pith. machine review for the scientific record. sign in

arxiv: 2605.08183 · v1 · submitted 2026-05-05 · 💻 cs.CV · cs.LG

Recognition: no theorem link

Sparsity Hurts: Simple Linear Adapter Can Boost Generalized Category Discovery

Bo Ye, Kai Gan, Min-Ling Zhang, Tong Wei

Pith reviewed 2026-05-12 01:08 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords generalized category discoverylinear adaptervision transformerfeature sparsityresidual adapterdistribution alignmentnovel categoriespre-trained models
0
0 comments X

The pith

Embedding a residual linear adapter in each ViT block boosts generalized category discovery by countering the harm from feature sparsity.

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

The paper establishes that non-linear adapters commonly used in adapting pre-trained vision models impair results on generalized category discovery tasks because they interact poorly with sparse features. In their place, a simple residual linear adapter placed inside every transformer block provides more flexible capacity for the model to adjust to unlabeled data containing both known and unknown classes. An added auxiliary loss aligns the output distributions to reduce bias toward seen categories. This yields consistent gains over partial fine-tuning and prompt-based methods across generic and fine-grained image datasets. A sympathetic reader would care because it shows a minimal change that unlocks better use of existing pre-trained models for discovering novel categories without extra overfitting risk.

Core claim

From the perspective of feature sparsity, non-linearity in conventional adapters impairs performance, whereas our linear adapter enhances it by enabling more flexible model capacity. LAGCD embeds a residual linear adapter into each ViT block and introduces an auxiliary distribution alignment loss to mitigate the negative impact of biased predictions between seen and novel categories.

What carries the argument

The residual linear adapter inserted into each ViT block, which adapts representations linearly to increase model capacity while avoiding non-linear transformations that degrade sparse features.

If this is right

  • The entire pre-trained ViT can be adapted more flexibly than with partial fine-tuning of only the final block.
  • Overfitting is reduced compared with visual prompt tuning because the linear adapter has constrained yet sufficient capacity.
  • An auxiliary loss that aligns seen and novel category distributions counters prediction bias without complicating the core architecture.
  • Consistent accuracy gains appear on both generic and fine-grained datasets over prior sophisticated GCD baselines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same sparsity argument could motivate linear adapters in other transfer settings where pre-trained features are known to be sparse.
  • Designers of future adapters might test linearity first before adding non-linearities, especially when adapting models for open-set recognition tasks.
  • The placement across all blocks suggests that full-model linear adaptation may generalize to other vision problems that require both retention of old knowledge and acquisition of new classes.

Load-bearing premise

The observed gains truly arise from the linear adapter's handling of feature sparsity rather than from the auxiliary loss, specific hyperparameter settings, or other unstated design elements.

What would settle it

An experiment in which the linear adapter is added to the model but the auxiliary distribution alignment loss is removed, and performance either stays flat or drops below the non-linear adapter baselines.

Figures

Figures reproduced from arXiv: 2605.08183 by Bo Ye, Kai Gan, Min-Ling Zhang, Tong Wei.

Figure 1
Figure 1. Figure 1: Performance comparisons with SimGCD [1], SPTNet [2], and AdaptGCD [3]. (a) We compare the number of tunable parameters of the ViT backbone, and accuracy averaged over four fine-grained and three generic datasets. (b) Seen and novel class accuracy of our method and existing methods on seven datasets. 1 INTRODUCTION R ECENTLY, semi-supervised learning (SSL) [4]–[10] has emerged as a promising paradigm for le… view at source ↗
Figure 3
Figure 3. Figure 3: Different thresholds t in Threshold-ReLU. 0 1.0 2.0 3.0 linear 55 60 65 Accuracy (%) FGVC-Aircraft All Seen Novel 0 1.0 2.0 3.0 linear 50 60 70 80 Accuracy (%) Stanford Cars All Seen Novel [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Different slopes β in ELU. explore different activation settings by varying the slope α ∈ {0, 0.01, 0.5, 0.99, 1}, the threshold t ∈ {−2, −1, 0, 1, 2}, and the slope β ∈ {0, 1, 2, 3}, while keeping all other hy￾perparameters consistent. In addition, to examine whether different adapters induce overfitting or underfitting, we compute feature similarity for each sample before and after fine-tuning the pre-tr… view at source ↗
Figure 2
Figure 2. Figure 2: Different slopes α in Leaky-ReLU. For experimental validation, we employ SimGCD [1] with adapters on two challenging fine-grained datasets FGVC-Aircraft [40] and Stanford Cars [41], where ReLU is replaced with Leaky-ReLU, Threshold-ReLU, or ELU. We linear -2 -1 0 1 2 t 45 50 55 60 65 70 Accuracy (%) FGVC-Aircraft All Seen Novel linear -2 -1 0 1 2 t 40 50 60 70 80 Accuracy (%) Stanford Cars All Seen Novel … view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison between ReluAdapter and LinearAdapter. Across different adapter scaling factors sa and bottleneck dimensions ˆd on the FGVC-Aircraft, LinearAdapter demonstrates superior overall accuracy. yield benefits. Furthermore, [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: A simple modification in PyTorch pseudo-code. Discussion. Further analysis reveals that the effects of hyperparameters differ between seen and novel categories. We find that variations cause only minor fluctuations in seen category accuracy, but have a significant impact on novel category accuracy, as shown in [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fluctuations in predicted bias between seen and novel categories. We track the change of predictions on the entire dataset to examine how the ratio of samples classified as seen versus novel categories changes across epochs. 25 50 75 100 125 150 175 Sorted Class Index 0 1 2 3 4 5 6 7 8 Predicted Distribution 1e 3 Stanford Cars w/o DA w DA Prior [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of predicted distributions without and with distribution alignment (DA). The x-axis indicates sorted class indices, with seen categories in the first half and novel categories in the second half. 4.3 Debiasing with Distribution Alignment GCD methods often suffer from severe bias arising from the asymmetric learning paces between seen and novel categories [12]. As shown in [PITH_FULL_IMAGE:figur… view at source ↗
Figure 10
Figure 10. Figure 10: Sensitivity of the scale sa of adapter. 16 32 64 96 128 Bottleneck Dimension 50 55 60 65 70 Accuracy (%) FGVC-Aircraft All Seen Novel 16 32 64 96 128 Bottleneck Dimension 50 60 70 80 90 Accuracy (%) Stanford Cars All Seen Novel [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Sensitivity of the number n of adapted blocks. 0.1 0.2 0.3 0.4 Alignment Scaling Factor 52 54 56 58 60 62 64 66 68 Accuracy (%) FGVC-Aircraft All Seen Novel 0.1 0.2 0.3 0.4 Alignment Scaling Factor 50 55 60 65 70 75 80 85 90 Accuracy (%) Stanford Cars All Seen Novel [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: FGVC-Aircraft Novel Categories. DINO ReLU Linear Seen Sample 1 Seen Sample 2 Original Image [PITH_FULL_IMAGE:figures/full_fig_p010_14.png] view at source ↗
Figure 16
Figure 16. Figure 16: Stanford Cars Novel Categories. DINO ReLU Linear Seen Sample 1 Seen Sample 2 Original Image [PITH_FULL_IMAGE:figures/full_fig_p010_16.png] view at source ↗
read the original abstract

Generalized Category Discovery (GCD) seeks to identify novel categories from unlabeled data while retaining the classification ability of seen categories. Prior GCD methods commonly leverage transferable representations from pre-trained models, adapting to downstream datasets via partial fine-tuning (updating only the final ViT block) and visual prompt tuning (appending learnable vectors to inputs). However, conventional partial fine-tuning offers limited flexibility, as it fails to adapt the entire model; meanwhile, visual prompt tuning is prone to overfitting, due to its sensitivity to initialization and inherently constrained capacity. To address these limitations, we propose LAGCD, a simple yet effective GCD approach that embeds a residual linear adapter into each ViT block. From the perspective of feature sparsity, we systematically show that non-linearity in conventional adapters impairs performance, whereas our linear adapter enhances it by enabling more flexible model capacity. We further introduce an auxiliary distribution alignment loss to mitigate the negative impact of biased predictions between seen and novel categories. Extensive experiments on both generic and fine-grained datasets confirm that LAGCD consistently improves performance over many sophisticated baselines. The source code is available at https://github.com/yebo0216best/LAGCD

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes LAGCD for Generalized Category Discovery, inserting a residual linear adapter into each ViT block. It claims that non-linearity in conventional adapters impairs performance due to increased feature sparsity, while the linear adapter improves results by enabling greater model flexibility. An auxiliary distribution alignment loss is added to mitigate biased predictions between seen and novel classes. Experiments on generic and fine-grained datasets report consistent gains over prior GCD baselines.

Significance. If the sparsity mechanism is causally validated, the work offers a simple alternative to partial fine-tuning and prompt tuning in GCD, potentially simplifying adapter design in transfer learning. Releasing source code aids reproducibility, but the absence of isolated mechanism tests limits the strength of the contribution.

major comments (2)
  1. [Analysis / Experiments] The central claim that non-linearity impairs performance via feature sparsity (abstract and analysis sections) requires explicit quantification of sparsity differences, such as L0 norms or activation histograms, between linear and non-linear adapters; end-to-end accuracy tables alone do not establish this as the load-bearing mechanism.
  2. [Experiments] The comparison between linear and non-linear adapters is potentially confounded by the auxiliary distribution alignment loss; a full factorial ablation (adapter type × with/without aux loss) is needed to isolate causality, as noted in the stress-test concern.
minor comments (1)
  1. [Abstract] The abstract states empirical improvements without referencing specific tables, datasets, or statistical tests; adding these cross-references would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for the constructive and detailed comments. We address each major comment below and commit to revisions that strengthen the mechanistic evidence and experimental isolation as requested.

read point-by-point responses

  1. Referee: [Analysis / Experiments] The central claim that non-linearity impairs performance via feature sparsity (abstract and analysis sections) requires explicit quantification of sparsity differences, such as L0 norms or activation histograms, between linear and non-linear adapters; end-to-end accuracy tables alone do not establish this as the load-bearing mechanism.

    Authors: We agree that explicit quantification is required to establish the sparsity mechanism as load-bearing. The manuscript demonstrates the performance advantage of linear adapters and discusses sparsity qualitatively, but we acknowledge that end-to-end tables and existing analysis do not provide direct metrics. In the revised manuscript we will add L0-norm statistics of post-adapter activations together with activation histograms comparing linear versus non-linear adapters across multiple ViT blocks and datasets. These additions will directly quantify the sparsity increase induced by non-linearity. revision: yes

  2. Referee: [Experiments] The comparison between linear and non-linear adapters is potentially confounded by the auxiliary distribution alignment loss; a full factorial ablation (adapter type × with/without aux loss) is needed to isolate causality, as noted in the stress-test concern.

    Authors: We appreciate the concern about potential confounding. The distribution alignment loss targets prediction bias between seen and novel classes and is intended to be complementary to the adapter choice. To isolate effects rigorously, the revised manuscript will include a complete factorial ablation: linear and non-linear adapters, each evaluated both with and without the auxiliary loss, on the primary generic and fine-grained benchmarks. Results will be presented in a dedicated table to confirm that the linear-adapter benefit holds independently of the loss. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes an empirical method (LAGCD) inserting a residual linear adapter per ViT block plus an auxiliary distribution alignment loss, then reports accuracy gains on GCD benchmarks. The 'systematic show' from the feature-sparsity perspective is an interpretive post-experiment explanation rather than a closed mathematical derivation or prediction that reduces to fitted inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing premises; the auxiliary loss is introduced separately to address bias. Experimental validation on external datasets supplies independent content, so the central claims remain non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the standard assumption that pre-trained ViT representations are transferable to GCD and that feature sparsity is the dominant factor determining adapter effectiveness; no new entities or fitted constants are introduced in the abstract.

axioms (1)
  • domain assumption Pre-trained Vision Transformer models yield transferable representations that can be adapted for generalized category discovery.
    Invoked as the starting point for all adaptation methods discussed.

pith-pipeline@v0.9.0 · 5512 in / 1264 out tokens · 49569 ms · 2026-05-12T01:08:56.670805+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 5 internal anchors

  1. [1]

    Parametric classification for gen- eralized category discovery: A baseline study,

    X. Wen, B. Zhao, and X. Qi, “Parametric classification for gen- eralized category discovery: A baseline study,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 16 590–16 600

    work page 2023

  2. [2]

    Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning.arXiv preprint arXiv:2403.13684, 2024

    H. Wang, S. Vaze, and K. Han, “Sptnet: An efficient alterna- tive framework for generalized category discovery with spatial prompt tuning,”arXiv preprint arXiv:2403.13684, 2024

    work page arXiv 2024

  3. [3]

    Adaptgcd: Multi-expert adapter tuning for generalized category discovery,

    Y. Qu, Y. Tang, C. Zhang, and W. Zhang, “Adaptgcd: Multi-expert adapter tuning for generalized category discovery,”arXiv preprint arXiv:2410.21705, 2024

    work page arXiv 2024

  4. [4]

    Pseudo-label: The simple and efficient semi- supervised learning method for deep neural networks,

    D.-H. Lee, “Pseudo-label: The simple and efficient semi- supervised learning method for deep neural networks,” inWork- shop on challenges in representation learning, ICML, 2013

    work page 2013

  5. [5]

    Regularization with stochastic transformations and perturbations for deep semi- supervised learning,

    M. Sajjadi, M. Javanmardi, and T. Tasdizen, “Regularization with stochastic transformations and perturbations for deep semi- supervised learning,” inNeural Information Processing Systems, 2016

    work page 2016

  6. [6]

    Temporal Ensembling for Semi-Supervised Learning,

    S. Laine and T. Aila, “Temporal Ensembling for Semi-Supervised Learning,”arXiv: Neural and Evolutionary Computing, 2016

    work page 2016

  7. [7]

    FixMatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, C.-L. Li, Z. Zhang, N. Carlini, E. D. Cubuk, A. Kurakin, H. Zhang, and C. Raffel, “FixMatch: Simplifying semi- supervised learning with consistency and confidence,” inNeural Information Processing Systems, 2020

    work page 2020

  8. [8]

    Towards realistic long-tailed semi-supervised learning: Consistency is all you need,

    T. Wei and K. Gan, “Towards realistic long-tailed semi-supervised learning: Consistency is all you need,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 3469–3478

    work page 2023

  9. [9]

    Erasing the bias: Fine-tuning foundation models for semi-supervised learning,

    K. Gan and T. Wei, “Erasing the bias: Fine-tuning foundation models for semi-supervised learning,” inForty-first International Conference on Machine Learning, 2024

    work page 2024

  10. [10]

    Semi-supervised clip adaptation by enforcing semantic and trapezoidal consistency,

    K. Gan, B. Ye, M.-L. Zhang, and T. Wei, “Semi-supervised clip adaptation by enforcing semantic and trapezoidal consistency,” in The Thirteenth International Conference on Learning Representations

    work page

  11. [11]

    Realistic evaluation of deep semi-supervised learning algorithms,

    A. Oliver, A. Odena, C. A. Raffel, E. D. Cubuk, and I. Good- fellow, “Realistic evaluation of deep semi-supervised learning algorithms,” inNeural Information Processing Systems, 2018

    work page 2018

  12. [12]

    Open-world semi-supervised learning,

    K. Cao, M. Brbic, and J. Leskovec, “Open-world semi-supervised learning,” inInternational Conference on Learning Representations, 2022

    work page 2022

  13. [13]

    Generalized category discovery,

    S. Vaze, K. Han, A. Vedaldi, and A. Zisserman, “Generalized category discovery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7492–7501

    work page 2022

  14. [14]

    Promptcal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery,

    S. Zhang, S. Khan, Z. Shen, M. Naseer, G. Chen, and F. S. Khan, “Promptcal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3479–3488

    work page 2023

  15. [15]

    A graph-theoretic framework for understanding open-world semi-supervised learning,

    Y. Sun, Z. Shi, and Y. Li, “A graph-theoretic framework for understanding open-world semi-supervised learning,”Advances in Neural Information Processing Systems, vol. 36, pp. 23 934–23 967, 2023

    work page 2023

  16. [16]

    No representation rules them all in category discovery,

    S. Vaze, A. Vedaldi, and A. Zisserman, “No representation rules them all in category discovery,”Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  17. [17]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  18. [18]

    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 confer- ence on computer vision, 2021, pp. 9650–9660

    work page 2021

  19. [19]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 709–727

    work page 2022

  20. [20]

    Parameter- efficient transfer learning for nlp,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Larous- silhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter- efficient transfer learning for nlp,” inInternational conference on machine learning. PMLR, 2019, pp. 2790–2799

    work page 2019

  21. [21]

    LoRA: Low-Rank Adaptation of Large Language Models

    E. J. Hu, Y. Shen, P . Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,”arXiv preprint arXiv:2106.09685, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  22. [22]

    Adaptformer: Adapting vision transformers for scalable visual recognition,

    S. Chen, C. Ge, Z. Tong, J. Wang, Y. Song, J. Wang, and P . Luo, “Adaptformer: Adapting vision transformers for scalable visual recognition,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 664–16 678, 2022

    work page 2022

  23. [23]

    Rectified linear units improve restricted boltzmann machines,

    V . Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines,” inProceedings of the 27th international confer- ence on machine learning (ICML-10), 2010, pp. 807–814

    work page 2010

  24. [24]

    Xcon: Learning with experts for fine-grained category discovery,

    Y. Fei, Z. Zhao, S. Yang, and B. Zhao, “Xcon: Learning with experts for fine-grained category discovery,”arXiv preprint arXiv:2208.01898, 2022

    work page arXiv 2022

  25. [25]

    Opencon: Open-world contrastive learning,

    Y. Sun and Y. Li, “Opencon: Open-world contrastive learning,” arXiv preprint arXiv:2208.02764, 2022

    work page arXiv 2022

  26. [26]

    Contrastive mean-shift learning for generalized category discovery,

    S. Choi, D. Kang, and M. Cho, “Contrastive mean-shift learning for generalized category discovery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 094–23 104

    work page 2024

  27. [27]

    Solving the catastrophic forgetting problem in general- ized category discovery,

    X. Cao, X. Zheng, G. Wang, W. Yu, Y. Shen, K. Li, Y. Lu, and Y. Tian, “Solving the catastrophic forgetting problem in general- ized category discovery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 880–16 889

    work page 2024

  28. [28]

    Flipped classroom: Aligning teacher attention with student in generalized category discovery,

    H. Lin, W. An, J. Wang, Y. Chen, F. Tian, M. Wang, Q. Wang, G. Dai, and J. Wang, “Flipped classroom: Aligning teacher attention with student in generalized category discovery,”Advances in Neural Information Processing Systems, vol. 37, pp. 60 897–60 935, 2024

    work page 2024

  29. [29]

    Hilo: A learning framework for generalized category discovery robust to domain shifts,

    H. Wang, S. Vaze, and K. Han, “Hilo: A learning framework for generalized category discovery robust to domain shifts,”arXiv preprint arXiv:2408.04591, 2024

    work page arXiv 2024

  30. [30]

    Bridging the gap: Learning pace synchronization for open-world semi-supervised learning,

    B. Ye, K. Gan, T. Wei, and M.-L. Zhang, “Bridging the gap: Learning pace synchronization for open-world semi-supervised learning,” inProceedings of the Thirty-Third International Joint Con- ference on Artificial Intelligence, 2024, pp. 5362–5370

    work page 2024

  31. [31]

    Long- tail learning with foundation model: Heavy fine-tuning hurts,

    J.-X. Shi, T. Wei, Z. Zhou, J.-J. Shao, X.-Y. Han, and Y.-F. Li, “Long- tail learning with foundation model: Heavy fine-tuning hurts,” in Proceedings of the 41st International Conference on Machine Learning, 2024

    work page 2024

  32. [32]

    Supervised contrastive learning,

    P . Khosla, P . Teterwak, C. Wang, A. Sarna, Y. Tian, P . Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,”Advances in neural information processing systems, vol. 33, pp. 18 661–18 673, 2020

    work page 2020

  33. [33]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning. PMLR, 2020, pp. 1597– 1607

    work page 2020

  34. [34]

    Masked siamese networks for label-efficient learning,

    M. Assran, M. Caron, I. Misra, P . Bojanowski, F. Bordes, P . Vincent, A. Joulin, M. Rabbat, and N. Ballas, “Masked siamese networks for label-efficient learning,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 456–473. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, AUGUST XX 13

    work page 2022

  35. [35]

    Pseudo-labeling and confirmation bias in deep semi-supervised learning,

    E. Arazo, D. Ortego, P . Albert, N. E. OConnor, and K. McGuinness, “Pseudo-labeling and confirmation bias in deep semi-supervised learning,” inInternational Joint Conference on Neural Networks, 2020

    work page 2020

  36. [36]

    Empirical evalua- tion of rectified activations in convolutional network

    B. Xu, “Empirical evaluation of rectified activations in convolu- tional network,”arXiv preprint arXiv:1505.00853, 2015

    work page arXiv 2015

  37. [37]

    Rectifier nonlinearities improve neural network acoustic models,

    A. L. Maas, A. Y. Hannun, A. Y. Nget al., “Rectifier nonlinearities improve neural network acoustic models,” inProc. icml, vol. 30, no. 1. Atlanta, GA, 2013, p. 3

    work page 2013

  38. [38]

    Zero-bias autoencoders and the benefits of co-adapting features

    K. Konda, R. Memisevic, and D. Krueger, “Zero-bias autoen- coders and the benefits of co-adapting features,”arXiv preprint arXiv:1402.3337, 2014

    work page arXiv 2014

  39. [39]

    Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

    D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (elus),”arXiv preprint arXiv:1511.07289, vol. 4, no. 5, p. 11, 2015

    work page Pith review arXiv 2015

  40. [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

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  41. [41]

    3d object represen- tations for fine-grained categorization,

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object represen- tations for fine-grained categorization,” inProceedings of the IEEE international conference on computer vision workshops, 2013, pp. 554– 561

    work page 2013

  42. [42]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

    work page 2009

  43. [43]

    Contrastive multiview coding,

    Y. Tian, D. Krishnan, and P . Isola, “Contrastive multiview coding,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16. Springer, 2020, pp. 776–794

    work page 2020

  44. [44]

    Caltech-ucsd birds 200,

    P . Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P . Perona, “Caltech-ucsd birds 200,” 2010

    work page 2010

  45. [45]

    Open-set recogni- tion: A good closed-set classifier is all you need?

    S. Vaze, K. Han, A. Vedaldi, and A. Zisserman, “Open-set recogni- tion: A good closed-set classifier is all you need?” 2021

    work page 2021

  46. [46]

    The herbarium challenge 2019 dataset,

    K. C. Tan, Y. Liu, B. Ambrose, M. Tulig, and S. Be- longie, “The herbarium challenge 2019 dataset,”arXiv preprint arXiv:1906.05372, 2019

    work page arXiv 2019

  47. [47]

    The Hungarian method for the assignment prob- lem,

    H. W. Kuhn, “The Hungarian method for the assignment prob- lem,”Naval Research Logistics Quarterly, 1955

    work page 1955

  48. [48]

    k-means++: The advantages of careful seeding,

    D. Arthur and S. Vassilvitskii, “k-means++: The advantages of careful seeding,” Stanford, Tech. Rep., 2006

    work page 2006

  49. [49]

    Autonovel: Automatically discovering and learning novel visual categories,

    K. Han, S.-A. Rebuffi, S. Ehrhardt, A. Vedaldi, and A. Zisserman, “Autonovel: Automatically discovering and learning novel visual categories,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6767–6781, 2021

    work page 2021

  50. [50]

    A unified objective for novel class discovery,

    E. Fini, E. Sangineto, S. Lathuili `ere, Z. Zhong, M. Nabi, and E. Ricci, “A unified objective for novel class discovery,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9284–9292

    work page 2021

  51. [51]

    Dynamic conceptional contrastive learning for generalized category discovery,

    N. Pu, Z. Zhong, and N. Sebe, “Dynamic conceptional contrastive learning for generalized category discovery,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7579–7588

    work page 2023

  52. [52]

    Learning semi-supervised gaussian mixture models for generalized category discovery,

    B. Zhao, X. Wen, and K. Han, “Learning semi-supervised gaussian mixture models for generalized category discovery,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 16 623–16 633

    work page 2023

  53. [53]

    Amend: Adap- tive margin and expanded neighborhood for efficient generalized category discovery,

    A. Banerjee, L. S. Kallooriyakath, and S. Biswas, “Amend: Adap- tive margin and expanded neighborhood for efficient generalized category discovery,” inProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, 2024, pp. 2101–2110

    work page 2024

  54. [54]

    Protogcd: Unified and unbiased prototype learning for generalized category discovery,

    S. Ma, F. Zhu, X.-Y. Zhang, and C.-L. Liu, “Protogcd: Unified and unbiased prototype learning for generalized category discovery,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

  55. [55]

    The perceptron: a probabilistic model for informa- tion storage and organization in the brain

    F. Rosenblatt, “The perceptron: a probabilistic model for informa- tion storage and organization in the brain.”Psychological review, vol. 65, no. 6, p. 386, 1958

    work page 1958

  56. [56]

    Efficient back- prop,

    Y. LeCun, L. Bottou, G. B. Orr, and K.-R. M ¨uller, “Efficient back- prop,” inNeural networks: Tricks of the trade. Springer, 2002, pp. 9–50

    work page 2002

  57. [57]

    Searching for Activation Functions

    P . Ramachandran, B. Zoph, and Q. V . Le, “Searching for activation functions,”arXiv preprint arXiv:1710.05941, 2017

    work page internal anchor Pith review arXiv 2017

  58. [58]

    Gaussian Error Linear Units (GELUs)

    D. Hendrycks and K. Gimpel, “Gaussian error linear units (gelus),”arXiv preprint arXiv:1606.08415, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016