arxiv: 2604.15678 · v1 · submitted 2026-04-17 · 💻 cs.CV

Recognition: unknown

HyCal: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning

Eunju Lee , MiHyeon Kim , JuneHyoung Kwon , Yoonji Lee , JiHyun Kim , Soojin Jang , YoungBin Kim

Authors on Pith no claims yet

Pith reviewed 2026-05-10 09:10 UTC · model grok-4.3

classification 💻 cs.CV

keywords few-shot class-incremental learningprototype calibrationdomain gravitycross-discipline learningCLIP embeddingstraining-free methodcontinual learningimbalanced data

0 comments

The pith

A training-free method blends cosine similarity and Mahalanobis distance on frozen CLIP embeddings to stabilize prototypes against domain imbalance in few-shot continual learning.

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

The paper establishes that data imbalance across heterogeneous disciplines creates a representational pull called Domain Gravity, where overrepresented domains distort class prototypes and degrade performance on underrepresented ones. It introduces the XD-VSCIL benchmark to capture this real-world heterogeneity and proposes HyCal as a calibration step that requires no training or parameter updates. By combining directional alignment from cosine similarity with covariance-aware scaling from Mahalanobis distance, the approach keeps prototypes stable on pre-trained embeddings. This matters because standard few-shot incremental methods assume uniform domains and balanced samples, conditions that rarely hold outside controlled benchmarks.

Core claim

HyCal, operating on frozen CLIP embeddings, combines cosine similarity and Mahalanobis distance to capture complementary geometric properties-directional alignment and covariance-aware magnitude-yielding stable prototypes under imbalanced heterogeneous conditions and mitigating Domain Gravity in cross-discipline variable few-shot class-incremental learning.

What carries the argument

Hybrid Prototype Calibration (HyCal), which blends cosine similarity for directional alignment with Mahalanobis distance for magnitude adjustment on frozen embeddings to counteract prototype drift.

If this is right

Existing FSCIL methods that assume homogeneous domains and balanced distributions become limited when applied to real-world cross-discipline data.
Training-free calibration on frozen embeddings preserves efficiency while still delivering retention and adaptation gains.
Prototype drift from overrepresented low-entropy domains can be reduced without retraining the underlying model.
The XD-VSCIL benchmark makes it possible to measure how well methods handle naturally occurring imbalance across disciplines.

Where Pith is reading between the lines

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

The same calibration principle might apply to other pre-trained embedding models if the two distance measures continue to provide complementary information.
Domains with very different visual entropy levels could be explicitly weighted during calibration to further reduce gravity effects.
Future continual-learning benchmarks should report domain-imbalance statistics alongside accuracy to reflect the conditions the paper identifies.
Combining HyCal with lightweight memory replay could test whether the training-free property holds when some adaptation is reintroduced.

Load-bearing premise

That combining cosine similarity and Mahalanobis distance on frozen embeddings is enough to capture the geometric properties needed to stabilize prototypes in every kind of heterogeneous and imbalanced setting.

What would settle it

Running HyCal on an extreme-imbalance test set where one domain supplies 90 percent of samples while another supplies 5 percent and checking whether accuracy on the minority domain still exceeds that of standard prototype-based FSCIL baselines.

Figures

Figures reproduced from arXiv: 2604.15678 by Eunju Lee, JiHyun Kim, JuneHyoung Kwon, MiHyeon Kim, Soojin Jang, Yoonji Lee, YoungBin Kim.

**Figure 1.** Figure 1: Overview of continual learning paradigms. (a) CIL: Class-incremental learning without domain shift. (b) FSCIL: Fixed shots per [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: t-SNE visualization showing prototype embeddings. Prototypes for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of three representative dataset distributions used in XD-VSCIL. (a)–(c): training-sample distributions under (a) highly [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Overview of HYCAL, our training-free prototype calibration method. For each XD-VSCIL task, class prototypes are constructed from frozen CLIP embeddings using a few sample images during training. At inference, HYCAL combines cosine similarity and Mahalanobis distance to compute test-to-prototype scores, enabling robust classification under domain shift and data imbalance. parameters, and no model expansion… view at source ↗

**Figure 5.** Figure 5: Task-per accuracy across incremental steps under the (a) Balanced-in-class domain and (b) Cross-scale imbalance settings. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.** Figure 6: Cosine–Mahalanobis: (a) Relationship and (b) Ranking where cosine-correct (green) and Mahalanobis-correct (red) samples [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Qualitative comparison of distance metrics. Pink points show samples only correctly classified by dynamic summation. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

**Figure 8.** Figure 8: Hyperparameter sensitivity analysis for α and β. Each curve shows the performance variation when sweeping one parameter while keeping the other fixed [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

**Figure 9.** Figure 9: Hyperparameter sensitivity analysis for λ and γ under the High-scale domain imbalance setting. Heatmaps show Last Acc., Avg Acc., and SCDE over different (λ, γ) configurations. 8.1. Robustness to domain order To examine the robustness of each method under variations in domain sequences, we evaluate all approaches across four random domain orders, derived from fixed random seeds \ifmmode \lbrace \else \text… view at source ↗

read the original abstract

Pretrained Vision-Language Models (VLMs) like CLIP show promise in continual learning, but existing Few-Shot Class-Incremental Learning (FSCIL) methods assume homogeneous domains and balanced data distributions, limiting real-world applicability where data arises from heterogeneous disciplines with imbalanced sample availability and varying visual complexity. We identify Domain Gravity, a representational asymmetry where data imbalance across heterogeneous domains causes overrepresented or low-entropy domains to disproportionately influence the embedding space, leading to prototype drift and degraded performance on underrepresented or high-entropy domains. To address this, we introduce Cross-Discipline Variable Few-Shot Class-Incremental Learning (XD-VSCIL), a benchmark capturing real-world heterogeneity and imbalance where Domain Gravity naturally intensifies. We propose Hybrid Prototype Calibration (HyCal), a training-free method combining cosine similarity and Mahalanobis distance to capture complementary geometric properties-directional alignment and covariance-aware magnitude-yielding stable prototypes under imbalanced heterogeneous conditions. Operating on frozen CLIP embeddings, HyCal achieves consistent retention-adaptation improvements while maintaining efficiency. Experiments show HyCal effectively mitigates Domain Gravity and outperforms existing methods in imbalanced cross-domain incremental learning.

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.

Desk Editor's Note private letter to a colleague

HyCal names Domain Gravity and offers a training-free cosine-plus-Mahalanobis fix on frozen CLIP, but the abstract gives no numbers or ablations so the practical gain is still unshown.

read the letter

The paper's main move is to define Domain Gravity as the skew that imbalanced cross-discipline data creates in VLM embedding spaces, then release the XD-VSCIL benchmark to make that setting testable. On top of that they give HyCal, a training-free prototype calibration that blends cosine similarity with Mahalanobis distance so the prototypes stay stable without any fine-tuning. That combination is the only concrete technical step they describe. It is new in the FSCIL literature and the benchmark itself could be useful for anyone who needs heterogeneous, imbalanced streams rather than the usual balanced single-domain splits. The training-free claim is also honest and keeps compute low, which matters for real deployment. The soft spot is that none of the performance claims are backed by numbers, baselines, or ablations in what we have. Without those it is impossible to tell whether the two distances really complement each other or whether the Mahalanobis term simply adds noise when few-shot covariance estimates are unstable. The stress-test worry about singular covariances in high-entropy domains is not answered here. This is for people already working on continual learning with pretrained vision-language models who want a new testbed for cross-domain imbalance. The benchmark and the framing are worth a look even if the method still needs proof. I would send it to referees so the experiments can be checked properly rather than desk-rejecting on the abstract alone.

Referee Report

2 major / 2 minor

Summary. The paper identifies 'Domain Gravity' as a representational asymmetry in cross-discipline few-shot class-incremental learning (XD-VSCIL) with pretrained VLMs such as CLIP, where data imbalance across heterogeneous domains causes prototype drift. It introduces the XD-VSCIL benchmark to capture real-world heterogeneity and imbalance, and proposes HyCal, a training-free prototype calibration method that combines cosine similarity (for directional alignment) and Mahalanobis distance (for covariance-aware magnitude) on frozen CLIP embeddings to yield stable prototypes. The central claim is that HyCal mitigates Domain Gravity, achieves consistent retention-adaptation improvements, maintains efficiency, and outperforms existing FSCIL methods in imbalanced cross-domain settings.

Significance. If the empirical claims hold, the work addresses a practically relevant gap in continual learning by providing an efficient, parameter-free approach for heterogeneous, imbalanced data without requiring model updates or additional training. The XD-VSCIL benchmark could serve as a useful testbed for future methods. However, the significance is tempered by the paper's introduction of the core phenomenon and benchmark, which creates dependence on the authors' framing, and by the absence of demonstrated robustness of the hybrid metric under violated embedding assumptions.

major comments (2)

[Abstract] Abstract: the claim of outperformance and effective mitigation of Domain Gravity is asserted without any quantitative results, baseline comparisons, statistical tests, ablation studies, or dataset details, rendering it impossible to evaluate whether the data supports the central claims.
[Method] Method description (hybrid distance formulation): the argument that cosine similarity and Mahalanobis distance capture complementary orthogonal geometric properties sufficient to neutralize Domain Gravity without any adaptation rests on the unverified assumption that frozen CLIP embeddings encode reliable second-order statistics across all disciplines and imbalance levels. No derivation, stability analysis, or ablation is provided to show the hybrid distance remains stable (rather than increasing prototype drift) when high-entropy domains yield near-singular covariances or few-shot estimates are noisy.

minor comments (2)

[§3] Clarify the precise combination rule for the two metrics (e.g., weighted sum, product, or other) and any hyperparameters involved, even if claimed to be training-free.
[§4] Provide the exact definition and construction of the XD-VSCIL benchmark, including how domains, imbalance ratios, and class splits are chosen, to allow reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comments point by point below, indicating where revisions will be made to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of outperformance and effective mitigation of Domain Gravity is asserted without any quantitative results, baseline comparisons, statistical tests, ablation studies, or dataset details, rendering it impossible to evaluate whether the data supports the central claims.

Authors: We agree that the abstract, being a concise overview, does not include specific quantitative results or details. The full manuscript contains these elements in the experiments section, including baseline comparisons, ablations, and XD-VSCIL dataset descriptions. We will revise the abstract to incorporate key quantitative highlights from our results (e.g., retention-adaptation gains) to better support the claims while maintaining brevity. revision: yes
Referee: [Method] Method description (hybrid distance formulation): the argument that cosine similarity and Mahalanobis distance capture complementary orthogonal geometric properties sufficient to neutralize Domain Gravity without any adaptation rests on the unverified assumption that frozen CLIP embeddings encode reliable second-order statistics across all disciplines and imbalance levels. No derivation, stability analysis, or ablation is provided to show the hybrid distance remains stable (rather than increasing prototype drift) when high-entropy domains yield near-singular covariances or few-shot estimates are noisy.

Authors: The hybrid metric is motivated by the complementary nature of directional (cosine) and covariance-aware (Mahalanobis) distances to counter prototype drift from domain imbalance, as motivated in the method section. We do not provide a formal derivation or explicit stability analysis for edge cases like near-singular covariances. Our empirical results across disciplines support stability, but we acknowledge the gap and will add a discussion on covariance regularization (e.g., shrinkage estimators), a brief stability argument, and an ablation on high-entropy/noisy few-shot regimes in the revised manuscript. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected in the claimed derivation.

full rationale

The paper introduces Domain Gravity as an observed representational asymmetry in heterogeneous imbalanced settings and defines the XD-VSCIL benchmark to study it, then proposes the training-free HyCal combination of cosine similarity and Mahalanobis distance on frozen CLIP embeddings. No equations, first-principles derivations, or predictions are shown that reduce by construction to fitted inputs, self-definitions, or self-citation chains. The complementarity of the two metrics is presented as a design choice whose efficacy is evaluated empirically on the new benchmark rather than asserted tautologically. This is a standard non-circular contribution pattern for a new problem formulation and method.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that CLIP embeddings are sufficiently rich for cross-domain prototype stability and that the two distance metrics are complementary without further justification or external validation provided in the abstract.

axioms (1)

domain assumption Pretrained VLMs such as CLIP produce embeddings that remain useful for incremental learning across heterogeneous domains when frozen.
The entire method operates exclusively on these frozen embeddings.

invented entities (1)

Domain Gravity no independent evidence
purpose: To name and explain representational asymmetry caused by data imbalance across disciplines.
Newly introduced concept used to motivate the problem and the need for HyCal.

pith-pipeline@v0.9.0 · 5529 in / 1333 out tokens · 37863 ms · 2026-05-10T09:10:55.470516+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 4 canonical work pages · 1 internal anchor

[1]

Distributions of angles in random packing on spheres.The Journal of Ma- chine Learning Research, 14(1):1837–1864, 2013

Tony Cai, Jianqing Fan, and Tiefeng Jiang. Distributions of angles in random packing on spheres.The Journal of Ma- chine Learning Research, 14(1):1837–1864, 2013. 5

2013
[2]

Online continual learning from imbalanced data

Aristotelis Chrysakis and Marie-Francine Moens. Online continual learning from imbalanced data. InInt. Conf. Mach. Learn.JMLR.org, 2020. 1

2020
[3]

Describing textures in the wild

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. InIEEE Conf. Comput. Vis. Pattern Recog., pages 3606–3613, 2014. 3, 4, 8

2014
[4]

The mnist database of handwritten digit images for machine learning research [best of the web].IEEE Signal Processing Magazine, 29(6):141–142, 2012

Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web].IEEE Signal Processing Magazine, 29(6):141–142, 2012. 4, 9

2012
[5]

Clip-adapter: Better vision-language models with feature adapters.Int

Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters.Int. J. Comput. Vis., 132(2):581–595, 2024. 1, 3

2024
[6]

Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning.Adv

Dipam Goswami, Yuyang Liu, Bartłomiej Twardowski, and Joost Van De Weijer. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning.Adv. Neural Inform. Process. Syst., 36:6582–6595, 2023. 1, 2, 3, 6, 7, 8, 5, 10

2023
[7]

Calibrating higher-order statistics for few-shot class-incremental learning with pre-trained vision transform- ers

Dipam Goswami, Bartłomiej Twardowski, and Joost Van De Weijer. Calibrating higher-order statistics for few-shot class-incremental learning with pre-trained vision transform- ers. InIEEE Conf. Comput. Vis. Pattern Recog., pages 4075– 4084, 2024. 1, 2, 3

2024
[8]

Informa- tion retrieval optimization for non-exemplar class incremen- tal learning

Shuai Guo, Yang Gu, Yuan Ma, Yingwei Zhang, Weining Weng, Jun Liu, Weiwei Dai, and Yiqiang Chen. Informa- tion retrieval optimization for non-exemplar class incremen- tal learning. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 717–726, 2024. 1

2024
[9]

Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 4, 8

2019
[10]

Dynamically anchored prompting for task-imbalanced continual learning

Chenxing Hong, Yan Jin, Zhiqi Kang, Yizhou Chen, Mengke Li, Yang Lu, and Hanzi Wang. Dynamically anchored prompting for task-imbalanced continual learning. InIJCAI, pages 4127–4135, 2024. 3

2024
[11]

Online continual learning via logit adjusted softmax.Trans

Zhehao Huang, Tao Li, Chenhe Yuan, Yingwen Wu, and Xi- aolin Huang. Online continual learning via logit adjusted softmax.Trans. Mach. Learn. Res., 2024, 2024. 1

2024
[12]

Open- clip, 2021

Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Han- naneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Open- clip, 2021. 3

2021
[13]

Deep learning of multi- element abundances from high-resolution spectroscopic data.Monthly Notices of the Royal Astronomical Society, 483(3):3255–3277, 2018

Henry W Leung and Jo Bovy. Deep learning of multi- element abundances from high-resolution spectroscopic data.Monthly Notices of the Royal Astronomical Society, 483(3):3255–3277, 2018. 4, 9

2018
[14]

The double-ellipsoid ge- ometry of clip

Meir Yossef Levi and Guy Gilboa. The double-ellipsoid ge- ometry of clip. InInt. Conf. Learn. Represent., pages 33999– 34019. PMLR, 2025. 2

2025
[15]

Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models

Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. InInt. Conf. Mach. Learn., pages 19730–19742, 2023. 3

2023
[16]

Clms: Bridging domain gaps in medical imaging segmentation with source-free continual learning for robust knowledge transfer and adaptation.Medical Image Analysis, 100:103404, 2025

Weilu Li, Yun Zhang, Hao Zhou, Wenhan Yang, Zhi Xie, and Yao He. Clms: Bridging domain gaps in medical imaging segmentation with source-free continual learning for robust knowledge transfer and adaptation.Medical Image Analysis, 100:103404, 2025. 1

2025
[17]

Hehai Lin, Hui Liu, Shilei Cao, Jing Li, Haoliang Li, and Wenya Wang

Peiyuan Liao, Xiuyu Li, Xihui Liu, and Kurt Keutzer. The artbench dataset: Benchmarking generative models with art- works.arXiv preprint arXiv:2206.11404, 2022. 4, 8

work page arXiv 2022
[18]

Lada: Scalable label-specific clip adapter for con- tinual learning

Mao-Lin Luo, Zi-Hao Zhou, Tong Wei, and Min-Ling Zhang. Lada: Scalable label-specific clip adapter for con- tinual learning. InInt. Conf. Mach. Learn., pages 41604– 41619. PMLR, 2025. 1, 2

2025
[19]

Fine-Grained Visual Classification of Aircraft

Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft.arXiv preprint arXiv:1306.5151, 2013. 4, 8

work page internal anchor Pith review arXiv 2013
[20]

John Wiley & Sons, 2024

Kanti V Mardia, John T Kent, and Charles C Taylor.Multi- variate analysis. John Wiley & Sons, 2024. 5

2024
[21]

Pip: Prototypes-injected prompt for federated class incremental learning

Muhammad Anwar Ma’sum, Mahardhika Pratama, Savitha Ramasamy, Lin Liu, Habibullah Habibullah, and Ryszard Kowalczyk. Pip: Prototypes-injected prompt for federated class incremental learning. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1670–1679, 2024. 1

2024
[22]

Ranpac: Random projections and pre-trained models for continual learning

Mark D McDonnell, Dong Gong, Amin Parvaneh, Ehsan Abbasnejad, and Anton Van den Hengel. Ranpac: Random projections and pre-trained models for continual learning. Adv. Neural Inform. Process. Syst., 36:12022–12053, 2023. 3, 6, 7, 8, 5, 10

2023
[23]

Continual learning using a kernel-based method over foundation mod- els

Saleh Momeni, Sahisnu Mazumder, and Bing Liu. Continual learning using a kernel-based method over foundation mod- els. InAAAI, pages 19528–19536, 2025. 3, 6, 7, 8, 5, 10

2025
[24]

Automated flower classification over a large number of classes

Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. InIn- dian Conference on Computer Vision, Graphics and Image Processing, 2008. 4, 9

2008
[25]

Interpreting the linear structure of vision-language model embedding spaces

Isabel Papadimitriou, Huangyuan Su, Thomas Fel, Sham M Kakade, and Stephanie Gil. Interpreting the linear structure of vision-language model embedding spaces. InSecond Con- ference on Language Modeling, 2025. 6, 2

2025
[26]

Understanding the feature norm for out- of-distribution detection

Jaewoo Park, Jacky Chen Long Chai, Jaeho Yoon, and An- drew Beng Jin Teoh. Understanding the feature norm for out- of-distribution detection. InInt. Conf. Comput. Vis., pages 1557–1567, 2023. 2

2023
[27]

Be- yond semantics: Rediscovering spatial awareness in vision- language models.arXiv preprint arXiv:2503.17349, 2025

Jianing Qi, Jiawei Liu, Hao Tang, and Zhigang Zhu. Be- yond semantics: Rediscovering spatial awareness in vision- language models.arXiv preprint arXiv:2503.17349, 2025. 6, 2

work page arXiv 2025
[28]

Learn- ing transferable visual models from natural language super- vision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InInt. Conf. Mach. Learn., pages 8748–8763. PmLR,
[29]

Liu, Abhijit Guha Roy, Shreyas Padhy, and Balaji Lakshminarayanan

Jie Ren, Stanislav Fort, Jeremiah Liu, Abhijit Guha Roy, Shreyas Padhy, and Balaji Lakshminarayanan. A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022, 2021. 7

work page arXiv 2021
[30]

Mos: Model surgery for pre- trained model-based class-incremental learning

Hai-Long Sun, Da-Wei Zhou, Hanbin Zhao, Le Gan, De- Chuan Zhan, and Han-Jia Ye. Mos: Model surgery for pre- trained model-based class-incremental learning. InAAAI, pages 20699–20707, 2025. 1

2025
[31]

Few-shot out of domain intent detection with covariance corrected ma- halanobis distance.AAAI Workshop on Uncertainty Reason- ing and Quantification in Decision Making, 2023

Jayasimha Talur, Oleg Smirnov, and Paul Missault. Few-shot out of domain intent detection with covariance corrected ma- halanobis distance.AAAI Workshop on Uncertainty Reason- ing and Quantification in Decision Making, 2023. 2

2023
[32]

Few-shot class- incremental learning

Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class- incremental learning. InIEEE Conf. Comput. Vis. Pattern Recog., pages 12183–12192, 2020. 1, 2

2020
[33]

What makes clip more robust to long-tailed pre-training data? a controlled study for transferable in- sights.Adv

Xin Wen, Bingchen Zhao, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. What makes clip more robust to long-tailed pre-training data? a controlled study for transferable in- sights.Adv. Neural Inform. Process. Syst., 37:36567–36601,
[34]

Defying imbalanced forgetting in class incremental learning

Shixiong Xu, Gaofeng Meng, Xing Nie, Bolin Ni, Bin Fan, and Shiming Xiang. Defying imbalanced forgetting in class incremental learning. InAAAI, pages 16211–16219, 2024. 1, 3

2024
[35]

Advancing cross- domain discriminability in continual learning of vision- language models.Adv

Yicheng Xu, Yuxin Chen, Jiahao Nie, Yusong Wang, Huip- ing Zhuang, and Manabu Okumura. Advancing cross- domain discriminability in continual learning of vision- language models.Adv. Neural Inform. Process. Syst., 37: 51552–51576, 2024. 1, 2, 3, 6, 7, 8, 5, 10

2024
[36]

Medmnist clas- sification decathlon: A lightweight automl benchmark for medical image analysis

Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist clas- sification decathlon: A lightweight automl benchmark for medical image analysis. InIEEE 18th International Sympo- sium on Biomedical Imaging (ISBI), pages 191–195, 2021. 3, 4, 9

2021
[37]

Boosting continual learning of vision-language models via mixture-of-experts adapters

Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. InIEEE Conf. Comput. Vis. Pattern Recog., pages 23219– 23230, 2024. 1, 2

2024
[38]

Class-incremental learning: A survey.IEEE Trans

Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De- Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey.IEEE Trans. Pattern Anal. Mach. Intell., 2024. 4

2024
[39]

Learning to prompt for vision-language models.Int

Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.Int. J. Comput. Vis., 130(9):2337–2348, 2022. 3 HYCAL: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning Supplementary Material

2022
[40]

Supplementary on the complementary roles and mutual information between cosine and Mahalanobis measures This supplementary section provides additional analysis supporting the complementary relationship between cosine similarity and Mahalanobis distance, as formalized in The- orem 1 and Theorem 2. While both measures are computed from the same embedding pa...
[41]

Additional ablation study To further evaluate the stability and design choices of HY- CAL, we conduct ablation studies on three components: ro- bustness to domain order, sensitivity to hyperparameters, and the effect of different image–text embedding fusion strategies. These analyses assess whether the method main- tains consistent performance under diffe...

1993
[42]

Efficiency analysis Because HYCALkeeps the pretrained backbone frozen and updates only class prototypes and regularized precision ma- trices for newly introduced classes, its computational cost scales with the number of new classes rather than with full model retraining. Unlike prior approaches that recompute statistics over all classes or domains in XD-V...
[43]

Implementation details We use the Vision Transformer (ViT-B/16) model with a frozen CLIP text encoder for all experi- ments

Detailed experimental setting 10.1. Implementation details We use the Vision Transformer (ViT-B/16) model with a frozen CLIP text encoder for all experi- ments. The model weights are loaded from the openai/clip-vit-base-patch16checkpoint. The image encoder’s parameters are kept frozen throughout all experiments. All experiments were conducted using Py- To...

1993
[44]

The Balanced-in-Class Domain setting results are provided in Tab

Numerical results of Balanced-in-class do- main and Cross-scale imbalance settings For completeness, we provide the numerical values for the Balanced-in-class domain and Cross-scale imbalance set- tings. The Balanced-in-Class Domain setting results are provided in Tab. 12, and the Cross-scale imbalance setting results are summarized in Tab. 13
[45]

Accordingly, its effectiveness depends on the quality of the underlying representation space

Limitations HYCALis designed for settings in which frozen pre- trained representations are already sufficiently informative and where prototype-level calibration is preferable to back- bone adaptation. Accordingly, its effectiveness depends on the quality of the underlying representation space. When newly arriving tasks come from domains that lie far out-...

1900