arxiv: 2511.12090 · v3 · submitted 2025-11-15 · 💻 cs.CV

Teaching Prompts to Coordinate: Hierarchical Layer-Grouped Prompt Tuning for Continual Learning

Shengqin Jiang , Tianqi Kong , Yuankai Qi , Haokui Zhang , Lina Yao , Quan Z. Sheng , Qingshan Liu , Ming-Hsuan Yang This is my paper

Pith reviewed 2026-05-17 22:00 UTC · model grok-4.3

classification 💻 cs.CV

keywords continual learningprompt tuningcatastrophic forgettinglayer groupingroot promptpre-trained modelsvision transformerstask adaptation

0 comments

The pith

Layer-grouped prompts generated from a single root reduce overwriting of prior task features in continual learning.

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

Standard prompt tuning attaches an independent learnable prompt to every layer of a frozen pre-trained model, giving each layer high flexibility to adapt to a new task. The paper shows this independence allows some layers to make changes that erase representations needed for earlier tasks, raising the risk of catastrophic forgetting when all past prompts are combined at inference. The proposed method instead groups layers so they share nearly identical prompts, which differ only by a position encoding term that keeps the original feature relationships inside each group. A single task-specific root prompt then generates the sub-prompts for every group, forcing all of them to be conditioned on the same learned vector and therefore more coordinated. Experiments on four vision benchmarks indicate the resulting stability gains yield favorable accuracy compared with prior prompt-based continual learning approaches.

Core claim

By making layers inside each group share roughly the same prompt adjusted only by position encoding, and by deriving every group sub-prompt from one shared task root prompt, the method coordinates updates across layers so that intrinsic feature pathways of the pre-trained model are preserved while still allowing adaptation to new tasks.

What carries the argument

Hierarchical layer-grouped prompt tuning in which same-group layers share prompts adjusted by position encoding and all sub-prompts are generated from one task-specific root prompt.

If this is right

Prompts inside a layer group remain similar enough to keep the pre-trained propagation pathways intact within that group.
Conditioning every sub-prompt on the same root increases synergy and lowers the chance that one layer will overwrite representations useful for past tasks.
Only a small number of additional parameters per task need to be stored and updated, leaving the backbone frozen.
The method is evaluated on four standard continual-learning vision benchmarks and reports competitive or better results than existing prompt-based approaches.

Where Pith is reading between the lines

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

The same grouping-plus-root idea could be tested on transformer language models to see whether it yields analogous stability in sequential text tasks.
Varying the number of groups or letting the model learn which layers belong together might reveal a better stability-plasticity trade-off than the current fixed grouping.
Combining the root-prompt generator with a small replay buffer of past examples could further reduce forgetting on long task sequences.

Load-bearing premise

Forcing sub-prompts to be generated from a shared root and grouping layers to share prompts will coordinate updates enough to protect previous feature pathways without removing the flexibility needed for new tasks.

What would settle it

Accuracy on earlier tasks falls sharply when a new task is introduced whose required features cut across the fixed layer groupings or when the position-encoding adjustment alone cannot prevent drift in those groups.

Figures

Figures reproduced from arXiv: 2511.12090 by Haokui Zhang, Lina Yao, Ming-Hsuan Yang, Qingshan Liu, Quan Z. Sheng, Shengqin Jiang, Tianqi Kong, Yuankai Qi.

**Figure 2.** Figure 2: Overview of our method for CL. The continual learning framework adapts a pre-trained model through prompt [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Final average accuracy of different methods after each incremental task. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 5.** Figure 5: Comparison of feature visualization of our method [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 4.** Figure 4: Ablation study on the intermediate feature dimension [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

Prompt-based continual learning methods fine-tune only a small set of additional learnable parameters while keeping the pre-trained model's parameters frozen. It enables efficient adaptation to new tasks while mitigating the risk of catastrophic forgetting. These methods typically attach one independent task-specific prompt to each layer of pre-trained models to locally modulate its features, ensuring that the layer's representation aligns with the requirements of the new task. However, although introducing learnable prompts independently at each layer provides high flexibility for adapting to new tasks, this overly flexible tuning could make certain layers susceptible to unnecessary updates. As all prompts till the current task are added together as a final prompt for all seen tasks, the model may easily overwrite feature representations essential to previous tasks, which increases the risk of catastrophic forgetting. To address this issue, we propose a novel hierarchical layer-grouped prompt tuning method for continual learning. It improves model stability in two ways: (i) Layers in the same group share roughly the same prompts, which are adjusted by position encoding. This helps preserve the intrinsic feature relationships and propagation pathways of the pre-trained model within each group. (ii) It utilizes a single task-specific root prompt to learn to generate sub-prompts for each layer group. In this way, all sub-prompts are conditioned on the same root prompt, enhancing their synergy and reducing independence. Extensive experiments across four benchmarks demonstrate that our method achieves favorable performance compared with several state-of-the-art methods.

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

Root prompt plus layer grouping adds a coordination structure to prompt-based continual learning, but thin experimental details leave the stability claims hard to assess.

read the letter

The main thing here is a proposal to coordinate prompts across layers by grouping them to share roughly the same sub-prompts, all generated from one task-specific root prompt and then adjusted by position encoding. This is meant to reduce the overwriting that happens when independent per-layer prompts accumulate and get updated freely. The authors contrast it with prior work that keeps prompts fully independent for maximum flexibility. What is new is this particular hierarchical conditioning and sharing rule; it is not just another per-layer prompt variant. They do a clear job spelling out the forgetting risk from too much independence and then offering two concrete mechanisms—group sharing to hold onto intrinsic pathways and root conditioning to increase synergy among the sub-prompts. That framing is straightforward and directly targets the stability-plasticity tension. The soft spots are in the support. The abstract claims favorable results against SOTA on four benchmarks, yet supplies no ablations on the grouping or root generator, no controls, and no significance numbers. The position-encoding adjustment is presented as the safeguard once sharing is enforced, but there is no analysis showing why a simple additive tweak is enough. The stress-test concern lands: forcing this coupling could over-constrain adaptability for new tasks, and the paper does not demonstrate that the trade-off stays neutral. This is for researchers already working on prompt tuning for vision continual learning. Someone looking for coordination tricks rather than fully independent prompts might find the grouping idea worth trying, but they would need the full experiments and ablations to decide if it actually delivers. I would send it for peer review. The core mechanism engages the literature honestly and is specific enough that referees could usefully pressure-test the empirical claims.

Referee Report

3 major / 2 minor

Summary. The paper proposes a hierarchical layer-grouped prompt tuning method for continual learning. It addresses the risk of catastrophic forgetting in prompt-based CL by (i) grouping layers to share roughly identical prompts that are modulated only by position encoding to preserve intra-group feature relationships and propagation pathways, and (ii) conditioning all group sub-prompts on a single task-specific root prompt to increase synergy and reduce independence among prompts. The abstract reports that this yields favorable performance against SOTA methods on four benchmarks.

Significance. If the empirical claims hold after proper validation, the work provides a concrete mechanism for coordinating prompt updates in CL without full per-layer independence, which could help stabilize prompt-based adaptation of frozen pre-trained models. The explicit linkage of grouping plus root conditioning to preservation of pre-trained pathways is a clear contribution over purely independent prompt tuning baselines.

major comments (3)

[§3.2] §3.2 (position encoding adjustment): the manuscript states that sharing prompts within groups plus additive position encoding 'helps preserve the intrinsic feature relationships and propagation pathways,' yet provides no derivation, analysis, or even a simple forward-pass argument showing why this adjustment is sufficient once sharing is enforced. This is load-bearing for the stability claim.
[§4] §4 (Experiments): the reported favorable results on four benchmarks are presented without ablation studies isolating the contribution of layer grouping versus root-prompt generation, without details on the number of random seeds or statistical significance tests, and without controls for implementation choices such as how the root prompt is optimized. These omissions prevent attribution of gains to the two proposed mechanisms.
[§3.1] §3.1 (root prompt and group-sharing rules): the coordination mechanisms are introduced ad hoc to mitigate the independence problem noted in the introduction, but no analysis quantifies the resulting stability-plasticity trade-off or shows that the added coupling does not erode adaptability for new tasks.

minor comments (2)

[§3] Notation for the root prompt, sub-prompts, and position encoding should be introduced with explicit equations and variable definitions in §3 to improve readability.
Figure captions and axis labels in the experimental figures would benefit from clearer indication of which curves correspond to the proposed method versus baselines.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful comments on our work. We address each of the major comments in detail below, outlining our planned revisions to the manuscript.

read point-by-point responses

Referee: [§3.2] §3.2 (position encoding adjustment): the manuscript states that sharing prompts within groups plus additive position encoding 'helps preserve the intrinsic feature relationships and propagation pathways,' yet provides no derivation, analysis, or even a simple forward-pass argument showing why this adjustment is sufficient once sharing is enforced. This is load-bearing for the stability claim.

Authors: We agree that a more rigorous justification is needed for the position encoding adjustment in §3.2. While the manuscript provides an intuitive explanation, it lacks a formal argument. In the revised manuscript, we will add a simple forward-pass analysis showing how the additive position encoding, combined with shared prompts, helps maintain the pre-trained model's feature propagation within groups. This will strengthen the stability claim. revision: yes
Referee: [§4] §4 (Experiments): the reported favorable results on four benchmarks are presented without ablation studies isolating the contribution of layer grouping versus root-prompt generation, without details on the number of random seeds or statistical significance tests, and without controls for implementation choices such as how the root prompt is optimized. These omissions prevent attribution of gains to the two proposed mechanisms.

Authors: The referee is correct that the experimental section would benefit from more comprehensive ablations and statistical details. We will revise §4 to include ablation studies that isolate the contributions of layer grouping and root-prompt generation. Additionally, we will report results averaged over multiple random seeds and include statistical significance tests. We will also clarify the optimization procedure for the root prompt to allow better attribution of the gains. revision: yes
Referee: [§3.1] §3.1 (root prompt and group-sharing rules): the coordination mechanisms are introduced ad hoc to mitigate the independence problem noted in the introduction, but no analysis quantifies the resulting stability-plasticity trade-off or shows that the added coupling does not erode adaptability for new tasks.

Authors: We acknowledge that the manuscript does not provide a quantitative analysis of the stability-plasticity trade-off resulting from the root prompt and group-sharing. In the revision, we will include an analysis that quantifies this trade-off, for instance by comparing forgetting rates and new task performance with and without the coordination mechanisms, to show that the added coupling improves stability while preserving adaptability. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method proposed by construction with empirical validation

full rationale

The paper introduces a hierarchical layer-grouped prompt tuning approach as a direct architectural solution to the stated problem of independent per-layer prompts causing overwriting of prior features. The claimed stability benefits are attributed explicitly to the design elements (layer grouping with position encoding and root-prompt conditioning), which are defined as part of the method rather than derived via equations or reductions that loop back to fitted inputs. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked in the provided text to justify the core choices. The work is evaluated empirically on four benchmarks, rendering the central contribution self-contained against external results rather than internally forced.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard continual learning assumptions plus two ad-hoc design choices for coordination; no new physical entities or fitted constants are introduced beyond typical prompt-tuning hyperparameters.

axioms (2)

domain assumption Task boundaries are known and prompts can be added sequentially without revisiting old data
Implicit in all prompt-based continual learning setups described in the abstract
ad hoc to paper Position encoding adjustment is sufficient to maintain intra-group feature relationships
Stated as the mechanism that preserves intrinsic pathways within each layer group

pith-pipeline@v0.9.0 · 5585 in / 1332 out tokens · 42882 ms · 2026-05-17T22:00:28.800043+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 1 internal anchor

[1]

Caprompt: Cyclic prompt aggregation for pre-trained model based class incremental learning,

Q. Li and J. Zhou, “Caprompt: Cyclic prompt aggregation for pre-trained model based class incremental learning,” inAAAI Conference on Artificial Intelligence, T. Walsh, J. Shah, and Z. Kolter, Eds., 2025, pp. 18 421–18 429

work page 2025
[2]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inIEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

work page 2016
[3]

BERT: pre-training of deep bidirectional transformers for language understanding,

J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional transformers for language understanding,” inNorth American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, J. Burstein, C. Doran, and T. Solorio, Eds., 2019, pp. 4171–4186

work page 2019
[4]

Learning to prompt for continual learning,

Z. Wang, Z. Zhang, C. Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. G. Dy, and T. Pfister, “Learning to prompt for continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 139–149

work page 2022
[5]

Coda-prompt: Con- tinual decomposed attention-based prompting for rehearsal-free continual learning,

J. S. Smith, L. Karlinsky, V . Gutta, P . Cascante-Bonilla, D. Kim, A. Arbelle, R. Panda, R. Feris, and Z. Kira, “Coda-prompt: Con- tinual decomposed attention-based prompting for rehearsal-free continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11 909–11 919

work page 2023
[6]

Consistent prompting for rehearsal- free continual learning,

Z. Gao, J. Cen, and X. Chang, “Consistent prompting for rehearsal- free continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 28 463–28 473

work page 2024
[7]

Convolu- tional prompting meets language models for continual learning,

A. Roy, R. Moulick, V . K. Verma, S. Ghosh, and A. Das, “Convolu- tional prompting meets language models for continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 616–23 626

work page 2024
[8]

Podnet: Pooled outputs distillation for small-tasks incremental learning,

A. Douillard, M. Cord, C. Ollion, T. Robert, and E. Valle, “Podnet: Pooled outputs distillation for small-tasks incremental learning,” inEuropean Conference on Computer Vision, vol. 12365. Springer, 2020, pp. 86–102

work page 2020
[9]

Cross-class feature augmentation for class incremental learning,

T. Kim, J. Park, and B. Han, “Cross-class feature augmentation for class incremental learning,” inAAAI Conference on Artificial Intelligence. AAAI Press, 2024, pp. 13 168–13 176

work page 2024
[10]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 40, no. 12, pp. 2935–2947, 2018

work page 2018
[11]

Elastic feature consolidation for cold start exemplar-free incremental learning,

S. Magistri, T. Trinci, A. Soutif-Cormerais, J. van de Weijer, and A. D. Bagdanov, “Elastic feature consolidation for cold start exemplar-free incremental learning,” inInternational Conference on Learning Representations, 2024

work page 2024
[12]

Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation,

H. Yan, L. Wang, K. Ma, and Y. Zhong, “Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024, pp. 23 670– 23 680

work page 2024
[13]

Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing,

P . Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig, “Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing,”ACM Computing Sur- veys, vol. 55, no. 9, pp. 195:1–195:35, 2023

work page 2023
[14]

Learning transferable visual models from natu- ral language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natu- ral language supervision,” inInternational Conference on Machine Learning, vol. 139. PMLR, 2021, pp. 8748–8763

work page 2021
[15]

PIVOT: prompting for video continual learning,

A. Villa, J. L. Alc ´azar, M. Alfarra, K. Alhamoud, J. Hurtado, F. C. Heilbron, A. Soto, and B. Ghanem, “PIVOT: prompting for video continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2023, pp. 24 214–24 223

work page 2023
[16]

Dualprompt: Complementary prompting for rehearsal-free continual learning,

Z. Wang, Z. Zhang, S. Ebrahimi, R. Sun, H. Zhang, C. Lee, X. Ren, G. Su, V . Perot, J. G. Dy, and T. Pfister, “Dualprompt: Complementary prompting for rehearsal-free continual learning,” inEuropean Conference on Computer Vision, S. Avidan, G. J. Brostow, M. Ciss ´e, G. M. Farinella, and T. Hassner, Eds., vol. 13686, 2022, pp. 631–648

work page 2022
[17]

S-prompts learning with pre- trained transformers: An occam’s razor for domain incremental learning,

Y. Wang, Z. Huang, and X. Hong, “S-prompts learning with pre- trained transformers: An occam’s razor for domain incremental learning,” inAdvances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., 2022

work page 2022
[18]

Hide-pet: Continual learning via hierarchical decomposition of parameter-efficient tun- ing,

L. Wang, J. Xie, X. Zhang, H. Su, and J. Zhu, “Hide-pet: Continual learning via hierarchical decomposition of parameter-efficient tun- ing,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 8, pp. 6687–6702, 2025

work page 2025
[19]

Self-expansion of pre- trained models with mixture of adapters for continual learning,

H. Wang, H. Lu, L. Yao, and D. Gong, “Self-expansion of pre- trained models with mixture of adapters for continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 10 087–10 098

work page 2025
[20]

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
[21]

The many faces of robustness: A critical analysis of out-of-distribution generalization,

D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, S. Parajuli, M. Guo, D. Song, J. Steinhardt, and J. Gilmer, “The many faces of robustness: A critical analysis of out-of-distribution generalization,” inIEEE/CVF International Conference on Computer Vision. IEEE, 2021, pp. 8320–8329

work page 2021
[22]

Natural adversarial examples,

D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt, and D. Song, “Natural adversarial examples,” inIEEE Conference on Computer Vision and Pattern Recognition. Computer Vision Foundation / IEEE, 2021, pp. 15 262–15 271

work page 2021
[23]

A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark

X. Zhai, J. Puigcerver, A. Kolesnikov, P . Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy et al., “A large-scale study of representation learning with the vi- sual task adaptation benchmark,”arXiv preprint arXiv:1910.04867, 2019

work page internal anchor Pith review arXiv 1910
[24]

Imagenet-21k pre- training for the masses,

T. Ridnik, E. B. Baruch, A. Noy, and L. Zelnik, “Imagenet-21k pre- training for the masses,” inNeural Information Processing Systems Track on Datasets and Benchmarks, J. Vanschoren and S. Yeung, Eds., 2021

work page 2021
[25]

Image BERT pre-training with online tokenizer,

J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. L. Yuille, and T. Kong, “Image BERT pre-training with online tokenizer,” inInternational Conference on Learning Representations, 2022

work page 2022
[26]

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,” inIEEE/CVF International Conference on Computer Vision. IEEE, 2021, pp. 9630–9640

work page 2021
[27]

Hierarchi- cal decomposition of prompt-based continual learning: Rethinking obscured sub-optimality,

L. Wang, J. Xie, X. Zhang, M. Huang, H. Su, and J. Zhu, “Hierarchi- cal decomposition of prompt-based continual learning: Rethinking obscured sub-optimality,” inAdvances in Neural Information Pro- cessing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., 2023

work page 2023
[28]

Revisiting class- incremental learning with pre-trained models: Generalizability and adaptivity are all you need,

D. Zhou, Z. Cai, H. Ye, D. Zhan, and Z. Liu, “Revisiting class- incremental learning with pre-trained models: Generalizability and adaptivity are all you need,”International Journal of Computer Vision, vol. 133, no. 3, pp. 1012–1032, 2025

work page 2025
[29]

Orthogonal subspace learning for language model continual learning,

X. Wang, T. Chen, Q. Ge, H. Xia, R. Bao, R. Zheng, Q. Zhang, T. Gui, and X. Huang, “Orthogonal subspace learning for language model continual learning,” inFindings of the Association for Compu- tational Linguistics: EMNLP, 2023, pp. 10 658–10 671

work page 2023
[30]

Inflora: Interference-free low-rank adaptation for continual learning,

Y. Liang and W. Li, “Inflora: Interference-free low-rank adaptation for continual learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 638–23 647

work page 2024
[31]

Plan: Proactive low-rank allo- IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE, VOL. *, NO. *, AUGUST 2025 9 cation for continual learning,

X. Wang, Z. Zhuang, and Y. Zhang, “Plan: Proactive low-rank allo- IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE, VOL. *, NO. *, AUGUST 2025 9 cation for continual learning,” inIEEE/CVF International Conference on Computer Vision, 2025, pp. 2909–2918

work page 2025
[32]

Adapter merging with centroid prototype mapping for scalable class-incremental learn- ing,

T. Fukuda, H. Kera, and K. Kawamoto, “Adapter merging with centroid prototype mapping for scalable class-incremental learn- ing,” inIEEE/CVF Conference on Computer Vision and Pattern Recog- nition. Computer Vision Foundation / IEEE, 2025, pp. 4884–4893

work page 2025
[33]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems, 2017, pp. 5998– 6008

work page 2017
[34]

Conditional positional encodings for vision transformers,

X. Chu, Z. Tian, B. Zhang, X. Wang, and C. Shen, “Conditional positional encodings for vision transformers,” inInternational Con- ference on Learning Representations, 2023

work page 2023
[35]

Attention guided CAM: visual explanations of vision transformer guided by self-attention,

S. Leem and H. Seo, “Attention guided CAM: visual explanations of vision transformer guided by self-attention,” inAAAI Conference on Artificial Intelligence, AAAI 2024, 2024, pp. 2956–2964

work page 2024