step: the headline 10.24%/61.44% gain is an arithmetic artifact of Table 1's summed 'Overall' columns — a report-level construction issue, not equation-level circularity.

  1. other [Abstract; Section 4.2 (Results); Table 1]
    "'ViRN achieves a 10.24% average accuracy gain over state-of-the-art methods.' (Abstract); 'ViRN achieves state-of-the-art performance, surpassing prior methods by 61.44%.' (Sec. 4.2); Table 1 'Overall' column: SLCA 395.16, LAE 462.82, ViRN 524.26 (SLCA's ESC-LT cell missing)."

    The two headline numbers are exact arithmetic identities of the paper's own Table 1. Each 'Overall' column is a sum of six per-dataset Top-1 accuracies (LAE 462.82; ViRN 524.26), so 524.26−462.82 = 61.44 is a summed difference of percentage points that Sec. 4.2 labels 'surpassing prior methods by 61.44%', and 10.24 = 61.44/6 is that summed difference divided by six, presented as an 'average accuracy gain'. Against the best per-dataset baseline the average absolute gain is about 4.94 points, and SLCA's missing ESC-LT cell is implicitly 0.

full rationale

The method's derivation chain is self-contained and not circular. Section 3 estimates class-conditional Gaussians from features (Eqs. 2-3), stabilizes them with a VAE ELBO (Eq. 5), computes closed-form 2-Wasserstein distances (Eq. 6), selects K nearest head classes (Eq. 7), assigns inverse-distance weights (Eq. 8), sets a sample-count fusion ratio (Eq. 9), and blends means and covariances (Eqs. 10-11). Equations 10-11 explicitly define the 'trilaterated' distribution as a weighted mixture of the class's own estimate and neighbors' estimates; the output is the definition of the procedure, not a quantity secretly contained in its inputs. Accuracy claims are empirical: they are measured on six external benchmarks against external baselines (iCaRL, DGR, SLCA, LAE) with fixed pre-trained backbones, so the per-dataset results are not forced by the derivation. No load-bearing self-citations exist: the reference list is entirely external (baselines, backbones, standard CL methods), and no prior work by these authors is invoked as a uniqueness theorem or as authority for the neighbor-fusion ansatz; the W2-neighborhood assumption is stated in the paper's own words and is empirically checkable. The flagged step is a report-level artifact: the 10.24% and 61.44% figures are exact arithmetic functions of Table 1's summed 'Overall' columns, with SLCA's missing ESC-LT cell counted implicitly as zero and a sum of percentage points mislabeled as a percentage gain; without error bars or multiple seeds, the headline margin is statistically unsupported. Two further weaknesses are correctness risks rather than circularity: hyperparameters (lambda=0.7, k=3) are reported without a validation protocol or sensitivity analysis, so selection on the test benchmarks cannot be excluded, and the VAE's stability for tail classes from a handful of samples is assumed. Both are evidentiary concerns, not self-referential derivations, so the circularity score remains low.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on distributional assumptions about the embedding space, the Gaussian class model, and the efficacy of VAE with minimal samples. Hyperparameters lambda and k are the main free choices, and their tuning is not disclosed.

free parameters (2)
  • lambda (fusion weight) = 0.7
    Fixed in Section 4.1; controls the balance between VI-estimated and reconstructed distributions (Eq. 9). No validation protocol described.
  • k (number of neighbors) = 3
    Fixed in Section 4.1; number of nearest neighbor distributions used in trilateration (Eq. 7). No sensitivity analysis or validation protocol.
assumptions (4)
  • domain assumption Semantically similar classes cluster in the PTM embedding space.
    The redistribution step relies on K-nearest neighbors by W2 distance to find classes whose distributions can be combined to reconstruct tail classes (Section 3.3, Eq. 7). If this geometric assumption fails, the method is ineffective.
  • domain assumption Class-conditional features are approximately Gaussian in the PTM latent space.
    The entire framework models each class as N(mu_c, Sigma_c) (Eq. 3) and uses closed-form 2-Wasserstein distance for Gaussians (Eq. 6).
  • domain assumption The variational autoencoder provides reliable posterior estimates of class means and covariances even with few samples.
    Section 3.2 claims VAE stabilizes tail-class distribution estimation, but the architecture and training are unspecified; the claim is assumed to hold for the method to work.
  • domain assumption Pretrained models (Wav2Vec, CLAP, DINOv2) provide transferable features for the downstream tasks.
    The method operates entirely in the pretrained feature space; if the PTM features are poor for a downstream dataset, the whole pipeline degrades.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViRN: Variational Inference and Distribution Trilateration for Long-Tailed Continual Representation Learning." pith.science (2026). https://pith.science/paper/42UVKZBR

@misc{pith2026250717368,
  author       = {Pith},
  title        = {Pith review of: ViRN: Variational Inference and Distribution Trilateration for Long-Tailed Continual Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42UVKZBR}},
  note         = {Machine review of arXiv:2507.17368}
}
read the original abstract

Continual learning (CL) with long-tailed data distributions remains a critical challenge for real-world AI systems, where models must sequentially adapt to new classes while retaining knowledge of old ones, despite severe class imbalance. Existing methods struggle to balance stability and plasticity, often collapsing under extreme sample scarcity. To address this, we propose ViRN, a novel CL framework that integrates variational inference (VI) with distributional trilateration for robust long-tailed learning. First, we model class-conditional distributions via a Variational Autoencoder to mitigate bias toward head classes. Second, we reconstruct tail-class distributions via Wasserstein distance-based neighborhood retrieval and geometric fusion, enabling sample-efficient alignment of tail-class representations. Evaluated on six long-tailed classification benchmarks, including speech (e.g., rare acoustic events, accents) and image tasks, ViRN achieves a 10.24% average accuracy gain over state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2507.17368 by the authors.

Figure 1
Figure 1. The trilateration method for reconstructing distributions. For this, we define the weight of each neighbor for dis￾tribution reconstruction, derived from the W2 distance as follows: ωj = 1/W2(i, j) Pk m∈Ni 1/W2(i, m) (8) Additionally, due to the significant difference in the number of samples between head and tail classes, we design an adap￾tive dynamic weighting factor α to control the fusion ratio between the VI e… view at source ↗
Figure 2
Figure 2. Accuracy of new class learning over sequential tasks. tailed scenarios by leveraging distributional trilateration. Additionally, ViRN narrows the gap with the balanced-data upper bound (524.26 vs. 538.52), demonstrating slight per￾formance degradation due to long-tailed data. This suggests that our distribution reconstruction mechanism effectively mitigates class imbalance, even in incremental settings. Learning of … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    wav2vec 2.0: A framework for self-supervised learning of speech representations

    Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pp.\ 12449--12460, 2020

    work page 2020

  3. [3]

    Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark

    Becker, S., Vielhaben, J., Ackermann, M., Müller, K.-R., Lapuschkin, S., and Samek, W. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute, 2023. ISSN 0016-0032. doi:https://doi.org/10.1016/j.jfranklin.2023.11.038. URL https://www.sciencedirect.com/science/article/pii/S00160...

    doi:10.1016/j.jfranklin.2023.11.038 2023

  4. [4]

    Learning imbalanced datasets with label-distribution-aware margin loss

    Cao, K., Wei, C., Gaidon, A., Arechiga, N., and Ma, T. Learning imbalanced datasets with label-distribution-aware margin loss. Curran Associates Inc., Red Hook, NY, USA, 2019

    work page 2019

  5. [5]

    Class-balanced loss based on effective number of samples

    Cui, Y., Jia, M., Lin, T.-Y., Song, Y., and Belongie, S. Class-balanced loss based on effective number of samples. In CVPR, pp.\ 9260--9269, 2019. doi:10.1109/CVPR.2019.00949

    arXiv 2019

  6. [6]

    Podnet: Pooled outputs distillation for small-tasks incremental learning

    Douillard, A., Cord, M., Ollion, C., Robert, T., and Valle, E. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pp.\ 86–102. Springer-Verlag, 2020. ISBN 978-3-030-58564-8. doi:10.1007/978-3-030-58565-5_6

    work page doi:10.1007/978-3-030-58565-5_6 2020

  7. [7]

    A unified continual learning framework with general parameter-efficient tuning

    Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., and Zhang, J. A unified continual learning framework with general parameter-efficient tuning. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 11449--11459, 2023. URL https://api.semanticscholar.org/CorpusID:257622675

    work page 2023

  8. [8]

    o., and van de Weijer, J

    Goswami, D., Liu, Y., Twardowski, B. o., and van de Weijer, J. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. In Advances in Neural Information Processing Systems, volume 36, pp.\ 6582--6595, 2023

    work page 2023

Show all 33 references
  1. [9]

    and Zhu, F

    He, J. and Zhu, F. M. Gradient reweighting: Towards imbalanced class-incremental learning. CVPR, pp.\ 16668--16677, 2024

    2024

  2. [10]

    D., Blei, D

    Hoffman, M. D., Blei, D. M., Wang, C., and Paisley, J. W. Stochastic variational inference. J. Mach. Learn. Res., 14: 0 1303--1347, 2012. URL https://api.semanticscholar.org/CorpusID:5652538

    2012

  3. [11]

    H., Lakhotia, K., Salakhutdinov, R., and Mohamed, A

    Hsu, W.-N., Bolte, B., Tsai, Y.-H. H., Lakhotia, K., Salakhutdinov, R., and Mohamed, A. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio, Speech and Lang. Proc., 29: 0 3451–3460, October 2021. ISSN 2329-9290. do...

    2021

  4. [12]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes, 2022. URL https://arxiv.org/abs/1312.6114

    2022 arXiv

  5. [13]

    C., Veness, J., Desjardins, G., Rusu, A

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N. C., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., and Hadsell, R. Overcoming catastrophic forgetting in neural networks. PNAS, 114: 0 3521 -...

    2016

  6. [14]

    Learning multiple layers of features from tiny images

    Krizhevsky, A. Learning multiple layers of features from tiny images. Technical report, 2009

    2009

  7. [15]

    and Yang, X

    Le, Y. and Yang, X. S. Tiny imagenet visual recognition challenge. 2015. URL https://api.semanticscholar.org/CorpusID:16664790

    2015

  8. [16]

    Modeling inter-class and intra-class constraints in novel class discovery

    Li, W., Fan, Z., Huo, J., and Gao, Y. Modeling inter-class and intra-class constraints in novel class discovery. In CVPR, pp.\ 3449--3458, 2023. doi:10.1109/CVPR52729.2023.00336

    2023

  9. [17]

    and Hoiem, D

    Li, Z. and Hoiem, D. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40 0 (12): 0 2935--2947, 2018. doi:10.1109/TPAMI.2017.2773081

    2018

  10. [18]

    Deep representation learning on long-tailed data: A learnable embedding augmentation perspective

    Liu, J., Sun, Y., Han, C., Dou, Z., and Li, W. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. CVPR, pp.\ 2967--2976, 2020. URL https://api.semanticscholar.org/CorpusID:211296758

    2020

  11. [19]

    Q., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.-Y

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H. Q., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.-Y. B., Li, S.-W., Misra, I., Rabbat, M. G., Sharma, V., Synnaeve, G., Xu, H., J \'e gou,...

    2023 arXiv

  12. [20]

    Piczak, K. J. ESC : Dataset for Environmental Sound Classification . In Proceedings of the 23rd Annual ACM Conference on Multimedia , pp.\ 1015--1018. ACM Press . ISBN 978-1-4503-3459-4. doi:10.1145/2733373.2806390. URL http://dl.acm.org/citation.cfm?doid=2733373.2806390

  13. [21]

    Rebuffi, S.-A., Kolesnikov, A., Sperl, G., and Lampert, C. H. icarl: Incremental classifier and representation learning. CVPR, pp.\ 5533--5542, 2016

    2016

  14. [22]

    Salamon, J., Jacoby, C., and Bello, J. P. A dataset and taxonomy for urban sound research. In 22nd ACM International Conference on Multimedia (ACM-MM'14) , pp.\ 1041--1044, Orlando, FL, USA, Nov. 2014

    2014

  15. [23]

    Diffu LT : Diffusion for long-tail recognition without external knowledge

    Shao, J., Zhu, K., Zhang, H., and Wu, J. Diffu LT : Diffusion for long-tail recognition without external knowledge. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=Kcsj9FGnKR

    2024

  16. [24]

    Long-tail learning with foundation model: heavy fine-tuning hurts

    Shi, J.-X., Wei, T., Zhou, Z., Shao, J.-J., Han, X.-Y., and Li, Y.-F. Long-tail learning with foundation model: heavy fine-tuning hurts. In ICML, 2024

    2024

  17. [25]

    Prototypical networks for few-shot learning

    Snell, J., Swersky, K., and Zemel, R. Prototypical networks for few-shot learning. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  18. [26]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, volume 37 of Proceedings of Machine Learning Research, pp.\ 2256--2265. PMLR, 07--09 Jul 2015

    2015

  19. [27]

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

    Wang, L., Zhang, X., Su, H., and Zhu, J. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46 0 (8): 0 5362--5383, 2024. doi:10.1109/TPAMI.2024.3367329

    2024

  20. [29]

    Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition

    Warden , P. Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition . ArXiv e-prints, April 2018. URL https://arxiv.org/abs/1804.03209

    2018 arXiv

  21. [30]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation

    Wu, Y., Chen, K., Zhang, T., Hui, Y., Berg-Kirkpatrick, T., and Dubnov, S. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. ICASSP, pp.\ 1--5, 2022

    2022

  22. [31]

    Yoon, J., Yang, E., Lee, J., and Hwang, S. J. Lifelong learning with dynamically expandable networks. ArXiv, abs/1708.01547, 2017. URL https://api.semanticscholar.org/CorpusID:3693512

    2017 arXiv

  23. [32]

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

    Zhang, G., Wang, L., Kang, G., Chen, L., and Wei, Y. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 19091--19101, 2023. URL https://api.semanticscholar.org/Corpu...

    2023

  24. [33]

    Continual learning with pre-trained models: A survey

    Zhou, D.-W., Sun, H.-L., Ning, J., Ye, H.-J., and Zhan, D.-C. Continual learning with pre-trained models: A survey. In IJCAI-24, pp.\ 8363--8371. International Joint Conferences on Artificial Intelligence Organization, 8 2024

    2024

  25. [34]

    Self-sustaining representation expansion for non-exemplar class-incremental learning

    Zhu, K., Zhai, W., Cao, Y., Luo, J., and Zha, Z.-J. Self-sustaining representation expansion for non-exemplar class-incremental learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 9286--9295, 2022. doi:10.1109/CVPR52688.2022.00908

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.