Towards Free Data Selection with General-Purpose Models

A desirable data selection algorithm can efficiently choose the most informative samples to maximize the utility of limited annotation budgets. However, current approaches, represented by active learning methods, typically follow a cumbersome pipeline that iterates the time-consuming model training and batch data selection repeatedly. In this paper, we challenge this status quo by designing a distinct data selection pipeline that utilizes existing general-purpose models to select data from various datasets with a single-pass inference without the need for additional training or supervision. A novel free data selection (FreeSel) method is proposed following this new pipeline. Specifically, we define semantic patterns extracted from inter-mediate features of the general-purpose model to capture subtle local information in each image. We then enable the selection of all data samples in a single pass through distance-based sampling at the fine-grained semantic pattern level. FreeSel bypasses the heavy batch selection process, achieving a significant improvement in efficiency and being 530x faster than existing active learning methods. Extensive experiments verify the effectiveness of FreeSel on various computer vision tasks. Our code is available at https://github.com/yichen928/FreeSel.

Paper

References (67)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer yLeE3/10 · confidence 4/52023-07-05

Summary

This paper presents a data selection methods without training, so saving lots of time than previous methods. By a publicly available pretrained model, FreeSel can select data from datasets with a single-pass inference, without need for additional training. This method uses comparison between semantic patterns of samples with distance-based sampling. Experiments on various tasks including classification, segmentation and detection verify the motivation.

Strengths

1. The proposed method requires no training to select useful examples, saving lots of time. 2. The experimental performance shows the priority about this method.

Weaknesses

1. The method is plain, to compare the similarities among input sample features. Though saving time, it cannot outperform other methods. 2. The downstream experiments are still using detection and segmentation models and backbones 5 or 6 years ago, the reviewer would like to see results on COCO, ADE20k and backbone like ViT, or DETR detectors. 3. The active learning which select samples from already annotated datasets is meaningless, more experiments about using FreeSel to select data from web or unannotated images are much more significant to show the effectiveness of this method.

Questions

Please see weaknesses. If authors address them, reviewer would like to change the rating.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Area Chair xpBf2023-08-16

Discussion?

Dear Reviewer yLeE, Any thoughts on authors' detailed response? Further questions / comments on the weakness points you mentioned? Thanks! Best, AC

Authorsrebuttal2023-08-21

Looking forward to more discussions and your post-rebuttal rating!

Dear reviewer, Thanks again for your suggestion to strengthen this work. As the rebuttal period is ending soon, we wonder if our response answers your questions and addresses your concerns. If yes, would you kindly consider raising the score? Thanks again for your very constructive and insightful feedback! Best, \ Authors

Reviewer 7NQc6/10 · confidence 4/52023-07-05

Summary

This work introduces a method named FreeSel, which is designed to efficiently select information-rich data points from an unlabeled pool. These selected data points can then be annotated and utilized for training deep learning models, thereby mitigating the costs and resources needed to annotate an entire dataset. The method primarily consists of two key components: feature extraction using a large-scale pretrained model (DINO), and the construction of semantic patterns based on intermediate features. Subsequently, a distance-based selection strategy is employed to identify the most diverse and informative data samples. The authors assert that the proposed approach addresses three critical aspects: Generality, by decoupling data selection from task-specific models; Efficiency, by enabling the selection of samples in a single pass; and Non-Supervision, by eliminating the need for annotations until the data selection process is complete. While the results on multiple datasets and tasks indicate that the proposed methodology holds promise, the paper seems to lack some important comparisons. Additionally, characterizing this work as an alternative to traditional active learning may not be entirely accurate.

Strengths

1. The idea of employing a large-scale pre-trained model for data selection is interesting. This is especially pertinent in today's environment, where there is a proliferation of large-scale foundational models. Harnessing these models to select data samples for training of downstream models could prove to be highly advantageous for addressing a wide range of problems. 2. The paper is well-structured and easy to understand 3. The inclusion of experiments across multiple tasks and datasets provides a comprehensive view of the empirical performance of the proposed approach

Weaknesses

1. The primary concern regarding this work pertains to its positioning. The paper seems to be more aligned with the domain of subset selection rather than active learning (AL). * AL methods, as the name suggests, actively learn which samples to select in a manner that is specific to the model at hand. While I concur with the authors' perspective that this iterative method is resource-intensive, drawing a direct comparison between the proposed "passive" selection method and traditional active learning in terms of training time may not be entirely appropriate. 2. There is a need for additional baselines to more effectively evaluate the merits of the proposed approach. * For instance, in the absence of any iterative active learning, if a random subset of data is selected and a model is trained on it, how does the accuracy of this model compare to one trained on samples selected using FreeSel? 3. There appears to be some conflation between the method (FreeSel) and the problem setting (data selection using large-scale pre-trained models). For example, the principle of Non-supervision (Line 48) seems more pertinent to the methodology than to the problem setting. 4. To gain a clearer understanding of the novelty and contributions of this work, it would be beneficial to compare the method with existing subset selection methods.Such comparisons do not necessarily need to be empirical but could include theoretical or conceptual comparisons. 5. It would be beneficial for the authors to clarify the positioning of FreeSel in relation to traditional active learning and to provide additional comparisons to strengthen the evaluation. References for consideration: 1. [Kaushal et al., Learning From Less Data: Diversified Subset Selection and Active Learning in Image Classification Tasks](https://arxiv.org/pdf/1805.11191) 2. [Chang et al., On Training Instance Selection for Few-Shot Neural Text Generation](https://arxiv.org/pdf/2107.03176) 3. [Birodkar et al., Semantic Redundancies in Image-Classification Datasets: The 10% You Don't Need](https://arxiv.org/pdf/1901.11409) 4. [Ramalingam, et al., Less is more: Selecting informative and diverse subsets with balancing constraints](https://arxiv.org/pdf/2104.12835)

Questions

1. What is the rationale for choosing CoreSet in Figure, there are much more recent and useful active learning strategies that are specifically proposed for the problem of object detection? 2. No comparisons were made with the works listed in `Data Selection with Pretrained Models` paragraph of related work though they seem very relevant, is there any specific reason for this? Including comparisons with these pertinent works could enhance the comprehensiveness and depth of the evaluation.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

All the limitations are clearly discussed

Reviewer SbPr6/10 · confidence 5/52023-07-06

Summary

This paper aims to introduce a unique method called FreeSel for data selection. This method is accompanied by a newly designed pipeline that incorporates existing general-purpose models, requiring minimal additional time. By leveraging a publicly available pre-trained model, FreeSel can select data from diverse datasets through a single-pass inference, eliminating the need for extra training or supervision. Specifically, the method involves extracting semantic patterns from intermediate features of the general-purpose model to capture nuanced local information within each image. Furthermore, it enables the selection of all data samples in a single pass using distance-based sampling at a fine-grained semantic pattern level.

Strengths

1- For the first time, this paper presents a novel model that introduces a data selection pipeline. This pipeline incorporates crucial principles of a robust feature extractor, including generality, efficiency, and non-supervision, while incurring minimal time costs. This innovative idea holds potential for application in active learning scenarios. 2- This approach leverages unsupervised extraction of intermediate features by harnessing the benefits of publicly available pre-trained vision transformer (LLM) models. These features are then utilized to define semantic patterns. 3- In contrast to existing active learning approaches, this method selects data examples for a given budget in a single pass, focusing on the diversity found in the local patterns extracted by the pre-trained model. This approach is considerably more time-effective compared to existing methods, as it avoids the need for multiple passes or iterations. 4- To demonstrate the generalizability of the proposed approach, experiments were conducted across various types of tasks, including image classification, object detection, and semantic segmentation. The results of these experiments reveal significant performance improvements, underscoring the effectiveness of the approach across different domains and applications.

Weaknesses

1- We have noticed a decrease in the performance of the proposed model as the number of samples increases, which is in contrast to several active learning approaches. We require a justification for this observation. Is the model sensitive to the size of the datasets? 2- In the paper, the variable "c" represents the number of pseudo categories, which defines the number of visual parts in an image. However, it is unclear how "c" is determined. Is it a fixed value for every image, or does it vary from image to image? 3- The proposed method relies solely on local features to identify semantic patterns, disregarding global features which can be crucial in certain cases. Completely disregarding global features may not be the most desirable approach.

Questions

Please answer all the concerns raised in the weaknesses section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have mentioned the limitations of the proposed approach in the conclusion section. The authors have also presented an ablation study to discriminate the contribution of pre-trained vision transformers (LLM) and other components. It is essential to understand that the performance of the proposed model is not only due to the capability of pre-trained LLM models.

Authorsrebuttal2023-08-21

Looking forward to more discussions and your post-rebuttal rating!

Dear reviewer, Thanks again for your suggestion to strengthen this work. As the rebuttal period is ending soon, we wonder if our response answers your questions and addresses your concerns. If yes, would you kindly consider raising the score? Thanks again for your very constructive and insightful feedback! Best, \ Authors

Reviewer xhjx3/10 · confidence 5/52023-07-14

Summary

This paper proposed a diversity-based method for active learning. Given an image, features are extracted from a pretrained visual transformer model (DeiT-S), features with [CLS] token attention above a threshold are kept and clustered into K clusters. Each cluster is represented by the average of features belonging to the cluster. The cluster center, named semantic pattern in the paper, is used in a distance-based probabilistic sampling framework (similar to kmeans++) for sample selection. The proposed method is evaluated in various visual tasks including object detection, semantic segmentation and image classification and obtained promising results.

Strengths

1. The proposed method utilized pretrained model for feature extraction and sample selection, eliminating the need to train a model at each AL cycle and thus is much more time efficient than traditional AL methods. 2. The method is evaluated on various tasks and datasets and shown promising results.

Weaknesses

1. The effectiveness of the proposed method is questionable. (1) From Fig.5 and Fig.7, it can be seen that the proposed method is always outperformed by competing methods at later stage, suggesting that the pretrained features become less effective in selecting informative samples as the training progresses. Performance at later stage is more important as active learning is only useful when your method can achieve comparable performance (e.g., 95%) of the fully-supervised baseline. (2) It is not clear how the method performs when the target domain has large domain shift to the pretrained domain. Currently the evaluation datasets are all natural images that are similar to ImageNet. I would expect that the effectiveness of the proposed method would be undermined when the domain shift becomes large. (3) Benchmarking is not comprehensive, missing BADGE [1] for image classification, EnmsDivproto [2] for object detection. 2. Technical contribution is limited. The proposed method combines existing techniques in a straightforward manner, lacking novelty and technical contribution. [1] Ash, Jordan T., et al. "Deep batch active learning by diverse, uncertain gradient lower bounds." ICLR2020. [2] Wu, Jiaxi, Jiaxin Chen, and Di Huang. "Entropy-based active learning for object detection with progressive diversity constraint." CVPR2022.

Questions

The proposed method is essentially doing class-wise aggregation to generate feature representation for an image, which has been explored in CDAL. Apart from the obvious difference that CDAL can obtain pseudo label directly from the target classifier while this paper need to employ clustering to generate pseudo labels, what are other differences and advantages of the proposed method compared to CDAL?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The authors discuss the limitation of the proposed method (i.e., cannot beat all state-of-the-art methods) in Section 6 and leave it for future work.

Reviewer xhjx2023-08-17

Dear authors and AC, I have read the rebuttal carefully. However, I still don't think the novelty and significance of the proposed method is sufficient for a conference like NeurIPS. Therefore, I will keep my original rating.

Authorsrebuttal2023-08-17

Definition of novelty and significance

Dear Reviewer xhjx, Thanks for your response. We respect your different opinions from us about the definition of novelty and significance. However, as emphasized in our rebuttal, we believe that novelty is never equivalent to technical complexity. A straightforward algorithm is always compelling as long as it has some desirable properties. Our novelty is mainly reflected by the brand-new pipeline for data selection with several superiorities. We believe a simple free data selection algorithm following a novel pipeline is much more significant for the community than many so-called "novel" and complicated traditional active learning methods. Best,\ Authors

Reviewer xhjx2023-08-18

Dear Authors, I am not saying that your work has no novelty and significance. I just don't think it is sufficient for a tier-1 conference like NeurIPS: the performance cannot beat state-of-the-art, benchmarking methods are incomplete, method-wise is just some simple combination of existing algorithms. The idea of using pretrained features for active learning is also straightforward.

Authorsrebuttal2023-08-16

Thank you and we are looking forward to your post-rebuttal feedback!

Dear AC and all reviewers: Thanks again for all the insightful comments and advice, which helped us improve the paper's quality and clarity. The discussion phase has been on for several days and we have not heard any post-rebuttal responses yet. We would love to convince you of the merits of the paper. Please do not hesitate to let us know if there are any additional clarifications that we can offer to make the paper better. We appreciate your comments and advice. Best, \ Authors

Reviewer 7NQc2023-08-19

I have carefully reviewed the initial submission, the authors' response, and the feedback from other reviewers. I appreciate the effort that has been invested in addressing the concerns raised, and I would like to thank the authors for the comparison with subset selection methods. The responses provide answerers to my questions, and in light of this, I have updated my rating accordingly. However, I would like to kindly request that the authors further refine the paper by including additional comparisons with relevant approaches. This will not only enhance the robustness of the paper but also provide readers with a clearer understanding of how the proposed method stands relative to existing solutions. Additionally, I encourage the authors to more explicitly position their work within the broader context of the field, either in the final version of this paper or in a future submission. Depending on this positioning, new baselines and comparisons may be necessary to strengthen the paper’s contributions and its distinction from other works in the field.

Authorsrebuttal2023-08-20

Thanks for your response

Dear Reviewer 7NQc, We sincerely thank you for your reply and for updating the rating. We also appreciate your valuable feedback about our paper. In the camera-ready version, we will follow your suggestions to further improve the paper writing and include additional comparisons. Best, Authors

Reviewer yLeE2023-08-21

I appreciate the authors' response. The rebuttal addressed part of my concerns. Hence, I am keeping my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC