Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Novel categories can be discovered by clustering in a human-readable concept space rather than opaque features, with competitive accuracy and built-in explanations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 17:31 UTC pith:ITLA762P

load-bearing objection Solid methods paper that puts NCD inside a CLIP-aligned concept bottleneck and actually ships readable cluster signatures; the “strict hypothesis restriction” claim is oversold by the paper’s own Remark 4.2. the 3 major comments →

arxiv 2607.04548 v1 pith:ITLA762P submitted 2026-07-05 cs.CV cs.AI

Explainable Novel Category Discovery in Semantic Concept Space

Ifrat Ikhtear Uddin , Yang Zhou , KC Santosh , Longwei Wang This is my paper
classification cs.CV cs.AI
keywords novel category discoveryconcept bottleneckexplainable AIvision-language alignmentself-labelingopen-world learningsemantic concept space
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most methods that find new categories in unlabeled images work in high-dimensional feature spaces that give no clue why a group of samples forms a class. This paper argues that discovery should happen inside a space whose axes are human-understandable visual attributes such as "four-legged" or "metallic." It learns those attributes without concept labels by aligning a projection layer with vision-language similarity scores, then runs the usual labeled-plus-unlabeled self-labeling objective entirely on concept activations. The result is that every discovered cluster comes with a stable concept signature and every instance with positive and negative concept evidence. A short theoretical argument shows that forcing decisions through this bottleneck strictly shrinks the set of allowable decision rules, ruling out many partitions that rely only on arbitrary feature correlations. On standard benchmarks the method stays close to strong non-interpretable baselines while uniquely supplying readable explanations.

Core claim

Routing both representation learning and pseudo-label assignment through a label-free semantic concept bottleneck yields novel-category partitions that are competitive in accuracy with feature-space methods and that are explainable by construction via stable concept signatures and instance-level concept evidence.

What carries the argument

The concept bottleneck: a projection that maps encoder features onto K human-readable concept activations (learned by alignment with vision-language similarities), after which all classification heads and Sinkhorn-based pseudo-labels operate only on those activations.

Load-bearing premise

The method assumes that vision-language similarity scores remain faithful concept supervisors for both known and never-before-seen categories; if those scores miss or mis-rank the attributes that actually define novel classes, both the explanations and the claimed inductive bias fail.

What would settle it

On a domain where the vision-language model systematically lacks the relevant attributes (for example fine-grained medical or industrial imagery), measure whether concept-space discovery still recovers coherent novel clusters and whether the resulting concept signatures match expert judgment; systematic failure on either metric would refute the claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes xNCD, a novel category discovery method that routes both representation learning and pseudo-label assignment through a label-free semantic concept bottleneck. Stage 1 pretrains an encoder on labeled data; Stage 2 learns a concept projection by aligning activations with CLIP image-text similarities via cubed cosine similarity and filtering; Stage 3 runs a UNO-style multi-view self-labeling objective entirely in concept space. The authors claim that this induces a strict restriction of the feature-space hypothesis class (Prop. 4.1), yields competitive task-aware and task-agnostic accuracy on CIFAR-10/100 and CUB-200, and supplies intrinsic cluster- and instance-level concept explanations that prior NCD methods lack.

Significance. If the central claims hold, the work is a useful contribution to open-world vision: it is the first systematic attempt to make NCD intrinsically interpretable by construction rather than post hoc, and it shows that competitive discovery accuracy can be retained under a concept bottleneck. Strengths include a clear three-stage pipeline, algorithms and hyperparameters sufficient for reproduction, means ± std over three runs, both task-aware and task-agnostic protocols, ablations on concept vocabulary size, and qualitative concept signatures. The theoretical framing is secondary; the practical value lies in the empirical demonstration that concept-space discovery can match or exceed strong baselines such as UNO while remaining human-readable.

major comments (3)
  1. Abstract and Prop. 4.1 claim that a concept bottleneck induces a strict restriction of H_feat via rank(W') ≤ K. Remark 4.2 immediately states that in all experimental settings K > C (CIFAR-10: 131>10; CIFAR-100: 422>100; CUB: 453>200), so the rank argument supplies no restriction and CLIP alignment is the sole operative guarantee. The abstract and introduction should be revised so that the advertised theoretical contribution matches the operative mechanism actually used in the experiments.
  2. Stage 2 (Eqs. 2–5, δ_align = 0.35) trains the concept map on all images using CLIP priors, yet no quantitative fidelity check of concept activations or CLIP similarity is reported on the unlabeled novel subset Du alone. Because both the claimed inductive bias and the intrinsic explanations rest on the assumption that CLIP remains a faithful supervisor for categories never seen as labeled classes, a held-out concept-alignment or concept-accuracy evaluation on Du is needed to support the central claim.
  3. Table 2 comparisons mix methods that differ in backbone and pretraining (ResNet-18/50 on the labeled split for xNCD versus DINO-pretrained ViT for GCD and related work). The paper notes this distinction but still presents overall accuracy as directly comparable. Either re-run key baselines under a matched backbone/pretraining protocol or clearly separate the comparison into architecture-matched and literature-reported columns so that the accuracy–interpretability trade-off can be assessed fairly.
minor comments (4)
  1. Figure 1 caption and §3.4 describe multiple unlabeled heads gn, but the main text does not state how the final head is selected at test time beyond “lowest training loss”; a one-sentence clarification would help.
  2. Appendix E defines Signature(i) with r = 10 while the main text (§3.4) uses r = 5; reconcile the two values.
  3. Table 1 reports strong novel-class accuracy on CIFAR-100 (84.10 % task-aware) that exceeds labeled accuracy; a short discussion of why the concept bottleneck appears more helpful for novel than known classes would strengthen the narrative.
  4. Typographical inconsistencies appear (e.g., “proposexNCD”, missing spaces after periods in several places); a careful proof-reading pass is needed.

Circularity Check

1 steps flagged

No load-bearing circularity: Prop. 4.1 is ordinary linear-algebra inclusion (with explicit experimental caveat that rank restriction is inactive); CLIP alignment and discovery objective are external, not self-defined.

specific steps
  1. other [Abstract + Prop. 4.1 / Remark 4.2]
    "Theoretically, we show that routing discovery through a semantic concept bottleneck induces a strict restriction of the feature-space hypothesis class... When K <min{C, d z}, the rank constraint eliminates all W' ... In our experimental settings K > C for all datasets ... so rank(W')≤C < K holds trivially and the rank argument provides no additional restriction. CLIP alignment then constitutes the sole operative interpretability guarantee"

    The abstract advertises a “strict restriction,” yet the paper itself immediately records that the rank condition fails in every reported experiment, leaving only the external CLIP prior. This is over-claim relative to the stated assumptions rather than a self-definitional loop or fitted-input-as-prediction; the mathematical inclusion itself remains non-circular. Flagged only as a minor presentational inconsistency that does not force the empirical results.

full rationale

The paper's derivation chain does not reduce any claimed prediction or first-principles result to its own inputs by construction. Proposition 4.1 establishes the standard containment H_concept \subseteq H_feat (and strictness under K < min{C, d_z}) via the factorization W' = A W_c; this is ordinary rank-linear-algebra, not a fitted quantity renamed as a theorem. Remark 4.2 candidly notes that the experimental regimes have K > C, so the rank argument supplies no restriction and CLIP alignment is the operative constraint—this is a scope caveat, not circularity. Concept learning maximizes cubed cosine similarity to an external CLIP image-text matrix (Eqs. 2–5) and filters at a fixed threshold; the discovery loss is the standard multi-view Sinkhorn self-labeling objective of UNO/SwAV applied to concept logits. Explanations are “by construction” only in the design sense that predictions are forced through the concept coordinates the authors chose to learn; that is intentional architecture, not a circular derivation. Minor self-citations to the authors’ prior explainability papers exist but are not invoked as uniqueness theorems or load-bearing premises for the NCD claims. Empirical accuracy numbers are measured against external baselines on public splits and do not recycle fitted parameters as “predictions.” Hence the circularity score is near zero.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on a small set of free thresholds, the standard NCD assumption that the number of novel classes is known, the domain assumption that CLIP supplies adequate concept priors, and the architectural invention of the concept-space discovery heads. No new physical entities are postulated; the main invented objects are the concept projection and the resulting signatures.

free parameters (4)
  • δ_align (concept retention threshold) = 0.35
    Hand-chosen cutoff (0.35) that discards poorly aligned concepts after Stage 2; directly controls the final concept vocabulary size and therefore the hypothesis-class restriction and explanation quality.
  • r (top-r concepts for signatures) = 5 (or 10)
    Number of activated/deactivated concepts used to characterize each cluster (r=5 or 10); chosen for readability rather than optimized against an external metric.
  • number of clustering heads N and overclustering factor = N=5, factor=3
    Architectural hyperparameters (N=5, overclustering factor 3) that affect discovery stability; selected by convention from prior multi-head NCD work rather than derived.
  • Sinkhorn ε and iteration count = ε=0.05, 3 iters
    Entropy regularization and iteration count for pseudo-label generation (ε=0.05, 3 iterations); standard but free choices that control cluster balance.
axioms (4)
  • domain assumption The number of novel categories Cu is known a priori.
    Stated in §3.1; standard in NCD but load-bearing for the unified softmax and Sinkhorn constraints.
  • domain assumption CLIP image-text cosine similarities constitute faithful dense supervision for human-interpretable visual concepts on both labeled and unlabeled images.
    Stage 2 (Eq. 2-5) treats CLIP as an external concept oracle; no independent verification that the retained concepts remain valid for novel classes.
  • ad hoc to paper A linear concept map c_ψ(x)=W_c φ_θ(x) (plus per-concept normalization) is an adequate bottleneck for discovery.
    Assumed throughout Prop. 4.1 and the architecture; non-linear concept maps are not explored.
  • domain assumption Known and novel category sets are disjoint and the labeled split is representative enough for knowledge transfer.
    Standard NCD problem statement (§3.1); required for the transfer claim.
invented entities (2)
  • Concept projection layer W_c trained by cubed-cosine alignment to CLIP no independent evidence
    purpose: Maps encoder features into a K-dimensional human-readable concept space without concept annotations.
    Core architectural invention; independent evidence is limited to the reported CLIP similarity statistics after filtering.
  • Cluster-level concept signatures and instance-level concept attributions no independent evidence
    purpose: Provide the human-readable explanations that distinguish xNCD from prior NCD methods.
    Defined post-hoc from mean concept activations (Appendix E); their semantic validity is not measured by an external human study.

pith-pipeline@v1.1.0-grok45 · 22091 in / 3424 out tokens · 48484 ms · 2026-07-11T17:31:23.460764+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Explainable Novel Category Discovery in Semantic Concept Space." pith.science (2026). https://pith.science/paper/ITLA762P

@misc{pith2026260704548,
  author       = {Pith},
  title        = {Pith review of: Explainable Novel Category Discovery in Semantic Concept Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITLA762P}},
  note         = {Machine review of arXiv:2607.04548}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Novel category discovery aims to identify unseen classes from unlabeled data by transferring knowledge from labeled categories, but most existing methods perform discovery in opaque latent feature spaces. As a result, they may separate novel categories accurately while providing little insight into what semantic evidence defines each discovered group. We propose xNCD, an explainable novel category discovery framework that performs both representation-based discovery and pseudo-label assignment directly in a structured semantic concept space. Instead of clustering arbitrary deep features, xNCD learns a label-free concept representation by aligning visual features with vision-language similarity priors from pretrained multimodal models, and then applies a unified labeled-and-unlabeled self-labeling objective over concept-space logits. This design makes each discovered category explainable by construction through stable concept signatures and instance-level concept evidence. Theoretically, we show that routing discovery through a semantic concept bottleneck induces a strict restriction of the feature-space hypothesis class, excluding a large family of unconstrained decision rules and biasing induced partitions toward semantically interpretable concept coordinates. Experiments on CIFAR-10, CIFAR-100, and CUB-200 demonstrate that xNCD preserves strong discovery performance while providing intrinsic explanations. Under task-agnostic evaluation, xNCD achieves 92.63% overall accuracy on CIFAR-10, close to UNO's 93.4%, and improves CIFAR-100 overall accuracy from 73.2% to 76.45%, while being the only compared method that provides human-readable cluster- and instance-level explanations.

Figures

Figures reproduced from arXiv: 2607.04548 by Ifrat Ikhtear Uddin, KC Santosh, Longwei Wang, Yang Zhou.

Figure 1
Figure 1. Figure 1: An input image x is processed by a pretrained encoder Eθ to produce features z ∈ R dz . The concept projection layer Wc maps these features to interpretable concept activations cˆ ∈ R K, where each dimension corresponds to a human-understandable attribute. Two classification heads operate on concept activations: a labeled head h for known categories and an unlabeled head g for novel categories discovery. L… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the xNCD framework. Stage 1: Encoder pretraining on labeled data with cross-entropy loss. Stage 2: Concept projection Wc is trained on all images by aligning concept activations with CLIP image-text similarities no concept annotations required. Stage 3: Novel categories are discovered by clustering in concept space, yielding interpretable clusters characterized by human-readable concept profile… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of concept space on CIFAR-10. Clear separation between all 10 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Concept activation profiles for discovered novel categories on CIFAR-10. The dog cluster [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Instance-level concept attribution for a CUB-200 bird image. Activated concepts (blue) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional cluster-level concept profiles for discovered novel categories [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional instance-level concept explanations for samples from CUB-200, CIFAR-10, and [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learning to Transmit: Volatility-Aware Predictive Communication for Energy-Efficient IoT Networks

    cs.IT 2026-07 conditional novelty 5.0

    Sensors using volatility-aware studentized residuals plus RLS online adaptation transmit up to 94.7% less IoT data while keeping reconstruction MAE at 0.35°C.

Reference graph

Works this paper leans on

57 extracted references · 15 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Novel class discovery without forgetting

    KJ Joseph, Sujoy Paul, Gaurav Aggarwal, Soma Biswas, Piyush Rai, Kai Han, and Vineeth N Balasubramanian. Novel class discovery without forgetting. InEuropean Conference on Computer Vision, pages 570–586. Springer, 2022

    2022

  2. [2]

    Autonovel: Automatically discovering and learning novel visual categories.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767–6781, 2021

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767–6781, 2021

    2021

  3. [3]

    Representation learning and nature encoded fusion for heterogeneous sensor networks.IEEE Access, 7:39227–39235, 2019

    Longwei Wang and Qilian Liang. Representation learning and nature encoded fusion for heterogeneous sensor networks.IEEE Access, 7:39227–39235, 2019

    2019

  4. [4]

    Congestion aware dynamic user association in heteroge- neous cellular network: A stochastic decision approach

    Longwei Wang, Wen Chen, and Jun Li. Congestion aware dynamic user association in heteroge- neous cellular network: A stochastic decision approach. In2014 IEEE International Conference on Communications (ICC), pages 2636–2640. IEEE, 2014

    2014

  5. [5]

    Explaining the behavior of neuron activations in deep neural networks.Ad Hoc Networks, 111:102346, 2021

    Longwei Wang, Chengfei Wang, Yupeng Li, and Rui Wang. Explaining the behavior of neuron activations in deep neural networks.Ad Hoc Networks, 111:102346, 2021

    2021

  6. [6]

    Exploration vs exploitation for distributed channel access in cognitive radio networks: A multi-user case study

    Longwei Wang, Xianfu Chen, Zhifeng Zhao, and Honggang Zhang. Exploration vs exploitation for distributed channel access in cognitive radio networks: A multi-user case study. In2011 11th International Symposium on Communications & Information Technologies (ISCIT), pages 360–365. IEEE, 2011

    2011

  7. [7]

    Deep reinforcement learning based computation offloading for mobility-aware edge computing

    Minyan Shi, Rui Wang, Erwu Liu, Zhixin Xu, and Longwei Wang. Deep reinforcement learning based computation offloading for mobility-aware edge computing. InInternational conference on communications and networking in china, pages 53–65. Springer International Publishing Cham, 2019

    2019

  8. [8]

    Improving robustness of deep neural networks via large-difference transformation.Neurocomputing, 450:411–419, 2021

    Longwei Wang, Chengfei Wang, Yupeng Li, and Rui Wang. Improving robustness of deep neural networks via large-difference transformation.Neurocomputing, 450:411–419, 2021

    2021

  9. [9]

    Looking beyond content: Modeling and detection of fake news from a social context perspective

    Kenan Xiao, Longwei Wang, Ashish Gupta, and Xiao Qin. Looking beyond content: Modeling and detection of fake news from a social context perspective. InProceedings of the 55th Hawaii International Conference on System Sciences 2022, pages 1–10, 2022

    2022

  10. [10]

    Layer-wise entropy analysis and visualization of neurons activation

    Longwei Wang, Peijie Chen, Chengfei Wang, and Rui Wang. Layer-wise entropy analysis and visualization of neurons activation. InInternational Conference on Communications and Networking in China, pages 29–36. Springer International Publishing Cham, 2019

    2019

  11. [11]

    Dense cross-connected ensemble convolutional neural networks for enhanced model robustness.arXiv preprint arXiv:2412.07022, 2024

    Longwei Wang, Xueqian Li, and Zheng Zhang. Dense cross-connected ensemble convolutional neural networks for enhanced model robustness.arXiv preprint arXiv:2412.07022, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Bridging interpretability and robustness using lime-guided model refinement.arXiv preprint arXiv:2412.18952, 2024

    Navid Nayyem, Abdullah Rakin, and Longwei Wang. Bridging interpretability and robustness using lime-guided model refinement.arXiv preprint arXiv:2412.18952, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Explainability- driven defense: grad-cam-guided model refinement against adversarial threats

    Longwei Wang, Ifrat Ikhtear Uddin, Xiao Qin, Yang Zhou, and KC Santosh. Explainability- driven defense: grad-cam-guided model refinement against adversarial threats. InProceedings of the AAAI Symposium Series (AAAI) 2025, volume 6, pages 49–57, 2025

    2025

  14. [14]

    Multi-scale unrectified push-pull with channel attention for enhanced corruption robustness

    Robin Narsingh Ranabhat, Longwei Wang, Xiao Qin, Yang Zhou, and KC Santosh. Multi-scale unrectified push-pull with channel attention for enhanced corruption robustness. InProceedings of the AAAI Symposium Series 2025, volume 6, pages 34–41, 2025

    2025

  15. [15]

    Expert-guided explainable few-shot learning for medical image diagnosis

    Ifrat Ikhtear Uddin, Longwei Wang, and KC Santosh. Expert-guided explainable few-shot learning for medical image diagnosis. InMICCAI Workshop on Data Engineering in Medical Imaging 2025, pages 95–104. Springer Nature Switzerland, 2025

    2025

  16. [16]

    Ecologically valid benchmarking and adaptive attention: Scalable marine bioacoustic monitoring.arXiv preprint arXiv:2509.04682, 2025

    Nicholas R Rasmussen, Rodrigue Rizk, Longwei Wang, and KC Santosh. Ecologically valid benchmarking and adaptive attention: Scalable marine bioacoustic monitoring.arXiv preprint arXiv:2509.04682, 2025. 10

    Pith/arXiv arXiv 2025

  17. [17]

    Toward carbon-neutral human ai: Rethinking data, computation, and learning paradigms for sustainable intelligence

    KC Santosh, Rodrigue Rizk, and Longwei Wang. Toward carbon-neutral human ai: Rethinking data, computation, and learning paradigms for sustainable intelligence. In2025 IEEE 7th International Conference on Cognitive Machine Intelligence (CogMI), 2025

    2025

  18. [18]

    Expert-guided explainable few-shot learning with active sample selection for medical image analysis.IEEE Journal of Biomedical and Health Informatics, 2026

    Longwei Wang, Ifrat Ikhtear Uddin, and KC Santosh. Expert-guided explainable few-shot learning with active sample selection for medical image analysis.IEEE Journal of Biomedical and Health Informatics, 2026

    2026

  19. [19]

    Acting flatterers via llms sycophancy: Combating clickbait with llms opposing-stance reasoning

    Chaowei Zhang, Xiansheng Luo, Zewei Zhang, Yi Zhu, Jipeng Qiang, and Longwei Wang. Acting flatterers via llms sycophancy: Combating clickbait with llms opposing-stance reasoning. InProceedings of the ACM Web Conference (WWW) 2026, pages 3195–3206, 2026

    2026

  20. [20]

    Bridging symmetry and robustness: On the role of equivariance in enhancing adversarial robustness

    Longwei Wang, Ifrat Ikhtear Uddin, Chaowei Zhang, Xiao Qin, and Yang Zhou. Bridging symmetry and robustness: On the role of equivariance in enhancing adversarial robustness. Advances in Neural Information Processing Systems (NeurIPS), 38:159102–159129, 2025

    2025

  21. [21]

    Channel- selected stratified nested cross-validation for clinically relevant eeg-based parkinson’s disease detection

    Nicholas R Rasmussen, Rodrigue Rizk, Longwei Wang, Arun Singh, and KC Santosh. Channel- selected stratified nested cross-validation for clinically relevant eeg-based parkinson’s disease detection. In2026 IEEE Conference on Artificial Intelligence (CAI), pages 91–97. IEEE, 2026

    2026

  22. [22]

    Winsor-cam: Human-tunable visual explanations from deep networks via layer-wise winsorization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    Casey Wall, Longwei Wang, Rodrigue Rizk, and KC Santosh. Winsor-cam: Human-tunable visual explanations from deep networks via layer-wise winsorization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  23. [23]

    Promoting shape bias in cnns: Frequency-based and contrastive regularization for corruption robustness

    Robin Narsingh Ranabhat, Longwei Wang, Amit Kumar Patel, and KC Santosh. Promoting shape bias in cnns: Frequency-based and contrastive regularization for corruption robustness. In International Conference on Intelligent Systems and Pattern Recognition, pages 16–26. Springer, 2025

    2025

  24. [24]

    Coswin: Convolu- tion enhanced hierarchical shifted window attention for small-scale vision.arXiv preprint arXiv:2509.08959, 2025

    Puskal Khadka, Rodrigue Rizk, Longwei Wang, and KC Santosh. Coswin: Convolu- tion enhanced hierarchical shifted window attention for small-scale vision.arXiv preprint arXiv:2509.08959, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Explainability-guided defense: Attribution-aware model refinement against adversarial data attacks

    Longwei Wang, Mohammad Navid Nayyem, Abdullah Al Rakin, KC Santosh, Chaowei Zhang, and Yang Zhou. Explainability-guided defense: Attribution-aware model refinement against adversarial data attacks. In2025 IEEE International Conference on Data Mining (ICDM), pages 1585–1592. IEEE, 2025

    2025

  26. [26]

    A unified objective for novel class discovery

    Enrico Fini, Enver Sangineto, Stéphane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. InProceedings of the IEEE/CVF international conference on computer vision, pages 9284–9292, 2021

    2021

  27. [27]

    Learning to discover novel visual categories via deep transfer clustering

    Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. InProceedings of the IEEE/CVF international conference on computer vision, pages 8401–8409, 2019

    2019

  28. [28]

    Semantic-guided novel category discovery

    Weishuai Wang, Ting Lei, Qingchao Chen, and Yang Liu. Semantic-guided novel category discovery. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5607–5614, 2024

    2024

  29. [29]

    Explaining deep neural networks and beyond: A review of methods and applications.Proceedings of the IEEE, 109(3):247–278, 2021

    Wojciech Samek, Grégoire Montavon, Sebastian Lapuschkin, Christopher J Anders, and Klaus- Robert Müller. Explaining deep neural networks and beyond: A review of methods and applications.Proceedings of the IEEE, 109(3):247–278, 2021

    2021

  30. [30]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature machine intelligence, 1(5):206–215, 2019

    Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature machine intelligence, 1(5):206–215, 2019

    2019

  31. [31]

    The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery.Queue, 16(3):31–57, 2018

    Zachary C Lipton. The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery.Queue, 16(3):31–57, 2018

    2018

  32. [32]

    Towards a rigorous science of interpretable machine learning

    Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 11

    Pith/arXiv arXiv 2017

  33. [33]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. InInternational conference on machine learning, pages 5338–5348. PMLR, 2020

    2020

  34. [34]

    Post-hoc concept bottleneck models.arXiv preprint arXiv:2205.15480, 2022

    Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models.arXiv preprint arXiv:2205.15480, 2022

    Pith/arXiv arXiv 2022

  35. [35]

    Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav)

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). InInternational conference on machine learning, pages 2668–2677. PMLR, 2018

    2018

  36. [36]

    Label-free concept bottleneck models.arXiv preprint arXiv:2304.06129, 2023

    Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models.arXiv preprint arXiv:2304.06129, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    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. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  38. [38]

    Learning to cluster in order to transfer across domains and tasks.arXiv preprint arXiv:1711.10125, 2017

    Yen-Chang Hsu, Zhaoyang Lv, and Zsolt Kira. Learning to cluster in order to transfer across domains and tasks.arXiv preprint arXiv:1711.10125, 2017

    Pith/arXiv arXiv 2017

  39. [39]

    Multi-class classification without multi-class labels.arXiv preprint arXiv:1901.00544, 2019

    Yen-Chang Hsu, Zhaoyang Lv, Joel Schlosser, Phillip Odom, and Zsolt Kira. Multi-class classification without multi-class labels.arXiv preprint arXiv:1901.00544, 2019

    Pith/arXiv arXiv 1901

  40. [40]

    Automatically discovering and learning new visual categories with ranking statistics.arXiv preprint arXiv:2002.05714, 2020

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics.arXiv preprint arXiv:2002.05714, 2020

    Pith/arXiv arXiv 2002

  41. [41]

    Generalized category discovery

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7492–7501, 2022

    2022

  42. [42]

    Open-world semi-supervised learning.arXiv preprint arXiv:2102.03526, 2021

    Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning.arXiv preprint arXiv:2102.03526, 2021

    Pith/arXiv arXiv 2021

  43. [43]

    Parametric classification for generalized category discovery: A baseline study

    Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. InProceedings of the IEEE/CVF international conference on computer vision, pages 16590–16600, 2023

    2023

  44. [44]

    Xcon: Learning with experts for fine-grained category discovery.arXiv preprint arXiv:2208.01898, 2022

    Yixin Fei, Zhongkai Zhao, Siwei Yang, and Bingchen Zhao. Xcon: Learning with experts for fine-grained category discovery.arXiv preprint arXiv:2208.01898, 2022

    Pith/arXiv arXiv 2022

  45. [45]

    Novel class discovery: an introduction and key concepts

    Colin Troisemaine, Vincent Lemaire, Stéphane Gosselin, Alexandre Reiffers-Masson, Joachim Flocon-Cholet, and Sandrine Vaton. Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028, 2023

    Pith/arXiv arXiv 2023

  46. [46]

    Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems, 33:9912–9924, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems, 33:9912–9924, 2020

    2020

  47. [47]

    Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    2013

  48. [48]

    Learning multiple layers of features from tiny images

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

    2009

  49. [49]

    The caltech-ucsd birds-200-2011 dataset

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011

    2011

  50. [50]

    The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    Harold W Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955. 12

    1955

  51. [51]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  52. [52]

    Clip-dissect: Automatic description of neuron repre- sentations in deep vision networks.arXiv preprint arXiv:2204.10965, 2022

    Tuomas Oikarinen and Tsui-Wei Weng. Clip-dissect: Automatic description of neuron repre- sentations in deep vision networks.arXiv preprint arXiv:2204.10965, 2022

    Pith/arXiv arXiv 2022

  53. [53]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  54. [54]

    Self-labelling via simultaneous clustering and representation learning.arXiv preprint arXiv:1911.05371, 2019

    Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning.arXiv preprint arXiv:1911.05371, 2019

    Pith/arXiv arXiv 1911

  55. [55]

    The national research platform: Stretched, multi-tenant, scientific kubernetes cluster

    Derek Weitzel, Ashton Graves, Sam Albin, Huijun Zhu, Frank Wuerthwein, Mahidhar Tatineni, Dmitry Mishin, Elham Khoda, Mohammad Sada, Larry Smarr, et al. The national research platform: Stretched, multi-tenant, scientific kubernetes cluster. InPractice and Experience in Advanced Research Computing 2025: The Power of Collaboration, pages 1–5, 2025. 13 A Con...

    2025

  56. [56]

    four-legged,

    Interpretable clustering: Samples are grouped together because they share similarconcept activations, not because of opaque feature correlations. A cluster of dogs emerges because its members all activate “four-legged,” “furry,” and “mammal” concepts. 2.Semantic consistency: The Sinkhorn-Knopp constraints operate in concept space, encour- aging each disco...

  57. [57]

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...