pith. sign in

arxiv: 2606.12023 · v1 · pith:SJ3DZXM5new · submitted 2026-06-10 · 💻 cs.CV

ViT-FREE: Efficient Face Recognition via Early Exiting and Synthetic Adaptation

Tahar Chettaoui , Guray Ozgur , Eduarda Caldeira , Naser Damer , Fadi Boutros This is my paper

Pith reviewed 2026-06-27 09:49 UTC · model grok-4.3

classification 💻 cs.CV
keywords early exitingvision transformerface recognitionefficient inferencetraining-freesynthetic adaptation
0
0 comments X

The pith

Exiting at layer 10 of a pretrained Vision Transformer speeds up face recognition by 20 percent with a 1.5 point accuracy drop.

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

Vision Transformers achieve strong face recognition but require substantial computation across all layers. The paper demonstrates that features refine progressively, so intermediate layers already hold discriminative information for verification. By attaching exit points without retraining, one can terminate early and save compute. On IJB-C, layer 10 exit delivers the 20 percent speedup at minimal cost. A follow-on method fine-tunes only the exit heads on synthetic images to lift shallow-exit quality.

Core claim

Pretrained ViTs for face recognition exhibit gradual feature refinement where patch embeddings and attention maps become increasingly aligned with the final representation, allowing direct verification from intermediate blocks via a training-free multi-exit framework.

What carries the argument

The ViT-FREE multi-exit framework that uses intermediate transformer encoder outputs for face verification without altering the backbone model.

If this is right

  • Later exits provide better accuracy while still reducing inference cost compared to full model.
  • Exiting at layer 10 achieves up to 20% speedup on IJB-C with 1.5 drop in performance.
  • ViT-FREE_FT with synthetic data adaptation improves shallow exits without affecting deeper ones or the backbone.
  • Uniform feature dimensionality across blocks enables attachment of verification heads at any depth.

Where Pith is reading between the lines

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

  • Similar early-exit logic could apply to ViT models in other vision tasks like object detection or segmentation.
  • Dynamic exiting per image based on confidence at early layers might further improve average speed.
  • The progressive convergence suggests that ViT depth is over-provisioned for many face recognition cases.

Load-bearing premise

Intermediate layers produce representations stable and discriminative enough for face verification because embeddings and attention maps align progressively with the final output.

What would settle it

An experiment on IJB-C or similar benchmark where the verification performance at layer 10 drops by more than 3 points compared to the full model would challenge the claimed trade-off.

Figures

Figures reproduced from arXiv: 2606.12023 by Eduarda Caldeira, Fadi Boutros, Guray Ozgur, Naser Damer, Tahar Chettaoui.

Figure 1
Figure 1. Figure 1: Cosine similarity of patch embeddings hi at different depths (0–11) in a ViT, averaged over the LFW benchmark. The blue bars show the similarity between each interme￾diate patch embedding hi and the final patch embedding h11, while the orange bars indicate the pairwise similarity between consecutive patch embeddings hi and hi+1. The x￾axis represents the depth of the ViT, and the y-axis shows the cosine si… view at source ↗
Figure 2
Figure 2. Figure 2: ViT-FREE inference pipeline. Intermediate patch em [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cosine similarity between intermediate feature em [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of ViT early exits at different depths (1–12) on [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attention maps of ViT-FREE at depths 1–12, cor [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: ViT-FREE and ViT-FREEF T efficiency–performance trade-off. SB represents the average over several small benchmarks, as defined in Section V-A. Each point corresponds to a specific early-exit depth, illustrating the trade-off between computational latency and FR performance. The red curve represents the baseline ViT-FREE, while the blue curve corresponds to the fine-tuned variant ViT-FREEF T . Fine-tuning s… view at source ↗
Figure 7
Figure 7. Figure 7: Cosine similarity of Attention maps at different depths [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

Vision Transformers (ViTs) have gained significant attention in computer vision and shown strong potential for face recognition (FR). However, their high computational cost makes deployment on resource-constrained devices challenging, motivating the need for methods that balance efficiency and accuracy. In this work, we investigate early exiting in pretrained ViTs as a simple yet effective training-free strategy for efficient FR inference. Leveraging the uniform feature dimensionality across transformer encoder blocks, we introduce ViT-FREE, a multi-exit framework that enables face verification directly from intermediate representations without modifying or retraining the backbone model, and thus, reducing inference cost. Empirically, we show that patch embeddings and attention maps evolve progressively across depth, exhibiting high similarity between consecutive ViT blocks and increasing alignment with the final representation. This indicates gradual feature refinement and attention convergence, suggesting that intermediate layers already provide stable and discriminative representations suitable for early exiting. Through extensive experiments on multiple FR benchmarks, we systematically analyze the accuracy-efficiency trade-off across exit depths. Our results demonstrate that later exits achieve a highly favorable balance, with exiting at layer 10 yielding up to a 20% speedup while incurring only a 1.5 drop in verification performance on benchmarks such as IJB-C. Also, we propose ViT-FREE_FT, a lightweight exit-specific fine-tuning strategy that adapts only the projection layers using a small synthetic dataset while keeping the transformer backbone frozen. This approach improves the performance of shallow exits while preserving the efficiency benefits and leaving deeper exits largely unaffected.

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 / 2 minor

Summary. The paper proposes ViT-FREE, a training-free multi-exit framework that performs face verification directly from intermediate representations of a pretrained ViT backbone by leveraging uniform feature dimensionality across encoder blocks. It reports empirical observations that patch embeddings and attention maps evolve progressively with high inter-block similarity and increasing alignment to the final layer, enabling early exits. Experiments on multiple FR benchmarks show that exiting at layer 10 yields up to 20% speedup with only a 1.5-point drop in verification performance (e.g., on IJB-C); a lightweight exit-specific fine-tuning variant (ViT-FREE_FT) using synthetic data is also introduced to improve shallow exits while keeping the backbone frozen.

Significance. If the reported accuracy-efficiency trade-offs hold under standard protocols, the work provides a practical, zero-retraining method to reduce ViT inference cost for face recognition on resource-constrained devices. The training-free core and the synthetic-adaptation extension are clear strengths; the empirical validation across benchmarks supports the central claim without circularity or invented parameters.

major comments (2)
  1. [Experiments] Experiments section: the central claim of a 20% speedup with 1.5-point verification drop at layer-10 exit is supported by benchmark measurements, but the manuscript provides no details on exact evaluation protocols, data splits, number of runs, or error bars; this information is required to verify reproducibility of the accuracy-efficiency curves.
  2. [§3] §3 (empirical observations): the justification for early exiting rests on progressive evolution of patch embeddings and attention maps, yet the text does not report quantitative similarity metrics (e.g., cosine similarity or attention-map correlation) across consecutive blocks or versus the final representation; without these numbers the link between the stated observation and the suitability of intermediate layers remains qualitative.
minor comments (2)
  1. [Method] Clarify in the method section how the verification head is applied to intermediate representations (e.g., whether the same projection layer is reused or a new one is attached per exit).
  2. [Figures/Tables] Table or figure captions should explicitly state the baseline model and hardware used for the reported speedup percentages.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive recommendation of minor revision and the constructive comments. We address each major point below.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the central claim of a 20% speedup with 1.5-point verification drop at layer-10 exit is supported by benchmark measurements, but the manuscript provides no details on exact evaluation protocols, data splits, number of runs, or error bars; this information is required to verify reproducibility of the accuracy-efficiency curves.

    Authors: We agree that these details are necessary for full reproducibility. In the revised manuscript we will add the precise evaluation protocols (standard IJB-C 1:1 verification protocol and equivalent protocols on other benchmarks), the data splits employed, the number of runs, and error bars or standard deviations on all reported metrics. revision: yes

  2. Referee: [§3] §3 (empirical observations): the justification for early exiting rests on progressive evolution of patch embeddings and attention maps, yet the text does not report quantitative similarity metrics (e.g., cosine similarity or attention-map correlation) across consecutive blocks or versus the final representation; without these numbers the link between the stated observation and the suitability of intermediate layers remains qualitative.

    Authors: We acknowledge that Section 3 currently presents the observations qualitatively. In the revision we will augment this section with quantitative metrics, specifically average cosine similarity between consecutive-block patch embeddings, cosine similarity of each intermediate representation to the final-layer representation, and correlation coefficients for the attention maps. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper is an empirical study of early-exiting in pretrained ViTs. All load-bearing claims (layer-10 exit: ~20% speedup, 1.5-point IJB-C drop) are direct measurements on benchmarks rather than derivations. The stated observation of progressive patch/attention evolution is presented as an empirical finding and is not used to derive any quantitative prediction that reduces to the same data by construction. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. The method is explicitly training-free for the core early-exit strategy.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on pretrained ViT backbones and empirical observations of progressive feature refinement; no free parameters, axioms, or invented entities are introduced.

pith-pipeline@v0.9.1-grok · 5819 in / 983 out tokens · 25056 ms · 2026-06-27T09:49:33.895953+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

55 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    X. An, X. Zhu, Y . Gao, Y . Xiao, Y . Zhao, Z. Feng, L. Wu, B. Qin, M. Zhang, D. Zhang, and Y . Fu. Partial FC: training 10 million identities on a single machine. InICCVW, pages 1445–1449. IEEE, 2021

    2021

  2. [2]

    Bakhtiarnia, Q

    A. Bakhtiarnia, Q. Zhang, and A. Iosifidis. Multi-exit vision trans- former for dynamic inference. InBMVC, page 81. BMV A Press, 2021

    2021

  3. [3]

    Bakhtiarnia, Q

    A. Bakhtiarnia, Q. Zhang, and A. Iosifidis. Single-layer vision transformers for more accurate early exits with less overhead.Neural Networks, 153:461–473, 2022

    2022

  4. [4]

    Bolya, C

    D. Bolya, C. Fu, X. Dai, P. Zhang, C. Feichtenhofer, and J. Hoffman. Token merging: Your vit but faster. InICLR. OpenReview.net, 2023

    2023

  5. [5]

    Boutros, E

    F. Boutros, E. Caldeira, T. Chettaoui, and N. Damer. Idperturb: En- hancing variation in synthetic face generation via angular perturbation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    2026

  6. [6]

    Boutros, N

    F. Boutros, N. Damer, F. Kirchbuchner, and A. Kuijper. Elasticface: Elastic margin loss for deep face recognition. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022, pages 1577–1586. IEEE, 2022

    2022

  7. [7]

    Chang, P

    S. Chang, P. Wang, M. Lin, F. Wang, D. J. Zhang, R. Jin, and M. Z. Shou. Making vision transformers efficient from A token sparsification view. InCVPR, pages 6195–6205. IEEE, 2023

    2023

  8. [8]

    Cheng, X

    Z. Cheng, X. Zhu, and S. Gong. Low-resolution face recognition. InACCV (3), volume 11363 ofLecture Notes in Computer Science, pages 605–621. Springer, 2018

    2018

  9. [9]

    Chettaoui, N

    T. Chettaoui, N. Damer, and F. Boutros. Froundation: Are foundation models ready for face recognition?Image Vis. Comput., 156:105453, 2025

    2025

  10. [10]

    E. D. Cubuk, B. Zoph, J. Shlens, and Q. V . Le. Randaugment: Practical automated data augmentation with a reduced search space. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR Workshops 2020, Seattle, WA, USA, June 14-19, 2020, pages 3008–3017. Computer Vision Foundation / IEEE, 2020

    2020

  11. [11]

    J. Dan, Y . Liu, H. Xie, J. Deng, H. Xie, X. Xie, and B. Sun. Transface: Calibrating transformer training for face recognition from a data- centric perspective. InICCV, pages 20585–20596. IEEE, 2023

    2023

  12. [12]

    Darcet, M

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski. Vision transform- ers need registers. InICLR. OpenReview.net, 2024

    2024

  13. [13]

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InCVPR, pages 4690–4699. Computer Vision Foundation / IEEE, 2019

    2019

  14. [14]

    J. Deng, J. Guo, J. Yang, N. Xue, I. Kotsia, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962–5979, Oct. 2022

    2022

  15. [15]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Trans- formers for image recognition at scale. InICLR. OpenReview.net, 2021

    2021

  16. [16]

    Graham, A

    B. Graham, A. El-Nouby, H. Touvron, P. Stock, A. Joulin, H. J ´egou, and M. Douze. Levit: a vision transformer in convnet’s clothing for faster inference. InICCV, pages 12239–12249. IEEE, 2021

    2021

  17. [17]

    J. Guo, K. Han, H. Wu, Y . Tang, X. Chen, Y . Wang, and C. Xu. CMT: convolutional neural networks meet vision transformers. InCVPR, pages 12165–12175. IEEE, 2022

    2022

  18. [18]

    Y . Guo, L. Zhang, Y . Hu, X. He, and J. Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In B. Leibe, J. Matas, N. Sebe, and M. Welling, editors,Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III, volume 9907 ofLecture Notes in Computer Science, pages 8...

    2016

  19. [19]

    Y . Han, G. Huang, S. Song, L. Yang, H. Wang, and Y . Wang. Dynamic neural networks: A survey.IEEE Trans. Pattern Anal. Mach. Intell., 44(11):7436–7456, 2022

    2022

  20. [20]

    M. I. Hosen and M. B. Islam. Himfr: A hybrid masked face recognition through face inpainting.CoRR, abs/2209.08930, 2022

    work page arXiv 2022

  21. [21]

    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. InICLR. OpenReview.net, 2022

    2022

  22. [22]

    G. B. Huang, M. Mattar, T. Berg, and E. Learned-Miller. Labeled Faces in the Wild: A Database forStudying Face Recognition in Unconstrained Environments. InWorkshop on Faces in ’Real-Life’ Images: Detection, Alignment, and Recognition, Marseille, France, Oct. 2008. Erik Learned-Miller and Andras Ferencz and Fr´ed´eric Jurie

    2008

  23. [23]

    Islam, M

    K. Islam, M. Z. Zaheer, and A. Mahmood. Face pyramid vision transformer. InBMVC, page 758. BMV A Press, 2022

    2022

  24. [24]

    M. Khan, M. Saeed, A. El-Saddik, and W. Gueaieb. Artrivit: Auto- matic face recognition system using vit-based siamese neural networks with a triplet loss. InISIE, pages 1–6. IEEE, 2023

    2023

  25. [25]

    M. Kim, Y . Su, F. Liu, A. Jain, and X. Liu. Keypoint relative position encoding for face recognition. InCVPR, pages 244–255. IEEE, 2024

    2024

  26. [26]

    Not all patches are what you need: Expediting vision transformers via token reorganizations.arXiv preprint arXiv:2202.07800, 2022

    Y . Liang, C. Ge, Z. Tong, Y . Song, J. Wang, and P. Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations.CoRR, abs/2202.07800, 2022

    work page arXiv 2022

  27. [27]

    X. Liu, H. Peng, N. Zheng, Y . Yang, H. Hu, and Y . Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In CVPR, pages 14420–14430. IEEE, 2023

    2023

  28. [28]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. InICLR (Poster). OpenReview.net, 2019

    2019

  29. [29]

    Matsubara, M

    Y . Matsubara, M. Levorato, and F. Restuccia. Split computing and early exiting for deep learning applications: Survey and research challenges.ACM Comput. Surv., 55(5):90:1–90:30, 2023

    2023

  30. [30]

    B. Maze, J. C. Adams, J. A. Duncan, N. D. Kalka, T. Miller, C. Otto, A. K. Jain, W. T. Niggel, J. Anderson, J. Cheney, and P. Grother. IARPA janus benchmark - C: face dataset and protocol. InICB, pages 158–165. IEEE, 2018

    2018

  31. [31]

    Mehta and M

    S. Mehta and M. Rastegari. Mobilevit: Light-weight, general-purpose, and mobile-friendly vision transformer. InICLR. OpenReview.net, 2022

    2022

  32. [32]

    Mishra and K

    P. Mishra and K. Sarawadekar. Polynomial learning rate policy with warm restart for deep neural network. InTENCON, pages 2087–2092. IEEE, 2019

    2087

  33. [33]

    Moschoglou, A

    S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia, and S. Zafeiriou. Agedb: the first manually collected, in-the-wild age database. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshop, volume 2, page 5, 2017

    2017

  34. [34]

    Nixon, P

    S. Nixon, P. Ruiu, M. Cadoni, A. Lagorio, and M. Tistarelli. Exploiting face recognizability with early exit vision transformers. InBIOSIG, LNI, pages 1–7. Gesellschaft f ¨ur Informatik e.V . / IEEE, 2023

    2023

  35. [35]

    Nixon, P

    S. Nixon, P. Ruiu, M. Cadoni, A. Lagorio, and M. Tistarelli. Assessing bias and computational efficiency in vision transformers using early exits.EURASIP J. Image Video Process., 2025(1):2, 2025

    2025

  36. [36]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. As- sran, N. Ballas, W. Galuba, R. Howes, P. Huang, S. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. J ´egou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski. Dinov2: Learning robust visual features without supe...

    2024

  37. [37]

    Phuong and C

    M. Phuong and C. Lampert. Distillation-based training for multi- exit architectures. In2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 1355–1364. IEEE, 2019

    2019

  38. [38]

    L. Qin, M. Wang, C. Deng, K. Wang, X. Chen, J. Hu, and W. Deng. Swinface: A multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation.IEEE Trans. Circuits Syst. Video Technol., 34(4):2223–2234, 2024

    2024

  39. [39]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. InICML, volume 139 ofProceedings of Machine Learning Research, pages 8748–8763. PMLR, 2021

    2021

  40. [40]

    Raghu, T

    M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. Do vision transformers see like convolutional neural networks? In M. Ranzato, A. Beygelzimer, Y . N. Dauphin, P. Liang, and J. W. Vaughan, editors,Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, Decembe...

    2021

  41. [41]

    Rasley, S

    J. Rasley, S. Rajbhandari, O. Ruwase, and Y . He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. InKDD, pages 3505–3506. ACM, 2020

    2020

  42. [42]

    Sengupta, J

    S. Sengupta, J. Chen, C. Castillo, V . Patel, R. Chellappa, and D. Ja- cobs. Frontal to profile face verification in the wild. In2016 IEEE Winter Conference on Applications of Computer Vision, WACV 2016, 2016 IEEE Winter Conference on Applications of Computer Vision, W ACV 2016. Institute of Electrical and Electronics Engineers Inc., May 2016. Publisher C...

    2016

  43. [43]

    Sun and G

    Z. Sun and G. Tzimiropoulos. Part-based face recognition with vision transformers. InBMVC, page 611. BMV A Press, 2022

    2022

  44. [44]

    Y . Tang, K. Han, Y . Wang, C. Xu, J. Guo, C. Xu, and D. Tao. Patch slimming for efficient vision transformers. InCVPR, pages 12155– 12164. IEEE, 2022

    2022

  45. [45]

    Teerapittayanon, B

    S. Teerapittayanon, B. McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural networks. InICPR, pages 2464–2469. IEEE, 2016

    2016

  46. [46]

    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.CoRR, abs/1706.03762, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  47. [47]

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu. Cosface: Large margin cosine loss for deep face recognition. InCVPR, pages 5265–5274. Computer Vision Foundation / IEEE Computer Society, 2018

    2018

  48. [48]

    Whitelam, E

    C. Whitelam, E. Taborsky, A. Blanton, B. Maze, J. C. Adams, T. Miller, N. D. Kalka, A. K. Jain, J. A. Duncan, K. Allen, J. Cheney, and P. Grother. IARPA janus benchmark-b face dataset. InCVPR Workshops, pages 592–600. IEEE Computer Society, 2017

    2017

  49. [49]

    Wolczyk, B

    M. Wolczyk, B. W ´ojcik, K. Balazy, I. T. Podolak, J. Tabor, M. Smieja, and T. Trzcinski. Zero time waste: Recycling predictions in early exit neural networks. InNeurIPS, pages 2516–2528, 2021

    2021

  50. [50]

    J. Xin, R. Tang, Y . Yu, and J. Lin. Berxit: Early exiting for BERT with better fine-tuning and extension to regression. In P. Merlo, J. Tiedemann, and R. Tsarfaty, editors,Proceedings of the 16th Con- ference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021, pages 91–104. Associ...

    2021

  51. [51]

    G. Xu, J. Hao, L. Shen, H. Hu, Y . Luo, H. Lin, and J. Shen. Lgvit: Dynamic early exiting for accelerating vision transformer. InACM Multimedia, pages 9103–9114. ACM, 2023

    2023

  52. [52]

    H. Yin, A. Vahdat, J. M. ´Alvarez, A. Mallya, J. Kautz, and P. Molchanov. Adavit: Adaptive tokens for efficient vision transformer. CoRR, abs/2112.07658, 2021

    work page arXiv 2021

  53. [53]

    Zheng and W

    T. Zheng and W. Deng. Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments. Technical Report 18-01, Beijing University of Posts and Telecommunications, February 2018

    2018

  54. [54]

    Cross-Age LFW: A Database for Studying Cross-Age Face Recognition in Unconstrained Environments

    T. Zheng, W. Deng, and J. Hu. Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments. CoRR, abs/1708.08197, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  55. [55]

    Zhong and W

    Y . Zhong and W. Deng. Face transformer for recognition.CoRR, abs/2103.14803, 2021

    work page arXiv 2021