Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective

[Cons 3. Providing performance comparisons with smaller models such as MobileNet.]

Thank you for the great suggestion. We conducted additional experiments on ImageNet pre-trained MobileNet and CIFAR100 using our method, HYDRA, and OMP. We observe that our method achieves {2.24%, 1.27%} higher accuracy than {HYDRA, OMP} at 50% sparsity level as shown in Table R11. The results are also included in the revision of our paper in Table A12. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R11. Performance comparison of our method, HYDRA, and OMP on ImageNet pre-trained MobileNet and CIFAR100.

[Cons 4. It would be insightful to explore the performance without pre-trained weights]

Thank you for the advice. We state that the pruning paradigm of our method which requires first finding masks in a trained model and then tuning the subnetwork to recover the accuracy can’t be used on models without pre-training. This is because a model without pre-training has random weights making the mask-finding stage of our method useless.

Additional experiments of training from scratch. To further alleviate reviewer hiB3’s concern, we conducted additional experiments by applying the VPNs pruning paradigm on SynFlow named VPNs w. SynFlow and pruning from scratch on ResNet-18 and CIFAR100. We observe that VPNs w. SynFlow(with VP) achieves {8.85%, 5.91%} higher accuracy than the original SynFlow (without VP) at {50%, 90%} sparsity levels as shown in Table R12, which indicates visual prompting significantly enhances models’ sparsification in the setting of training from scratch. We also report the results in the revision of our paper in Table A13. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R12. Performance comparison of VPNs w. SynFlow and SynFlow pruning from scratch on ResNet-18 and CIFAR100.

[Cons 5. It might be perplexing why the curve representing "ours" appears as a straight line.]

Thank you for the great point. We use a similar plot format as in [15]. Regarding reviewer hiB3’s concern, we deleted the “Our Best” dashed lines in all plots in the revision of our paper.

[15] Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Mingyi Hong, Yanzhi Wang, and Sijia Liu. Advancing model pruning via bi-level optimization. Advances in Neural Information Processing Systems, 35:18309–18326, 2022a.

Many thanks to reviewer hiB3 for acknowledging that our method achieves “remarkable efficiency” on CIFAR10 and CIFAR100, our writing proficiency is “commendable”, our graphics are “well-crafted”, and our narrative is “compelling”. We sincerely appreciate all constructive suggestions, which help us to improve our paper further. To address reviewer hiB3’s concerns, we provide pointwise responses below.

[Cons 1. Only presenting experiments on CIFAR-10 and CIFAR-100 and conducting experiments on large datasets is essential.]

This is a possible misunderstanding of our experiment setting. As we state in the Abstract, Line 15; Section 4.1, Paragraph 1; Section 4.2, Paragraph 1; Figure 4, we conduct our experiments on eight datasets which are Tiny-ImageNet, Food101, CIFAR100, CIFAR10, DTD, Flowers102, StanfordCars, and OxfordPets. We also display the results of pruning on ImageNet and fine-tuning on Tiny-ImageNet and CIFAR100 in Figure 6.

Additional experiments on ImageNet-1K. To further alleviate reviewer hiB3’s concern, we conducted additional experiments on ImageNet-1K using our method and some of the best baselines such as HYDRA and OMP on ImageNet-1K pre-trained ResNet-18 network. The empirical results are shown in Table R10, our method achieves the best accuracy at 50% and 90% sparsity levels, which illustrates the effectiveness of our method on ImageNet. We also report the results in the revision of our paper in Table A5. Due to the time limitation of rebuttal, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R10. Performance comparison of our method, HYDRA, and OMP on ImageNet pre-trained ResNet-18 and ImageNet.

[Cons 2. Our method uses a resolution of 32.]

Thanks for pointing out the question. We are aware of the role of resolution in the experiments and our responses to this concern are as follows.

1. [We use a resolution of 224] We kindly indicate that we unify the resolution of 224 in our method and the baselines in all of our results in Figure 4 \ 5 \ 6 \ 7 \ 8. As displayed in Section 4.1, Paragraph 3, Line 5, we use an input size of 224 and pad size of 16, which corresponds to Figure 3 that i = 224 and p = 16. This is also illustrated in our code, the main.py file, line 30, we use an input size of 224. For reviewer hiB3’s concern, we significantly polished our paper in Section 4.1, Paragraph 3 in our revision by specifying both our method and baselines using a resolution of 224 to avoid misunderstanding.

2. [We have results on datasets that have a resolution of 224] We gently state that we present results on Tiny-ImageNet which has an original resolution of 224 in Figure 4 \ 5 \ 6. The findings of these experiments also demonstrate our method is superior to all the baselines at an original resolution of 224 circumstances.

3. [Additional results on ImageNet] The results on ImageNet in Table R10 also illustrate that our method achieves superior performance on ImageNet.

Dear Reviewer hiB3,

We thank reviewer hiB3 time for the review and constructive comments. We really hope to have a further discussion with the reviewer hiB3 to see if our response solves the concerns.

In our response, we have (1) clarified the resolution of our method; (2) conducted additional experiments on ImageNet, MobileNet, and models without pre-trained weights, further indicating the superiority of our method.

We genuinely hope reviewer hiB3 could kindly check our response. Thanks!

Best wishes,

Authors

Thank you for recognizing that our idea is “easy to follow and effective” and our performance is “interesting”. To address reviewer eRqj’s concerns, we detailed our responses below.

[Cons 1. Focusing more on the structured sparse case.]

Thank you for the suggestion. We kindly point out that our main contributions focus on introducing the new data-model co-design pruning paradigm and we respectfully argue that focusing on unstructured pruning is highly meaningful. Unstructured pruning is extremely useful as both a mathematical prototype and an empirical testbed for new SNN algorithms and it is also receiving increasingly better support in practice. The evidence is as follows:

1. [Better performance than structured pruning] As the finest-grained and most flexible sparsity level, unstructured sparsity is superior to other more structured forms of sparsity [7].

2. [Widely used on nonGPU hardware] Unstructured sparsity has widely proven its practical relevance on nonGPU hardware, such as CPUs or customized accelerators. For instance, in the range of 70-90% high unstructured sparsity, XNNPACK [8] has already shown significant speedups over dense baselines on smartphone processors.

3. [Receive increasing support] The hardware support of unstructured sparsity may be relatively limited on “off-the-shelf” commodity GPUs/TPUs, but it keeps improving quickly over the years. For example, advanced GPU kernels such as NVIDIA cuSPARSE [9] and Sputnik [10] have built the momentum to better support finer-grained sparsity.

4. [Our method achieves time reduction in mask finding] We demonstrate that our method (unstructured pruning) achieves significant time reduction in terms of mask finding compared to other pruning methods like HYDRA and BiP as shown in Figure 9.

[Cons 2. The speedup ratio should use latency.]

Thank you for the advice. We use FLOPs following [11, 12] and the FLOPs are also widely used in calculating the speedup ratio. To alleviate reviewer eRqj’s concern, we provide the latency speedup ratio of our method on the Quadro RTX 6000 GPU in Table R6. Our method achieves latency speedup ratios of 1.1× and 1.2× at 10% and 20% channel-wise sparsity respectively without compromising the performance relative to the dense network, as indicated in Figure 8. We kindly point out that our main contributions focus on introducing the new data-model co-design pruning paradigm and unstructured pruning, we provide the results of structured pruning only to indicate a potential research direction. We also include the latency results in the revised paper in Table A9.

Table R6. The latency and FLOPs of VPNs structured pruning on ImageNet pre-trained ResNet-18 and CIFAR100.

[Cons 3. More recent structured pruning methods should be compared.]

Thank you for the suggestion. We gently indicate that we have considered several recent state-of-the-art methods such as DepGraph which is a very strong baseline published in 2023 in Section 4.2, Paragraph 1, Line 3.

Additional experiments on structured pruning. To further address reviewer eRqj’s concern, we conducted additional experiments on ImageNet pre-trained ResNet-18 and CIFAR100 using GReg [13] and LAMP [14] which are two recent structured pruning methods. We observe that our method has the best performance as shown in Table R7. We add the new results in our revised paper in Table A10. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version. If the reviewer can kindly point out more methods, we will also add them in our final version.

Table R7. Performance comparison of our method, GReg, and LAMP on ImageNet pre-trained ResNet-18 and CIFAR100 at 20% and 50% channel-wise sparsity levels.

[Cons 4. Our method essentially uses lower resolution input and needs to compare the baselines with lower resolution.]

This might be a misunderstanding of our setting and we respond to the concern from the following aspects:

1. [The resolution of our method and baselines is always 224] In our design, the resolution of our method is always 224 which equals the resolution of the input images in all baselines. Although the visual prompt as learnable parameters are added to the margin of the image, the full image pixel information is preserved. Evidence can be found in Section 4.1, Paragraph 3, Line 5 that we use an input size of 224 and a pad size of 16, and in Figure 3 that i = 224 and p = 16.

2. [Additional experiments of baselines using lower resolution and pruning fewer parameters] To further alleviate reviewer eRqj's concern, we conducted additional experiments on some of the best baselines such as HYDRA and OMP with lower resolution and pruning fewer parameters on ImageNet pre-trained ResNet-18 and CIFAR100. HYDRA and OMP use an input resolution of 192 and prune 13k (the number of parameters in the visual prompt) fewer parameters than our method. We find that our method achieves {3.94%, 3.69%} higher accuracy than {HYDRA, OMP} at 90% sparsity level from the results displayed in Table R8, which indicates our method is superior to baselines using lower resolution and pruning fewer parameters. We also add the results in the revision of our paper in Table A11. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R8. Performance comparison of our method, HYDRA, and OMP on ImageNet pre-trained ResNet-18 and CIFAR100. HYDRA and OMP use 192 as the input resolution and prune 13k fewer parameters than our method.

[Cons 5. Whether our method will deteriorate in applications that are sensitive to resolution such as object detection.]

This is an interesting direction to explore. We provide detailed responses below.

1. [Our method is insensitive to resolution] As illustrated above, our method and the baselines all use a resolution of 224 in the original setting and our method surpasses all baselines as shown in Figure 4. In the additional experiments displayed in Table R8, our method is still better than the baselines using a resolution of 192. No matter when the baselines use the resolution of 224 or 192, our method is consistently better than the baselines, demonstrating our method is insensitive to resolution.

2. [Superior performance on object detection] To further alleviate reviewer eRqj's concern, we conducted supplemental experiments on Pascal VOC 2007, which is a well-known object detection task. We compare our method to HYDRA and OMP on YOLOv4 with ImageNet pre-trained ResNet-18 backbone. Our method achieves {3.78%, 2.67%} higher AP than {HYDRA, OMP} at 90% sparsity level as shown in Table R9, which indicates the superiority of our method in resolution-sensitive tasks. The results are also included in the revision of our paper in Table A6. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R9. AP comparison of our method, HYDRA, and OMP on YOLOv4 with ImageNet pre-trained ResNet-18 backbone and Pascal VOC 2007.

[7] Mao, Huizi, Han, Song, Pool, Jeff, Li, Wenshuo, Liu, Xingyu, Wang, Yu & Dally, William J 2017 Exploring the granularity of sparsity in convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 13–20.

[8] Elsen, Erich, Dukhan, Marat, Gale, Trevor & Simonyan, Karen 2020 Fast sparse convnets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14629–14638.

[9] Valero-Lara, Pedro, Martínez-Pérez, Ivan, Sirvent, Raül, Martorell, Xavier & Pena, Antonio J 2018 Nvidia gpus scalability to solve multiple (batch) tridiagonal systems implementation of cuthomasbatch. In Parallel Processing and Applied Mathematics: 12th International Conference, PPAM 2017, Lublin, Poland, September 10- 13, 2017, Revised Selected Papers, Part I, pp. 243–253. Springer.

[10] Gale, Trevor, Zaharia, Matei, Young, Cliff & Elsen, Erich 2020 Sparse gpu kernels for deep learning. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pp. 1–14. IEEE.

[11] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In Proceedings of the IEEE international conference on computer vision, pp. 2736–2744, 2017.

[12] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16091–16101, 2023.

[13] Huan Wang, Can Qin, Yulun Zhang, and Yun Fu. Neural pruning via growing regularization. arXiv preprint arXiv:2012.09243, 2020.

[14] Jaeho Lee, Sejun Park, Sangwoo Mo, Sungsoo Ahn, and Jinwoo Shin. Layer-adaptive sparsity for the magnitude-based pruning. arXiv preprint arXiv:2010.07611, 2020.

Dear Reviewer eRqj,

We thank reviewer eRqj time for the review and constructive comments. We really hope to have a further discussion with the reviewer eRqj to see if our response solves the concerns.

In our response, we have (1) interpreted the contribution of our method and the significance of unstructured pruning; (2) conducted additional experiments on structured pruning, object detection, and baselines with lower resolution, further indicating the superiority of our method.

We genuinely hope reviewer eRqj could kindly check our response. Thanks!

Best wishes,

Authors

Above all, the rebuttal has not addressed my concerns at all. I have to lower my score.

Dear Reviewer eRqj,

We genuinely appreciate your dedicated time and effort in reviewing our work. As the final day of the discussion period approaches, we kindly request that you share any additional questions or concerns you may have. We are eager to engage in further discussions with you.

If our responses have adequately addressed your concerns, we kindly ask that you consider raising the score of our work. Again, we are truly grateful for your valuable time and efforts.

Warm regards,

The Authors

[Cons 3. How about learning additional parameters without using the visual prompt?]

Thank you for the suggestions. To the best of our understanding, the reviewer suggests comparing our method (with VP) to the baselines which learn more parameters than ours. Here is our response to the questions.

1. [Current results] As displayed in Figure 4, our method achieves better performance at 80% sparsity than other baselines at {40%, 50%, 60%, 70%} sparsity levels on CIFAR100, CIFAR10, Food101, Flowers102, DTD, and StanfordCars.

2. [Additional experiments of the baselines learning additional parameters] We implemented supplemental experiments on ImageNet pre-trained ResNet-18 on CIFAR100, using our method, HYDRA, and OMP. HYDRA and OMP learn an additional 13k (the number of parameters in the visual prompt) parameters than our method. We find that our method obtains {3.46%, 3.06%} higher accuracy than {HYDRA, OMP} at 90% sparsity level as displayed in Table R4, which manifests that our method is better than the baselines with additional parameters. The results are also reported in our revised paper in Table A7. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R4. Performance comparison of our method, HYDRA, and OMP on ImageNet pre-trained ResNet-18 and CIFAR100. HYDRA and OMP learn 13k additional parameters than our method.

[Cons 4. How would visual prompt be different from data augmentation?]

Visual prompt is essentially different from data augmentation, and the reasons are as follows:

1. [Purpose] Visual prompts are usually designed to improve model adaption. While data augmentations are used to avoid overfitting.

2. [Methodology] Visual prompts can be optimized in a data-driven way. However, most data augmentations are data-agnostic.In our method, visual prompting is the key to enabling the data-model exploration of crucial sparsity topologies.

3. [Additional experiments of more data augmentations] To further address reviewer sGzb’s concern, we compare our method without mix-ups to HYDRA and OMP with mix-ups on ImageNet pre-trained ResNet-18 and CIFAR10. Our method achieves {5.46%, 3.50%} higher accuracy than {HYDRA, OMP} at 90% sparsity level as shown in Table R5, which indicates that visual prompting is better than using more data augmentation. We add the results to the revised paper in Table A8. Due to the time limitation, we provide outcomes only at 50% and 90% sparsity levels, more sparsity levels are promised in our final version.

Table R5. Performance comparison of our method (without mix-up), HYDRA (with mix-up), and OMP (with mix-up) on ImageNet pre-trained ResNet-18 and CIFAR10.

[3] Hidenori Tanaka, Daniel Kunin, Daniel L Yamins, and Surya Ganguli. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in neural information processing systems, 33:6377–6389, 2020.

[4] Zhaozhuo Xu, Zirui Liu, Beidi Chen, Yuxin Tang, Jue Wang, Kaixiong Zhou, Xia Hu, and An-shumali Shrivastava. Compress, then prompt: Improving accuracy-efficiency trade-off of llm inference with transferable prompt. arXiv preprint arXiv:2305.11186, 2023.

[5] Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19133–19143, 2023.

[6] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16816–16825, 2022a.

Thank you for recognizing that our results and the direction of our work are “interesting”. To address reviewer sGzb’s concerns, we provide detailed responses below.

[Cons 1. The statement of “data-model co-design” is not concrete.]

Thanks for putting forward the point. Specifically, data-model co-design indicates both the input data and the model are optimized in pruning. This paradigm of our method is significantly different from the baselines which are model-centric.

1. [Model-centric design] Focusing on searching and preserving crucial weights by analyzing network topologies (Abstract, Line 6) with fixed inputs as illustrated in [3].
2. [Data-model co-design] Optimizing weight masks and visual prompts jointly by combining data-centric design that constructs well-designed prompts (Section 1, Paragraph 3, Line 3) and model-centric design. More explanation of data-model co-design can be found in Section 1, Paragraph 5, Point 2; Figure 2; and Section 3, Paragraph 1.

To further alleviate reviewer sGzb’s concern, we significantly polished our paper in Section 1, Paragraph 5, and Section 3, Paragraph 1 by interpreting the data-model co-design more concretely to avoid misunderstanding in our revision.

[Cons 2. Why is the visual prompt essential?]

Thank you for the question. The significance of the visual prompt is detailed as follows:

1. [Motivation]

• Xu et al. [4] demonstrate that prompts can recover compressed LLMs which illuminates the efficacy of post-pruning prompts in enhancing the performance of compressed LLMs. Observing that post-pruning prompts bolster the efficiency and performance of LLMs, It is only natural to inquire about the impact of VPs on vision models.

• We observe from Figure 1 that post-pruning prompts escalate the performance of the subnetworks before fine-tuning. However, neither of these settings consistently surpasses the standard no-prompting approach (pruning + fine-tuning). We postulate visual prompts are capable of enhancing the sparsification of vision models by utilizing them in a different way.

2. [Empirical evidence]

• As shown in Figure 4 and Figure 5, we demonstrate the essential role of VP by comparing our method (with VP) to HYDRA (without VP) on eight downstream datasets and three architectures.

• As shown in Figure 7, we observe that VP combined with existing prunings consistently surpasses their original counterpart (Section 4.2, Paragraph 4), which further illustrates the essential of VP.

3. [Previous literature]

• [5, 6] showcases visual prompting as a parameter-efficient method that substantially improves the generalization and accuracy of vision models over multiple tasks. We believe these advantages brought by visual prompts also benefit pruning.

Dear Reviewer sGzb,

We thank reviewer sGzb time for the review and constructive comments. We really hope to have a further discussion with the reviewer sGzb to see if our response solves the concerns.

In our response, we have (1) interpreted the data-model co-design and the importance of visual prompts concretely; (2) conducted additional experiments on baselines learning more parameters and with more data augmentations, further indicating the superiority of our method.

We genuinely hope reviewer sGzb could kindly check our response. Thanks!

Best wishes,

Authors

Dear Reviewer sGzb,

We genuinely appreciate your dedicated time and effort in reviewing our work. As the final day of the discussion period approaches, we kindly request that you share any additional questions or concerns you may have. We are eager to engage in further discussions with you.

If our responses have adequately addressed your concerns, we kindly ask that you consider raising the score of our work. Again, we are truly grateful for your valuable time and efforts.

Warm regards,

The Authors

Many thanks to reviewer 8b5d for acknowledging that our proposal is “intriguing”, our experiments “validate the effectiveness” of our method, our writing and presentation are “quite good”, and we demonstrate the “transferability” of our approach. We sincerely appreciate all constructive suggestions, which help us to improve our paper further. To address reviewer 8b5d’s questions, we provide pointwise responses below.

[Cons 1. Comparing models with visual prompts to those without might not be entirely fair.]

We respectfully disagree. We have tried our best to guarantee fairness in the empirical results shown in Figure 7. Specifically, in the setting of Figure 7:

1. [For our method (VPNs)] We leverage VP in mask finding and subnetwork tuning.
2. [For baselines (VPNs w. Random, VPNs w. OMP, VPNs w. LTH)] We apply the same VP as our method to Random, OMP, and LTH in mask finding and subnetwork tuning.

Both our method and baselines use exactly the same fine-tuning of subnetwork weights and VP. But our method apparently surpasses all baselines with visual prompting at all sparsity levels as shown in Figure 7, which indicates the superiority of our method in the “entirely fair” setting.

Additional experiments of baselines with VP. To further address reviewer 8b5d’s concern, we conduct supplemental experiments using our method and some of the best baselines such as LTH and OMP with VP on ImageNet pre-trained ResNet-18 network and Tiny-imageNet. We observe that our method is superior to baselines with VP as displayed in Table R1. We also include the results in the revision of our paper in Table A4. Due to the time limitation of rebuttal, we provide outcomes at 50% and 90% sparsity levels, and more sparsity levels are promised in our final version.

Table R1. Performance comparison of our method, VPNs w. LTH, and VPNs w. OMP on ImageNet pre-trained ResNet-18 and Tiny-ImageNet.

[Cons 2. Why not directly conduct experiments on ImageNet itself?]

Thank you for the great advice. We provided results of pruning on ImageNet and fine-tuning on downstream datasets for transfer study in Figure 6.

Additional experiments on ImageNet. To further alleviate reviewer 8b5d’s concern, we conducted additional experiments on ImageNet using our method and some of the best baselines such as HYDRA and OMP, on ImageNet pre-trained ResNet-18 network. Our method has {0.52%, 2.87%} higher accuracy than {HYDRA, OMP} at 50% sparsity level as shown in Table R2, which illustrates the effectiveness of our method on ImageNet. We also report the results in the revision of our paper in Table A5. Due to the time limitation of rebuttal, we provide outcomes at 50% and 90% sparsity levels, more sparsity levels are promised in our final version.

Table R2. Performance comparison of our method, HYDRA, and OMP on ImageNet pre-trained ResNet-18 and ImageNet.

[Cons 3. How can the proposed approach be used for other tasks, such as detection?]

Thank you for the great point. Following the reviewer’s suggestion, we carried out additional experiments on Pascal VOC 2007 [1], which is one of the most widely used datasets for object detection tasks.

Additional experiments on Pascal VOC 2007. We compared our method to some of the best baselines such as HYDRA and OMP on YOLOv4 [2] which uses ImageNet pre-trained ResNet-18 as the backbone. The outcomes are presented in Table R3. Our method achieves {3.78%, 2.67%} higher AP than {HYDRA, OMP} at 90% sparsity level, which demonstrates the superiority of our method on object detection. We add the results to the revised paper in Table A6. Due to the time limitation, we provide outcomes only at two sparsity levels, more sparsity levels are promised in our final version.

Table R3. AP comparison of our method, HYDRA, and OMP on YOLOv4 with ImageNet pre-trained ResNet-18 backbone and Pascal VOC 2007.

[1] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–338, 2010.

[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020.

Dear Reviewer 8b5d,

We thank reviewer 8b5d time for the review and constructive comments. We really hope to have a further discussion with the reviewer 8b5d to see if our response solves the concerns.

In our response, we have (1) clarified the fairness of our setting and carried out supplemental experiments on baselines with visual prompting; (2) conducted additional experiments on ImageNet and object detection tasks, further demonstrating the superiority of our method.

We genuinely hope reviewer 8b5d could kindly check our response. Thanks!

Best wishes,

Authors

Dear Reviewer 8b5d,

We genuinely appreciate your dedicated time and effort in reviewing our work. As the final day of the discussion period approaches, we kindly request that you share any additional questions or concerns you may have. We are eager to engage in further discussions with you.

If our responses have adequately addressed your concerns, we kindly ask that you consider raising the score of our work. Again, we are truly grateful for your valuable time and efforts.

Warm regards,

The Authors