Pith. sign in

REVIEW 5 major objections 5 minor 46 references

Empirical Evidences for the Effects of Feature Diversity in Open Set Recognition and Continual Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Training a classifier on a task that forces it to use more features—shape as well as color—improves its ability to detect novel classes, retain old knowledge, and learn new classes, according to controlled experiments on synthetic images.

desk verdict Small controlled study of feature diversity in OSR/CL that has a real idea but never measures diversity and undercuts its own retention claim—worth refereeing as a critique exercise, not citable as evidence. read the letter →

arxiv 2508.13005 v1 pith:25QG4NJC submitted 2025-08-18 cs.CV cs.LG

Jiawen Xu , Odej Kao This is my paper
classification cs.CVcs.LG
keywords opensetrecognitioncontinuallearningclass-incrementalfeaturediversitycatastrophicforgettingforwardtransferMahalanobisdistanceCKA
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 asks whether feature diversity—having a classifier rely on a richer set of discriminative cues—is itself what makes open set recognition and continual learning work, rather than a side effect of specific algorithms. To test this, the authors train small CNNs on synthetic color-and-shape images: one model (E1) can solve its task using color alone, while a second (E2) is forced to also use shape. Across three measurements—histogram separation of Mahalanobis distances between known and unknown classes, CKA feature similarity before and after learning a new task, and linear-probing accuracy on new classes—the model that learned more diverse features detects unknown classes better, forgets past features less, and transfers better to new tasks with overlapping features. The authors conclude that learning as diverse features as possible from observed data, while preserving current-task accuracy, is a useful principle for both problems.

What carries the argument

The load-bearing setup is a controlled task composition on synthetic images, where each class is one color-shape combination and the background is black during training. E1's two classes can be separated by color alone; E2's third class forces the network to encode shape as well, making 'feature diversity' a property of the task rather than of a diversity-promoting loss. The evaluation machinery consists of three measures: the normalized histogram distance D_hist between Mahalanobis distances of close-set and open-set features (Equation 2), centered kernel alignment (CKA) between Task 0 features before and after Task 1, and linear probing accuracy on Task 1 with frozen features. These metric

What would settle it

Measure the effective rank or redundancy of the penultimate-layer features of E1 and E2 on the same synthetic data: the claimed mechanism predicts E2 is measurably more diverse. A control experiment replacing the third class in E2 with a duplicate of an existing class, so no new feature is required, should show no OSR or continual-learning gains if diversity, not task difficulty, is the cause.

Watch

Extended reading notes

Core claim

The central claim is that feature diversity is causally beneficial in two settings that are usually studied separately. On the synthetic dataset, the task is engineered: E1 distinguishes blue circles from red rectangles, which color alone can do, while E2 adds red circles, so correct classification requires shape as well as color. The paper shows that E2's feature space separates open-set classes from close-set classes at least as well as E1's under the normalized histogram distance D_hist of Mahalanobis distances, and that this gap appears early in training and persists. In continual learning, models initialized from E2 retain more of Task 0's features according to CKA and achieve higher li

Load-bearing premise

The paper's mechanism assumes that forcing E2 to classify a red circle actually produces more diverse features than E1's color-only solution (no direct diversity metric verifies this), and that small CKA differences indicate meaningful forgetting even though the paper reports old-class testing accuracy around 0.5 in every continual learning experiment.

Editorial extensions

If this is right

  • If the claim is right, OSR methods that encourage diverse features are not just heuristics: they address a direct cause of poor novelty detection.
  • In class-incremental learning, a base model trained with more diverse features should suffer less catastrophic forgetting of past-task features, and should adapt faster when new classes overlap old features.
  • Because future tasks are unknown, it is reasonable to train on observed data to preserve as much feature diversity as possible while keeping current-task accuracy; diversity becomes a design objective rather than an accident.
  • The same logic should carry over to real data, but only where feature overlaps between old and new classes exist; when new classes introduce entirely novel features, base feature diversity alone gives no transfer gain.

Reading between the lines

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

  • A direct test is missing from the paper: applying an explicit diversity measure (effective rank, redundancy, or feature count) to E1 vs E2 penultimate-layer features. If E2 is not measurably more diverse, the causal story would need revision; this is an editorial next step.
  • Feature diversity likely has a cost: forcing a network to encode more features can lower closed-set accuracy or reduce margin on easy classes. The paper's 'as diverse as possible while maintaining performance' is a trade-off that remains to be quantified for real datasets.
  • The CKA-based forgetting results are suggestive, but the paper itself notes old-class accuracy is near chance after Task 1; a stronger test would rerun the comparison with a replay-free base that can actually classify old classes, then measure accuracy decay.
  • Entangled features explain the E9/E10 case: pink ellipses are close to red circles in E2, so apparent transfer may be color similarity rather than genuine reusable shape features. Disentangling color and shape would make feature-diversity effects cleaner.
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

5 major / 5 minor

Summary. The paper presents a controlled study on synthetic color/shape images to ask whether learning more diverse features helps open set recognition (OSR) and continual learning (CL). Two base classifiers are compared: E1, a binary classifier that can succeed using color alone, and E2, a three-class classifier that must also use shape. OSR is evaluated by the histogram distance Dhist between Mahalanobis distances of close-set and open-set features (Eq. 2, Fig. 3). For CL, the base models are extended to new tasks; feature forgetting is measured by CKA between Task-0 features before/after Task 1 (Table 3), and forward transfer by linear probing accuracy (Table 4). The authors conclude that diverse features improve OSR, reduce feature forgetting, and help learn new tasks.

Significance. If the claimed effects were established, the paper would provide a simple, controlled empirical justification for the common heuristic of promoting feature diversity in OSR and CL. The synthetic setup is a strength: it allows the experimenter to know which input features are task-relevant, and the use of representation-level metrics (CKA, Mahalanobis distances) is appropriate for probing internal representations. However, the current evidence is not sufficient to support the causal claims. The independent variable (feature diversity) is never directly measured, the retention result is contradicted by the authors' own report of chance-level old-task accuracy, and the quantitative results are single-run with no error bars. The paper is better treated as a preliminary study than as definitive empirical evidence.

major comments (5)
  1. [Section 3.2, Table 2a] Feature diversity is never measured. The paper infers that E2 learns more diverse features solely because its task includes red circles, forcing the model to use shape in addition to color. No direct diversity metric (e.g., effective rank, participation ratio, feature redundancy, or mutual information between features) is reported for the learned representations. The confusion matrices in Fig. 2 show that E2 has difficulty distinguishing red rectangles from red circles, but this is evidence about task difficulty, not about the diversity of the learned feature set. Alternative explanations for any behavioral difference—number of classes, optimization trajectory, or decision-boundary complexity—are not ruled out.
  2. [Section 3.3, first paragraph] The authors state: 'the testing accuracy of the new models on the observed data in all experiments is around 0.5, i.e., almost a guess.' This directly contradicts the abstract's claim that increased feature diversity facilitates 'the retention of previously learned data.' No actual retention accuracy is reported; the paper instead switches to CKA feature similarity. CKA may be a useful diagnostic, but it is not calibrated to any task-level retention measure, and the reader cannot infer from it that previously learned data are retained. The claims in the abstract and conclusion need to be restricted to 'feature forgetting' as measured by CKA, or the experiments need to include a setting where old-task accuracy is above chance.
  3. [Table 3 and Section 3.3.1] The CKA results are too weak and inconsistent to support the forgetting claims. All reported values are small (≈0.009–0.074), and the ordering varies across rows. For E7/E8, the E2(2) CKA is 0.0159, below the E1 value of 0.017, while E2(3) is 0.074; the choice between reporting E2(2) and E2(3) is not justified. No standard errors, confidence intervals, or replicated runs are provided, so the differences (e.g., 0.0127 vs. 0.016) may be noise. A principled comparison would require multiple random seeds and a calibrated measure of feature forgetting.
  4. [Equation (2) and Figure 3] The OSR metric Dhist depends on the free parameter Nbins (number of histogram bins), but no sensitivity analysis is given. The authors state that Dhist values for E2 are 'larger or at least equal' to those for E1, but the figure is a set of single curves without error bars or statistical tests. It is unclear whether the small visual differences are reproducible or significant. The paper should report numerical Dhist values, vary Nbins, and provide multiple runs.
  5. [Sections 3.2 and 3.3] The experimental design confounds feature diversity with task composition. E1 is a binary classifier, while E2 is a three-class classifier; the two also differ in which classes are present. Any observed improvements in OSR or forward transfer could be due to the larger number of training classes, the particular class statistics, or the need to separate red circles from red rectangles, rather than to 'feature diversity' as a general property. A cleaner control would compare models with the same number of classes and the same data, but different feature diversity (e.g., by varying the redundancy of available cues), or directly measure diversity of the learned features to demonstrate that E2 is indeed more diverse.
minor comments (5)
  1. [Abstract / Section 1] The abbreviation 'ODD' should be 'OOD' (out-of-distribution).
  2. [Section 3.1] The sentence 'the connection between feature diversity and continual learning as well as continual learning' appears to have a duplicated phrase; likely one occurrence should be 'open set recognition.'
  3. [Figure 2] The caption lists epochs 0, 10, 20, 30, 50, and 100, but the panels show 0, 10, 30, 50, and 100 (missing epoch 20). Please align the caption with the figure.
  4. [Table 3] The column headers 'E2(2)' and 'E2(3)' are not defined in the table caption; the text explains them but the caption should be self-contained.
  5. [Throughout] The terms 'open-sets' and 'close-sets' are unconventional; 'open-set samples' and 'closed-set samples' would be clearer.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; claims are empirical, though 'feature diversity' is inferred from task design rather than measured.

full rationale

The paper performs controlled experiments and contains no derivation chain that reduces its conclusions to its inputs. No parameter is fitted and then reported as a prediction; no uniqueness theorem or ansatz is imported from prior work; the only self-citation ([42]) appears in related work and is not load-bearing. The central independent variable, feature diversity, is operationalized by task composition ('E2 models are encouraged to learn more comprehensive features than in E1') and is never directly measured; this is a construct-validity and confound concern (class count and task difficulty are not controlled), not a circular computation. Similarly, the retention claim relies on CKA values despite the paper's admission that observed-data accuracy is 'around 0.5, i.e., almost a guess'; that is an evidential floor effect, not a circular step. Because the conclusions are empirical observations rather than derivations, no circularity score above 2 is warranted.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an unmeasured construct (feature diversity), a proxy OSR metric, and a forgetting metric that is not tied to task accuracy. These are domain assumptions or ad hoc operationalizations, not formal derivations.

free parameters (1)
  • Nbins in Dhist
    The number of histogram bins for Dhist is not reported and can change the overlap value, affecting the OSR comparison.
assumptions (4)
  • ad hoc to paper E2 learns more diverse features than E1 because its task forces shape plus color while E1 only needs color.
    The paper assumes the task design creates feature diversity but never measures feature diversity directly. Location: Section 3.2.
  • domain assumption CKA between features before and after Task 1 quantifies feature forgetting.
    CKA measures representation similarity, but the paper equates low CKA with forgetting without checking task accuracy, which contradicts the stated chance-level accuracy. Location: Section 3.3.1.
  • domain assumption Mahalanobis distance histogram overlap (Dhist) is a valid OSR performance measure.
    Dhist is used as a proxy for open set detection without computing standard metrics such as AUROC. Location: Section 3.2.
  • domain assumption CNN color bias implies E1 relies primarily on color features.
    The paper depends on Geirhos et al. [14] and Singh et al. [38] to infer feature reliance; confusion matrices provide supporting but indirect evidence. Location: Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Empirical Evidences for the Effects of Feature Diversity in Open Set Recognition and Continual Learning." pith.science (2026). https://pith.science/paper/25QG4NJC

@misc{pith2026250813005,
  author       = {Pith},
  title        = {Pith review of: Empirical Evidences for the Effects of Feature Diversity in Open Set Recognition and Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/25QG4NJC}},
  note         = {Machine review of arXiv:2508.13005}
}
read the original abstract

Open set recognition (OSR) and continual learning are two critical challenges in machine learning, focusing respectively on detecting novel classes at inference time and updating models to incorporate the new classes. While many recent approaches have addressed these problems, particularly OSR, by heuristically promoting feature diversity, few studies have directly examined the role that feature diversity plays in tackling them. In this work, we provide empirical evidence that enhancing feature diversity improves the recognition of open set samples. Moreover, increased feature diversity also facilitates both the retention of previously learned data and the integration of new data in continual learning. We hope our findings can inspire further research into both practical methods and theoretical understanding in these domains.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 4 linked inside Pith

  1. [1]

    Memory aware synapses: Learning what (not) to forget

    Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139–154, 2018. 2

    2018

  2. [2]

    Make continual learning stronger via c-flat

    Ang Bian, Wei Li, Hangjie Yuan, Mang Wang, Zixiang Zhao, Aojun Lu, Pengliang Ji, Tao Feng, et al. Make continual learning stronger via c-flat. Advances in Neural Information Processing Systems, 37:7608–7630, 2024. 1

    2024

  3. [3]

    Semi- supervised novelty detection.The Journal of Machine Learn- ing Research, 11:2973–3009, 2010

    Gilles Blanchard, Gyemin Lee, and Clayton Scott. Semi- supervised novelty detection.The Journal of Machine Learn- ing Research, 11:2973–3009, 2010. 1

    2010

  4. [4]

    Dark experience for general continual learning: a strong, simple baseline

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. In Advances in Neural Information Processing Systems , pages 15920– 15930. Curran Associates, Inc., 2020. 1

    2020

  5. [5]

    Open-set recognition with gaussian mixture variational autoencoders

    Alexander Cao, Yuan Luo, and Diego Klabjan. Open-set recognition with gaussian mixture variational autoencoders. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 6877–6884, 2021. 2

    2021

  6. [6]

    On measuring the distance between histograms

    Sung-Hyuk Cha and Sargur N Srihari. On measuring the distance between histograms. Pattern Recognition, 35(6): 1355–1370, 2002. 3

    2002

  7. [7]

    Novelty de- tection via contrastive learning with negative data augmenta- tion

    Chengwei Chen, Yuan Xie, Shaohui Lin, Ruizhi Qiao, Jian Zhou, Xin Tan, Yi Zhang, and Lizhuang Ma. Novelty de- tection via contrastive learning with negative data augmenta- tion. arXiv preprint arXiv:2106.09958, 2021. 1

    arXiv 2021

  8. [8]

    Learning open set network with discriminative reciprocal points

    Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 507–522. Springer, 2020. 1

    2020

Show all 46 references
  1. [9]

    Is forgetting less a good inductive bias for for- ward transfer? arXiv preprint arXiv:2303.08207, 2023

    Jiefeng Chen, Timothy Nguyen, Dilan Gorur, and Arslan Chaudhry. Is forgetting less a good inductive bias for for- ward transfer? arXiv preprint arXiv:2303.08207, 2023. 4

    2023 arXiv

  2. [10]

    Imagenet: A large-scale hierarchical im- age database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical im- age database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR’09), pages 248–255. IEEE,

    2009

  3. [11]

    Reducing network agnostophobia

    Akshay Raj Dhamija, Manuel G ¨unther, and Terrance Boult. Reducing network agnostophobia. Advances in Neural In- formation Processing Systems, 31, 2018. 1

    2018

  4. [12]

    Re- visiting deep ensemble for out-of-distribution detection: A loss landscape perspective

    Kun Fang, Qinghua Tao, Xiaolin Huang, and Jie Yang. Re- visiting deep ensemble for out-of-distribution detection: A loss landscape perspective. International Journal of Com- puter Vision, 132(12):6107–6126, 2024. 1

    2024

  5. [13]

    Gen- erative openmax for multi-class open set classification

    Zongyuan Ge, Sergey Demyanov, and Rahil Garnavi. Gen- erative openmax for multi-class open set classification. In British Machine Vision Conference 2017 (BMVC’17) , Lon- don, UK, 2017. 1

    2017

  6. [14]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In Interna- tional conference on learning representations, 2018. 2

    2018

  7. [15]

    Learning a neural- network-based representation for open set recognition

    Mehadi Hassen and Philip K Chan. Learning a neural- network-based representation for open set recognition. In Proceedings of the 2020 SIAM International Conference on Data Mining, pages 154–162. SIAM, 2020. 1

    2020

  8. [16]

    Learning a unified classifier incrementally via rebalancing

    Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831–839,

  9. [17]

    Overcoming catastrophic forgetting for continual learning via model adaptation

    Wenpeng Hu, Zhou Lin, Bing Liu, Chongyang Tao, Zheng- wei Tao Tao, Dongyan Zhao, Jinwen Ma, and Rui Yan. Overcoming catastrophic forgetting for continual learning via model adaptation. In International conference on learn- ing representations, 2019. 2

    2019

  10. [18]

    Compacting, picking and growing for unforgetting continual learning.Ad- vances in neural information processing systems , 32, 2019

    Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning.Ad- vances in neural information processing systems , 32, 2019. 2

    2019

  11. [19]

    Class- incremental learning by knowledge distillation with adaptive feature consolidation

    Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class- incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 16071–16080, 2022. 2

    2022

  12. [20]

    Large-scale video classification with convolutional neural networks

    Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Pro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 1

    2014

  13. [21]

    Overcoming catastrophic forgetting in neu- ral networks

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

    2017

  14. [22]

    Similarity of neural network represen- tations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network represen- tations revisited. In International conference on machine learning, pages 3519–3529. PMlR, 2019. 4

    2019

  15. [23]

    A simple unified framework for detecting out-of-distribution samples and adversarial attacks

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural infor- mation processing systems, 31, 2018. 3

    2018

  16. [24]

    Applications of machine learning to machine fault diagnosis: A review and roadmap.Mechanical Systems and Signal Processing, 138:106587, 2020

    Yaguo Lei, Bin Yang, Xinwei Jiang, Feng Jia, Naipeng Li, and Asoke K Nandi. Applications of machine learning to machine fault diagnosis: A review and roadmap.Mechanical Systems and Signal Processing, 138:106587, 2020. 1

    2020

  17. [25]

    Con- trastive continual learning with importance sampling and prototype-instance relation distillation

    Jiyong Li, Dilshod Azizov, Shangsong Liang, et al. Con- trastive continual learning with importance sampling and prototype-instance relation distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13554– 13562, 2024. 1

    2024

  18. [26]

    Learn to grow: A continual structure learn- ing framework for overcoming catastrophic forgetting

    Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learn- ing framework for overcoming catastrophic forgetting. In 5 International conference on machine learning, pages 3925–

  19. [27]

    Provable and efficient continual representation learn- ing

    Yingcong Li, Mingchen Li, M Salman Asif, and Samet Oy- mak. Provable and efficient continual representation learn- ing. arXiv preprint arXiv:2203.02026, 2022. 1

    2022 arXiv

  20. [28]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelli- gence, 40(12):2935–2947, 2017. 2

    2017

  21. [29]

    Few-shot open-set recognition using meta- learning

    Bo Liu, Hao Kang, Haoxiang Li, Gang Hua, and Nuno Vasconcelos. Few-shot open-set recognition using meta- learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8798– 8807, 2020. 3

    2020

  22. [30]

    Catastrophic inter- ference in connectionist networks: The sequential learning problem

    Michael McCloskey and Neal J Cohen. Catastrophic inter- ference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation , pages 109–165. Elsevier, 1989. 2

    1989

  23. [31]

    Class anchor clustering: A loss for distance- based open set recognition

    Dimity Miller, Niko Suenderhauf, Michael Milford, and Feras Dayoub. Class anchor clustering: A loss for distance- based open set recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV’21), pages 3570–3578, 2021. 1

    2021

  24. [32]

    Open set learning with counterfac- tual images

    Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfac- tual images. In Proceedings of the European Conference on Computer Vision (ECCV’18), pages 613–628, 2018. 1

    2018

  25. [33]

    C2ae: Class conditioned auto-encoder for open-set recognition

    Poojan Oza and Vishal M Patel. C2ae: Class conditioned auto-encoder for open-set recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2307–2316, 2019. 2

    2019

  26. [34]

    Generative-discriminative feature representations for open-set recognition

    Pramuditha Perera, Vlad I Morariu, Rajiv Jain, Varun Man- junatha, Curtis Wigington, Vicente Ordonez, and Vishal M Patel. Generative-discriminative feature representations for open-set recognition. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Reco...

    2020

  27. [35]

    A review of novelty detection

    Marco AF Pimentel, David A Clifton, Lei Clifton, and Li- onel Tarassenko. A review of novelty detection. Signal pro- cessing, 99:215–249, 2014. 1

    2014

  28. [36]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE con- ference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 1, 2

    2001

  29. [37]

    Overcoming catastrophic forgetting with hard attention to the task

    Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International conference on machine learning, pages 4548–4557. PMLR, 2018. 1

    2018

  30. [38]

    As- sessing the importance of colours for cnns in object recogni- tion

    Aditya Singh, Alessandro Bay, and Andrea Mirabile. As- sessing the importance of colours for cnns in object recogni- tion. arXiv preprint arXiv:2012.06917, 2020. 2

    2012 arXiv

  31. [39]

    Dice: Leveraging sparsification for out-of-distribution detection

    Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution detection. In European conference on computer vision, pages 691–708. Springer, 2022. 1

    2022

  32. [40]

    Ex- ploring diverse representations for open set recognition

    Yu Wang, Junxian Mu, Pengfei Zhu, and Qinghua Hu. Ex- ploring diverse representations for open set recognition. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 5731–5739, 2024. 1

    2024

  33. [41]

    Learning to prompt for con- tinual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jen- nifer Dy, and Tomas Pfister. Learning to prompt for con- tinual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139–149,

  34. [42]

    Openincre- ment: A unified framework for open set recognition and deep class-incremental learning

    Jiawen Xu, Claas Grohnfeldt, and Odej Kao. Openincre- ment: A unified framework for open set recognition and deep class-incremental learning. In Proceedings of the IEEE/CVF international conference on computer vision , pages 3303– 3311, 2023. 1

    2023

  35. [43]

    Convolutional prototype network for open set recognition

    Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, Qing Yang, and Cheng-Lin Liu. Convolutional prototype network for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2358–2370, 2020. 1

    2020

  36. [44]

    Openood: Benchmarking generalized out-of-distribution detection

    Jingkang Yang, Pengyun Wang, Dejian Zou, Zitang Zhou, Kunyuan Ding, Wenxuan Peng, Haoqi Wang, Guangyao Chen, Bo Li, Yiyou Sun, et al. Openood: Benchmarking generalized out-of-distribution detection. Advances in Neu- ral Information Processing Systems, 35:32598–32611, 2022. 1

    2022

  37. [45]

    Generalized out-of-distribution detection: A survey

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. Inter- national Journal of Computer Vision , 132(12):5635–5662,

  38. [46]

    Classification- reconstruction learning for open-set recognition

    Ryota Yoshihashi, Wen Shao, Rei Kawakami, Shaodi You, Makoto Iida, and Takeshi Naemura. Classification- reconstruction learning for open-set recognition. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4016–4025, 2019. 2

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.