Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts

T0 review · 4 major / 5 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 The paper claims that prototype-based self-explainable models, previously designed for data-rich classification, can be made to work in few-shot settings by introducing a concept-guided Mixture of LoRA Experts, multi-level feature fusion…

desk verdict A plausible integration of known components whose headline claim is contradicted by its own tables; fixable, but the comparison needs to be made fair. read the letter →

arxiv 2506.04673 v1 pith:K3JA27RI submitted 2025-06-05 cs.CV

Zhong Ji , Rongshuai Wei , Jingren Liu , Yanwei Pang , Jungong Han This is my paper
classification cs.CV
keywords few-shotclassificationself-explainablemodelsprototypicalconceptlearningmixtureofexpertslow-rankadaptationparameter-efficientfine-tuningdiscriminationinterpretableimage
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

The paper is trying to establish that prototypical concept learning, the machinery behind 'this looks like that' explanations, can be made to work in few-shot classification if the adaptation of a frozen pre-trained backbone is routed through a concept-guided mixture of low-rank experts. It identifies two failure causes in low-data regimes: parametric imbalance, where full fine-tuning lets backbone updates dominate and distort concept manifolds, and representation misalignment, where pre-trained features are not tuned for part-level concept discrimination. The proposed LE-ProtoPNet addresses both with four coordinated components and reports 4.2%-8.7% relative accuracy gains over prior self-explainable models in 5-way 5-shot classification across six benchmarks. A sympathetic reader would take away that interpretability and few-shot accuracy may not be in tension, provided the concept module actively steers feature adaptation rather than being appended after the fact.

What carries the argument

The load-bearing object is the Prototypical Concept-guided Mixture of LoRA Experts (PCM). A Mixture of LoRA Experts is an ensemble of low-rank adapters, each a pair of small matrices updating a frozen weight matrix, with a gating network that selects and weights experts for each input. The paper's innovation is to realign that gating vector with the prototypical concept matrix through an attention step before routing, so each expert gradually specializes in features for a subset of concepts. Around this core sit a multi-level feature aggregation module that fuses low-, mid-, and high-level features with the final output, and a concept discrimination loss that enforces separation among concept activations; these together carry the argument that few-shot accuracy and interpretability improve jointly.

What would settle it

Re-run the five self-explainable baselines under exactly the conditions used for LE-ProtoPNet, namely the DINO backbone, the same episodic training scheme, the same 80-epoch schedule, and the same support and query construction, then compare 5-way 5-shot accuracy; if the 4.2%-8.7% relative gains shrink to noise or reverse, the headline advantage is an artifact of unequal experimental conditions.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two obstacles to using prototypical concept learning in low-data regimes, parametric imbalance and representation misalignment, can both be removed by routing a frozen pre-trained backbone's adaptation through a concept-guided mixture of low-rank experts. In LE-ProtoPNet, the prototypical concept learner scores image parts against a learnable concept matrix; that activation signal is fed into the gating of the Mixture of LoRA Experts, so each expert specializes in a subset of concepts. Features from early, middle, and late blocks are fused to preserve information that few-shot training would otherwise lose, and an InfoNCE-style concept discrimination loss pushes concepts into separated regions. The paper reports that this combination outperforms the compared self-explainable models on CUB-200-2011, mini-ImageNet, CIFAR-FS, Stanford Cars, FGVC-Aircraft, and DTD, with relative 5-way 5-shot gains of 4.2%-8.7%, while retaining prototype-based 'this looks like that' explanations.

Load-bearing premise

The reported accuracy gains assume the comparison methods were evaluated under the same backbone and training protocol; the paper states only the DINO backbone for its own model and gives no backbone or training schedule for ProtoPNet, TesNet, Eval-ProtoPNet, ST-ProtoPNet, and ProtosViT, so part of the gap may reflect setup, not the proposed modules.

Editorial extensions

If this is right

  • If the reported gains hold, prototype-based explanations no longer have to be sacrificed when only a handful of labeled images per class are available.
  • Concept-guided routing is doing real work: the ablation shows it raises 1-shot CUB accuracy from 78.53 with plain MoLE to 80.96 with PCM, suggesting expert specialization by concept is a distinct mechanism, not just added capacity.
  • The design points to a general recipe for self-explainable models: freeze the backbone, adapt with low-rank experts, and let the concept module act as the routing signal, which reduces training cost while preserving explanations.
  • On coarse-grained datasets such as mini-ImageNet, the margin over the strongest baseline is smaller or negative in some shot settings, indicating the benefit concentrates where part-level concepts matter most.

Reading between the lines

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

  • I would expect the concept-guidance benefit to concentrate on part-based tasks: if PCM were ablated on DTD or other texture-dominated classes, the gap between MoLE and PCM should shrink relative to CUB and Aircraft.
  • The interpretability claim could be quantified by comparing the learned concepts to human part annotations on CUB; the paper's visual example supports but does not measure this.
  • The LoRA-expert-per-concept design suggests a natural extension to continual or open-set few-shot learning, where new concepts could be added as new experts without retraining the backbone.
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 LE-ProtoPNet, a few-shot prototypical concept classification (FSPCC) framework that combines a Mixture of LoRA Experts (MoLE) with a prototypical concept learner. The method includes concept-guided gating, multi-level feature aggregation, and a concept discrimination loss. The authors report experiments on six benchmarks (CUB-200-2011, mini-ImageNet, CIFAR-FS, Stanford Cars, FGVC-Aircraft, DTD) and claim consistent superiority over existing self-explainable models, with 4.2%-8.7% relative gains in 5-way 5-shot classification.

Significance. The paper addresses a relevant and timely problem: making self-explainable models viable in few-shot settings. The proposed architecture is coherent and the ablation study on CUB indicates that each module contributes to accuracy, which is a useful engineering result. The interpretability visualization in Figure 5 is a strength. However, the central comparative claim is currently not supported by the paper's own tables, and the comparison protocol is insufficiently controlled. If these issues are corrected, the method could be a practical contribution to interpretable few-shot classification.

major comments (4)
  1. [IV-A, Tables II and III] The abstract's claim of '4.2%-8.7% relative gains in 5-way 5-shot classification' and of 'consistently outperforming existing SEMs' is contradicted by the paper's own results. Computing relative gains versus the best baseline per dataset from Tables II and III gives approximately: CUB +6.9%, Stanford Cars +4.0%, FGVC-Aircraft +2.2%, CIFAR-FS +8.4%, DTD +4.0%, and mini-ImageNet -0.24% (Eval-ProtoPNet 87.38 vs. LE-ProtoPNet 87.17). Thus the claimed range excludes Cars, Aircraft, and DTD, and the mini-ImageNet result is a loss, not a gain. The text in Section IV-A even acknowledges that the method 'still slightly trails behind Eval-ProtoPNet in certain settings.' This is a load-bearing inconsistency because the paper's main contribution is the claimed consistent superiority. The authors must correct the abstract and all related claims, and either report the results faithfully or provide evidence that the method indeed achieves the claimed gains.
  2. [IV (Experiment Setup)] The experimental comparison is not controlled. The manuscript states 'we adopt the DINO as the backbone' only for the proposed model, and does not specify the backbone, training protocol, or input resolution used for ProtoPNet, TesNet, Eval-ProtoPNet, ST-ProtoPNet, or ProtosViT in Tables II-III. Without this information, the reported accuracy gaps may reflect the strength of the DINO backbone rather than the proposed modules. The authors should re-run all baselines under the same backbone and training protocol, or clearly disclose each baseline's backbone and provide a controlled analysis (e.g., a variant of LE-ProtoPNet using the same DINO backbone but without the proposed modules, as partially shown in Table IV, is not a substitute for a fair external comparison).
  3. [III-D, Eq. (26)] The concept discrimination loss is described as a 'geometry-aware concept discrimination loss that enforces orthogonality among concepts,' but Eq. (26) is an InfoNCE-style softmax over the C concept activations for each sample. This loss encourages a sparse or concentrated activation pattern per sample; it does not impose any orthogonality constraint on the concept matrix P or any pairwise separation between concept vectors. The claimed mechanism is therefore not what is implemented. The authors should either replace the loss with an actual orthogonality/separability regularizer (e.g., a term encouraging PP^T to approximate the identity) or revise the description to accurately state that the loss promotes per-sample sparsity of concept activations. This is a load-bearing point for one of the four claimed contributions.
  4. [Tables II-IV] The paper reports point estimates without confidence intervals or significance tests. In particular, the mini-ImageNet 5-shot difference between LE-ProtoPNet (87.17%) and Eval-ProtoPNet (87.38%) is only 0.21 percentage points, and the claimed 'consistent outperformance' cannot be assessed without an error bar. The authors should report means and standard deviations over multiple random seeds or episodes, as is standard in few-shot classification literature, especially when making comparative claims of superiority.
minor comments (5)
  1. [Eq. (1)] The denominator in Eq. (1) uses squared L2 norms (||F||_2^2 ||P||_2^2), which is not the standard cosine similarity; Eq. (5) uses the correct cosine formula. Please check and correct the normalization.
  2. [Eq. (10)] The normalization term in the denominator of Eq. (10) is written as sum_j e_j, but the sentence immediately below states that the normalization uses e_j ⊙ g_j. These notations are inconsistent and should be aligned.
  3. [III-C, V, Table V, Table VI] There are several typos that should be fixed: 'cenarios' in Section III-C, 'metheds' in Section V, and 'Nubmer' in the headers of Tables V and VI.
  4. [III-B] The statement that 'we introduce PEFT into the SEMs for the first time' is too strong and likely inaccurate; prior prototype-based methods may have used parameter-efficient adaptations. Please soften or substantiate this claim with a more thorough survey.
  5. [IV-C] The hyperparameter analysis (number of concepts, number of experts, and lambda) is performed on mini-ImageNet or CUB only, but the chosen values are applied to all six datasets. Please report sensitivity on at least one additional dataset to demonstrate that the settings transfer.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the abstract's accuracy claim is internally inconsistent with the paper's own tables, but that is a correctness and evaluation issue, not circularity.

full rationale

This is an empirical engineering paper, and there is no derivation chain that reduces a predicted quantity to a fitted input. The classification rule in Eqs. 4-5 is standard prototype matching, and L_CD in Eq. 26 is an InfoNCE-style training loss; both are optimized on base classes and evaluated on held-out novel classes, so the reported accuracies are out-of-sample measurements rather than by-construction consequences of the objective. I found no load-bearing self-citation: the cited SEM/FSL/MoE works are external, and no uniqueness theorem or author-imported ansatz is used to force the architecture. I flag two manuscript passages that are not circularity but affect the strength of the claims. First, Section IV-A states 'it still slightly trails behind Eval-ProtoPNet [45] in certain settings,' and Table III shows Eval-ProtoPNet at 87.38% versus LE-ProtoPNet at 87.17% on mini-ImageNet 5-shot, with similar Eval-ProtoPNet wins at 10-shot and 20-shot; the abstract's '4.2%-8.7% relative gains' therefore cannot be recovered from the paper's own tables. Second, the Experiment Setup only says 'we adopt the DINO as the backbone' for the proposed model and does not state the backbone or training protocol used for the baselines, so the relative margin may reflect setup differences. These are evaluation-fairness and internal-consistency concerns, not circular reasoning. The interpretability visualizations use the model's own trained concept activation maps, but this is the standard SEM evaluation protocol rather than a definitional reduction of the accuracy claim.

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

The central claim rests on standard few-shot assumptions, a pretrained backbone, and hyperparameters selected on the evaluation benchmarks. No new physical entities or forces are introduced. The main unverified premise is that softmax separation of concept activations yields human-meaningful, orthogonal concepts.

free parameters (5)
  • number_of_concepts_C = 312
    Tuned on mini-ImageNet, Table V; best 1-shot/5-shot; applied across all datasets.
  • number_of_experts_E = 3
    Tuned on mini-ImageNet, Table VI; best across shot settings.
  • lcd_coefficient_lambda = 0.003 (1-shot), 0.001 (5-shot)
    Tuned on CUB, Fig. 4; two separate values used for the two task settings.
  • lora_rank_alpha_dropout = 8, 32, 0.1
    Chosen hyperparameters for LoRA experts; no sensitivity analysis reported.
  • temperature_tau_and_kappa = not reported
    Softmax temperature in Eq. (2) and InfoNCE temperature in Eq. (26) control activation sharpness; values omitted.
assumptions (4)
  • domain assumption DINO pretrained backbone provides transferable visual features that few-shot concept learning can build on.
    The entire method freezes the DINO backbone and adds LoRA experts; the claimed gains depend on DINO features being a strong base. Section IV Experiment Setup.
  • domain assumption Episodic few-shot evaluation with base/novel class disjointness is a valid measure of generalization.
    Standard few-shot protocol, but the paper adds concept learning without demonstrating that learned concepts transfer across the base-to-novel split. Section III-A.
  • domain assumption Softmax/InfoNCE separation over concept activations increases interpretability and orthogonality.
    The concept discrimination loss is assumed to produce disentangled, human-aligned concepts; no human study or concept purity metric is reported. Section III-D.
  • domain assumption Feature maps from low/mid/high blocks can be reshaped and fused without introducing distribution shift.
    Multi-level aggregation concatenates features of different depths; the paper does not analyze feature distribution mismatch. Section III-C.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts." pith.science (2026). https://pith.science/paper/K3JA27RI

@misc{pith2026250604673,
  author       = {Pith},
  title        = {Pith review of: Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K3JA27RI}},
  note         = {Machine review of arXiv:2506.04673}
}
read the original abstract

Self-Explainable Models (SEMs) rely on Prototypical Concept Learning (PCL) to enable their visual recognition processes more interpretable, but they often struggle in data-scarce settings where insufficient training samples lead to suboptimal performance.To address this limitation, we propose a Few-Shot Prototypical Concept Classification (FSPCC) framework that systematically mitigates two key challenges under low-data regimes: parametric imbalance and representation misalignment. Specifically, our approach leverages a Mixture of LoRA Experts (MoLE) for parameter-efficient adaptation, ensuring a balanced allocation of trainable parameters between the backbone and the PCL module.Meanwhile, cross-module concept guidance enforces tight alignment between the backbone's feature representations and the prototypical concept activation patterns.In addition, we incorporate a multi-level feature preservation strategy that fuses spatial and semantic cues across various layers, thereby enriching the learned representations and mitigating the challenges posed by limited data availability.Finally, to enhance interpretability and minimize concept overlap, we introduce a geometry-aware concept discrimination loss that enforces orthogonality among concepts, encouraging more disentangled and transparent decision boundaries.Experimental results on six popular benchmarks (CUB-200-2011, mini-ImageNet, CIFAR-FS, Stanford Cars, FGVC-Aircraft, and DTD) demonstrate that our approach consistently outperforms existing SEMs by a notable margin, with 4.2%-8.7% relative gains in 5-way 5-shot classification.These findings highlight the efficacy of coupling concept learning with few-shot adaptation to achieve both higher accuracy and clearer model interpretability, paving the way for more transparent visual recognition systems.

Figures

Figures reproduced from arXiv: 2506.04673 by the authors.

Figure 1
Figure 1. The key challenges limiting model generalization in FSPCC. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A comprehensive flowchart of the Few-Shot Prototypical Concept Classification framework with a prototypical concept learner. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of the specific module composition of our FSPCC framework,including PCM module, MFA module and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Study on the impact of coefficient λ of LCD on CUB. 3) Coefficient λ of LCD: We investigated the im￾pact of the coefficient λ of LCD on the model’s performance through experiments, with the specific results presented in [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: An example of the interpretability of our model demonstrated on CUB dataset including a support image and two query images. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 30 canonical work pages

  1. [1]

    This looks like that: deep learning for interpretable image recognition,

    C. Chen, O. Li, D. Tao, A. Barnett, C. Rudin, and J. K. Su, “This looks like that: deep learning for interpretable image recognition,”Advances in Neural Information Processing Sys- tems, vol. 32, pp. 8930–8941, 2019

    work page 2019

  2. [2]

    Learning transferable conceptual prototypes for interpretable unsupervised domain adaptation,

    J. Gao, X. Ma, and C. Xu, “Learning transferable conceptual prototypes for interpretable unsupervised domain adaptation,” IEEE Transactions on Image Processing, vol. 33, pp. 5284– 5297, 2023

    work page 2023

  3. [3]

    Describe, spot and explain: Interpretable representation learning for discriminative visual reasoning,

    C.-S. Lin and Y .-C. F. Wang, “Describe, spot and explain: Interpretable representation learning for discriminative visual reasoning,”IEEE Transactions on Image Processing, vol. 32, pp. 2481–2492, 2023

    work page 2023

  4. [4]

    An interpretable deep architecture for similarity learning built upon hierarchical concepts,

    X. Gao, T. Mu, J. Y . Goulermas, J. Thiyagalingam, and M. Wang, “An interpretable deep architecture for similarity learning built upon hierarchical concepts,”IEEE Transactions on Image Processing, vol. 29, pp. 3911–3926, 2020

    work page 2020

  5. [5]

    Explainability enhanced object detection transformer with feature disentanglement,

    W. Yu, R. Liu, D. Chen, and Q. Hu, “Explainability enhanced object detection transformer with feature disentanglement,” IEEE Transactions on Image Processing, vol. 33, pp. 6439– 6454, 2024

    work page 2024

  6. [6]

    Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai,

    A. B. Arrieta, N. D ´ıaz-Rodr´ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garc ´ıa, S. Gil-L ´opez, D. Molina, R. Benjaminset al., “Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai,”Information Fusion, vol. 58, pp. 82–115, 2020

    work page 2020

  7. [7]

    Prediction with visual evidence: Sketch classification explanation via stroke- level attributions,

    S. Liu, J. Li, H. Zhang, L. Xu, and X. Cao, “Prediction with visual evidence: Sketch classification explanation via stroke- level attributions,”IEEE Transactions on Image Processing, vol. 32, pp. 4393–4406, 2023

    work page 2023

  8. [8]

    Joint answering and explanation for visual commonsense reasoning,

    Z. Li, Y . Guo, K. Wang, Y . Wei, L. Nie, and M. Kankanhalli, “Joint answering and explanation for visual commonsense reasoning,”IEEE Transactions on Image Processing, vol. 32, pp. 3836–3846, 2023. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

    work page 2023

Show all 46 references
  1. [9]

    Interpretable image recognition by constructing transparent embedding space,

    J. Wang, H. Liu, X. Wang, and L. Jing, “Interpretable image recognition by constructing transparent embedding space,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 895–904

    2021

  2. [10]

    A grass- mann manifold handbook: Basic geometry and computational aspects,

    T. Bendokat, R. Zimmermann, and P.-A. Absil, “A grass- mann manifold handbook: Basic geometry and computational aspects,”Advances in Computational Mathematics, vol. 50, no. 1, p. 6, 2024

    2024

  3. [11]

    Pip- net: Patch-based intuitive prototypes for interpretable image classification,

    M. Nauta, J. Schl ¨otterer, M. Van Keulen, and C. Seifert, “Pip- net: Patch-based intuitive prototypes for interpretable image classification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2744– 2753

    2023

  4. [12]

    Protos- vit: Visual foundation models for sparse self-explainable clas- sifications,

    H. Turb ´e, M. Bjelogrlic, G. Mengaldo, and C. Lovis, “Protos- vit: Visual foundation models for sparse self-explainable clas- sifications,”arXiv preprint arXiv:2406.10025, 2024

    2024 arXiv

  5. [13]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning ...

    2021

  6. [14]

    Learning to prompt for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision-language models,”International Journal of Computer Vision, vol. 130, no. 9, pp. 2337–2348, 2022

    2022

  7. [15]

    Conditional prompt learning for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Conditional prompt learning for vision-language models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 16 816–16 825

    2022

  8. [16]

    Visual-language prompt tuning with knowledge-guided context optimization,

    H. Yao, R. Zhang, and C. Xu, “Visual-language prompt tuning with knowledge-guided context optimization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6757–6767

    2023

  9. [17]

    Prompt- aligned gradient for prompt tuning,

    B. Zhu, Y . Niu, Y . Han, Y . Wu, and H. Zhang, “Prompt- aligned gradient for prompt tuning,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 15 659–15 669

    2023

  10. [18]

    Clip-adapter: Better vision-language models with feature adapters,

    P. Gao, S. Geng, R. Zhang, T. Ma, R. Fang, Y . Zhang, H. Li, and Y . Qiao, “Clip-adapter: Better vision-language models with feature adapters,”International Journal of Computer Vision, vol. 132, no. 2, pp. 581–595, 2024

    2024

  11. [19]

    Tip-adapter: Training-free clip-adapter for better vision-language modeling,

    R. Zhang, R. Fang, W. Zhang, P. Gao, K. Li, J. Dai, Y . Qiao, and H. Li, “Tip-adapter: Training-free clip-adapter for better vision-language modeling,”arXiv preprint arXiv:2111.03930, 2021

    2021 arXiv

  12. [20]

    Task resid- ual for tuning vision-language models,

    T. Yu, Z. Lu, X. Jin, Z. Chen, and X. Wang, “Task resid- ual for tuning vision-language models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 10 899–10 909

    2023

  13. [21]

    Adaptive mixtures of local experts,

    R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, “Adaptive mixtures of local experts,”Neural Computation, vol. 3, no. 1, pp. 79–87, 1991

    1991

  14. [22]

    Hierarchical mixtures of experts and the em algorithm,

    M. I. Jordan and R. A. Jacobs, “Hierarchical mixtures of experts and the em algorithm,”Neural Computation, vol. 6, no. 2, pp. 181–214, 1994

    1994

  15. [23]

    Nonlinear models using dirichlet process mixtures,

    B. Shahbaba and R. Neal, “Nonlinear models using dirichlet process mixtures,”Journal of Machine Learning Research, vol. 10, no. 63, pp. 1829–1850, 2009

    2009

  16. [24]

    A parallel mixture of svms for very large scale problems,

    R. Collobert, S. Bengio, and Y . Bengio, “A parallel mixture of svms for very large scale problems,”Neural Computation, vol. 14, no. 5, pp. 1105–1114, 2002

    2002

  17. [25]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,”arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  18. [26]

    Gshard: Scaling giant models with conditional computation and automatic sharding,

    D. Lepikhin, H. Lee, Y . Xu, D. Chen, O. Firat, Y . Huang, M. Krikun, N. Shazeer, and Z. Chen, “Gshard: Scaling giant models with conditional computation and automatic sharding,” arXiv preprint arXiv:2006.16668, 2020

    2006 arXiv

  19. [27]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

    W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,”Journal of Machine Learning Research, vol. 23, no. 120, pp. 1–39, 2022

    2022

  20. [28]

    Adap- tive gating in mixture-of-experts based language models,

    J. Li, Q. Su, Y . Yang, Y . Jiang, C. Wang, and H. Xu, “Adap- tive gating in mixture-of-experts based language models,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, H. Bouamor, J. Pino, and K. Bali, Eds. Singapore: Association for ...

    2023

  21. [29]

    DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models,

    D. Dai, C. Deng, C. Zhao, R. Xu, H. Gao, D. Chen, J. Li, W. Zeng, X. Yu, Y . Wu, Z. Xie, Y . Li, P. Huang, F. Luo, C. Ruan, Z. Sui, and W. Liang, “DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models,” inProceedings of the 62nd Annual Meeti...

    2024

  22. [30]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9650–9660

    2021

  23. [31]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El- Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  24. [32]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.”International Conference on Learning Representations, vol. 1, no. 2, p. 3, 2022

    2022

  25. [33]

    Adalora: Adaptive budget allocation for parameter-efficient fine-tuning,

    Q. Zhang, M. Chen, A. Bukharin, N. Karampatziakis, P. He, Y . Cheng, W. Chen, and T. Zhao, “Adalora: Adaptive budget allocation for parameter-efficient fine-tuning,”arXiv preprint arXiv:2303.10512, 2023

    2023 arXiv

  26. [34]

    The power of scale for parameter-efficient prompt tuning,

    B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter-efficient prompt tuning,” inProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, M.-F. Moens, X. Huang, L. Specia, and S. W.- t. Yih, Eds. Online and Punta Cana, Domini...

    2021

  27. [35]

    Prefix-tuning: Optimizing continuous prompts for generation,

    X. L. Li and P. Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), C...

    2021

  28. [36]

    Sparse low-rank adaptation of pre-trained language models,

    N. Ding, X. Lv, Q. Wang, Y . Chen, B. Zhou, Z. Liu, and M. Sun, “Sparse low-rank adaptation of pre-trained language models,” inProceedings of the 2023 Conference on Empiri- cal Methods in Natural Language Processing, H. Bouamor, J. Pino, and K. Bali, Eds. Singapore: Associatio...

    2023

  29. [37]

    Enhancing few-shot out-of-distribution detection with pre- trained model features,

    J. Dong, Y . Yao, W. Jin, H. Zhou, Y . Gao, and Z. Fang, “Enhancing few-shot out-of-distribution detection with pre- trained model features,”IEEE Transactions on Image Process- ing, vol. 33, pp. 6309–6323, 2024

    2024

  30. [38]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  31. [39]

    The Caltech-UCSD Birds-200-2011 Dataset,

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The Caltech-UCSD Birds-200-2011 Dataset,” California In- stitute of Technology, Tech. Rep. CNS-TR-2011-001, 2011

    2011

  32. [40]

    Fine-grained visual classification of aircraft,

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi, “Fine-grained visual classification of aircraft,”arXiv preprint arXiv:1306.5151, 2013

    2013 arXiv

  33. [41]

    Matching networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstraet al., “Matching networks for one shot learning,”Advances in Neu- ral Information Processing Systems, vol. 29, pp. 3630–3638, 2016

    2016

  34. [42]

    Meta-learning with differentiable closed-form solvers,

    L. Bertinetto, J. F. Henriques, P. H. Torr, and A. Vedaldi, “Meta-learning with differentiable closed-form solvers,”arXiv preprint arXiv:1805.08136, 2018

    2018 arXiv

  35. [43]

    Few-shot learning with metric-agnostic conditional embeddings,

    N. Hilliard, L. Phillips, S. Howland, A. Yankov, C. D. Corley, and N. O. Hodas, “Few-shot learning with metric-agnostic conditional embeddings,”arXiv preprint arXiv:1802.04376, 2018

    2018 arXiv

  36. [44]

    Describing textures in the wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 3606–3613. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    2014

  37. [45]

    Evaluation and improvement of interpretability for self-explainable part-prototype networks,

    Q. Huang, M. Xue, W. Huang, H. Zhang, J. Song, Y . Jing, and M. Song, “Evaluation and improvement of interpretability for self-explainable part-prototype networks,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 2011–2020

    2023

  38. [46]

    Learning support and trivial prototypes for interpretable image classification,

    C. Wang, Y . Liu, Y . Chen, F. Liu, Y . Tian, D. J. McCarthy, H. Frazer, and G. Carneiro, “Learning support and trivial prototypes for interpretable image classification,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 2062–2072. Zhong J...

    2023

Pith tools

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