Towards Accelerated Model Training via Bayesian Data Selection

Mislabeled, duplicated, or biased data in real-world scenarios can lead to prolonged training and even hinder model convergence. Traditional solutions prioritizing easy or hard samples lack the flexibility to handle such a variety simultaneously. Recent work has proposed a more reasonable data selection principle by examining the data's impact on the model's generalization loss. However, its practical adoption relies on less principled approximations and additional holdout data. This work solves these problems by leveraging a lightweight Bayesian treatment and incorporating off-the-shelf zero-shot predictors built on large-scale pre-trained models. The resulting algorithm is efficient and easy to implement. We perform extensive empirical studies on challenging benchmarks with considerable data noise and imbalance in the online batch selection scenario, and observe superior training efficiency over competitive baselines. Notably, on the challenging WebVision benchmark, our method can achieve similar predictive performance with significantly fewer training iterations than leading data selection methods.

Paper

Similar papers

Peer review

Reviewer UrxY5/10 · confidence 3/52023-07-06

Summary

This paper followed up reducible hold-out loss selection (RHO-LOSS) method for data selection and improved it by proposing a reasonable approximation for the non-trivial objective function and eliminating the need for extra hold-out data.

Strengths

1. Authors present both theoretical results (the lower bound for hard-to-estimate objective) and empirical results (algorithm and experiment results).

Weaknesses

1. I am not fully convinced by the paper that their proposed method out-performs existed SOTA, namely RHO-LOSS much, specifically from the results in table 1 and figure 1. I wonder if authors can provide more intuitive comparisons between the proposed method the RHO-LOSS, either empirically show effectiveness and efficiency of the proposed method, or theoretically show the proposed lower bound is more solid than approximation used by RHO-LOSS. 2. The ablation study needs more experiment result (more datasets and more model structure) to get valid conclusion. I also wonder why compare with baseline instead of RHO-LOSS.

Questions

In addition to questions addressed in weakness section, I have following questions for the authors: 1. How is target accuracy selected? Specifically for table 3, why select 30% and 40% as target accuracy? 2. I am curious if each training epoch for RHO-LOSS and your proposed method take same amount of time considering they have different data selection steps? 3. It seems the proposed method performs well on classification datasets with large number of classes (CIFAR 100, ILSVRC12 and WebVision datasets) compared with dataset with smaller number of classes (CIFAR 10). I wonder authors can share some ideas from the perspective of the designed algorithm about this. 4. I wonder if the authors have look into how zero-shot predictor affects the method performance. In other word, is the proposed method robust to poorly-performed predictor? If not, how much will the model be impacted?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, authors have addressed their limitation, namely the method performance hugely depends on effectiveness of the zero-shot predictor.

Reviewer sAWs7/10 · confidence 4/52023-07-08

Summary

The paper builds on recently proposed work that accelerates training through online batch selection using generalisation loss as selection criterion, by using LaPlace approximation for a stronger Bayesian approximation and using off-the-shelf pre-trained models. The paper presents theory deriving their selection function and then demonstrating its efficacy on a number of datasets in the vision setting.

Strengths

- Novelty on top of RHO-Loss: (1) using pre-trained models as validation model, (2) LaPlace approximation for stronger Bayesian approximation than point-estimate models. - Well-written paper, that is clean and easy to read. - Good ablations for how to tune the introduced zero-shot predictor as a validation model. - Decent performance gains over baselines.

Weaknesses

- The novelty over RHO-Loss is appreciated, but it should be corrected in the narrative that RHO-Loss does not require access to clean holdout data. This is incorrect as far as I can tell, and Mindermann et al. note this in their paper. - Would be nice to see additional datasets and settings as per the original paper (Language and Clothing1M, however WebVision seems to be a bigger/better dataset) - In the practical acceleration/algorithm section they note that it's not much slower than [31]. I think a more empirical measurement of the difference would be appreciated, further more how much time does this method add over normal training? Does it really accelerate training in terms of time? - It's not clear whether the zero-shot predictor or the laplace approximation is providing the gain? An ablation here seems necessary. If it's just the zero-shot predictor, it's seems like there's not much of a difference here besides slapping a pre-trained model on as a validation model?

Questions

As per above.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

N/A

Reviewer GG6S6/10 · confidence 4/52023-07-27

Summary

This paper studies data selection methods because training examples may be of different importance/quality. By selecting a subset of high-quality/high-usefulness examples, the model’s performance can be improved when training on this subset. In Particular, the paper proposed a method by leveraging a light-weight Bayesian treatment and incorporating off-the-shelf zero-shot predictors.

Strengths

1. The paper is clearly motivated. 2. The proposed method is well explained. 3. The experiments cover different baselines, models and datasets.

Weaknesses

The problem of data selection has been studied extensively in the ML community, which verifies the importance of this problem. Some recent works ([1,2] from reference list below) proposed data selection methods that do not only focus on hard/easy examples, but also consider data distribution and its impact on the overall loss (e.g., see Theorem 1 from [2]). It looks like these methods may already have partially addressed the problem studied in this paper. It would be better if the authors could include some discussion about these methods and possibly compare against them in the experiments. Reference: [1] Xia, Xiaobo, et al. "Moderate coreset: A universal method of data selection for real-world data-efficient deep learning." ICLR 2023. [2] Zheng, Haizhong, et al. "Coverage-centric Coreset Selection for High Pruning Rates." ICLR 2023.

Questions

Please see my comments in Weaknesses

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

None

Reviewer agRJ7/10 · confidence 4/52023-07-29

Summary

This works is situated in the field of **robust generalization**; in particular it studies the problem of how models that were trained on noisy or imbalanced data perform on clean data. They achieve this via **online batch selection**, and in particular they contribute to the development of a **Bayesian framework** for searching for the best datapoints to use in each mini-batch among some canditate points sampled from the training set. In more details, they extend a SotA method, called **RHO-LOSS** [1]. The authors pinpoint some approximation choices that [1] makes in order to make the principled Bayesian approach to data selection practical. As a result, they design an algorithm which, in contrast to [1], they claim 1) to make **less crude approximations about predictive posteriors**, by using variational Laplacian approximations, MC sampling (and other) instead of pure point estimates, and 2) **not to rely on the existence of oracle clean data** during the training of their model. Their experiments demonstrate **improvements in robust generalization and training acceleration for image classification tasks** over other methods in the literature (incl. [1]). [1] Mindermann, Sören, et al. "Prioritized training on points that are learnable, worth learning, and not yet learnt." International Conference on Machine Learning. PMLR, 2022.

Strengths

The authors contribute to an important direction for robust training methods, in particular developping a bit further the Bayesian data selection framework. 1. The paper is well-written and motivated. The derivations are easy to follow and the deferrals to the appendix are used appropriately. Presentation can further be improved if there was somewhere a paragraph (perhaps in the appendix) with the list of approximation steps taken (for future reference), e.g. MC sample of a variational posterior, based on Laplace approximation of the last layer of a neural network (for which the backbone is taken as a point estimate), and the Laplace approximation further does not use the true Gaussian, but a Gauss-Newton matrix + potentially a KFAC approximation of it. 2. Authors identify correctly potential improvement points of [1], and proceed in mitigating them. In this way, they introduce a novel approximation scheme to the Bayesian framework of data selection, which is effective and computationally-lightweight. 3. Experiments seem to advocate in favor of using their method over [1] or other methods, they are extensive and on various dataset scales in image classification tasks.

Weaknesses

1. Proposed method overly claims that it does not use oracle clean data, however it might still depend on clean data via the unsupervised pre-trained zero-shot classification proxy that they use. In particular, CLIP-R50 might have been trained in a superset of datasets (like ImageNet) which contains CIFAR10/100. In that case, information from oracle clean data has been stored in the pretrained model. While I believe that this is a weaker assumption, there needs to be an explicit statement of this assumption and a quantitative assessment of this potential pitfall. To which extent would a pretrained model on the same noisy/imbalanced dataset as the benchmark task be useful? Nonetheless, the authors acknowledge this potential limitation of an underperforming pretrained model at the last section. The following (lightweight) experiments can be further performed to augment their arguments or awareness of the limitation: - A zero-shot CLIP baseline is provided, I think they should also provide with a finetuned version using linear probing and training on imbalanced/noisy CIFAR10/100 with uniform sampling. This way it will be more clear that the improvement are more due to their method and not because of an overpowered pretrained model. - Consider CIFAR10/100 experiments using an unsupervised pretrained model, with linear probing or kNN (so that it is zero-shot if you want that) for the oracle model, via for instance some simple SSL method like MoCov2 [2] on the same noisy/imbalanced training set. I will increase the assessed score if such experiments are performed and reported. 2. Ablation study reveals sensitivity to some hyperparameters. As a result, model selection and hyperparameter tuning are important for the success of the method. How was model selection performed? Was the validation split iid to the training set or was it a clean/balanced dataset? This needs to be reported clearly, but not addressed in this paper. 3. It would be nice to have a more extensive ablation study on the effects of the approximation choices in terms of final test accuracy, spanning from crude approximation schemes to the one finally used by the authors. [2] Chen, Xinlei, et al. "Improved baselines with momentum contrastive learning." arXiv preprint arXiv:2003.04297 (2020).

Questions

### Questions * (related to weakness 2) How was model selection performed? * How were target accuracies chosen for training acceleration results? * Lines 151-152: What do we miss by replacing the Hessian matrix in the Laplacian approximation by the Gauss-Newton? * Lines 131-132 about “the recent trend of exploring the potential of pretrained models” (in robust generalization) needs some citations. * What does Table 4 in the Appendix refer to? Please complete the appendix to explain. ### Typos * y-axis on Fig.2,3 and 4 are not in % of test accuracy as indicated.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

The authors derive a novel approximation scheme to Bayesian data selection for training robust models (Strength 2), with favorable results over past literature in image classifiacation tasks (Strength 3) and the writing is overall great (Strength 1). There is a concern about the applicability of the method in cases where a pretrained model does not exist for the task at hand, in which case we need to find another proxy or probably pretrain it (Weakness 1), and how much actually the method makes no use of priviledged clean/balanced datasets for training (Weakness 1) and validation (Weakness 2). The limitation about the pretrained model is mentioned briefly, but it would be appreciated if it is expanded and experimented on with the ablation studies suggested in the Weakness section above.

Reviewer agRJ2023-08-11

Thank you for your responses

**Q1**: Regarding the point of “superset” dataset, the web-scale collection of image-text pairs almost certainly is exposed to images of CIFAR10 categories (airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks). The authors can try to disprove this. The more CLIP, BLIP and SWAG are trained in larger-scale in-the-wild datasets, the more difficult it is to claim zero-shot generalization to ood data. This is not up to the authors to disprove/approve, they just have to provide a convincing enough ablation, that in light of this their method still provides with further benefits. Thank you for the linear probing experiments, I would have been more convinced if you provided a *finetuned version* of the R50-CLIP network instead. The second experiment, however, provides with enough ablation against RHO-Loss. **Q2**: Please be clear about the validation protocol, it is important to clarify (if not address with the same resources as training). **Q7**: Please clarify in the appendix as well Table 4, since there is no text accompanying it. Thank you for answering the rest of my questions, I raise my score to 7.

Authorsrebuttal2023-08-13

Thank you for your feedback

We appreciate your detailed comments and suggestions. Regarding the zero-shot predictor, we totally understand what you're trying to convey. As suggested, we provide a further baseline—the finetuned version of the R50-CLIP, with the results detailed below: | | CIFAR-10 | CIFAR-10* | CIFAR-100 | CIFAR-100* | | ------------------------------------ | -------- | --------- | --------- | ---------- | | Fine-tuning R50-CLIP with uniform sampling | 87.6 | 85.3 | 59.1 | 57.6 | | Proposed | 91.4 | 91.3 | 63.3 | 61.4 | We will incorporate your suggestions on validation protocol and Table 4 in the final revision and continually polish our paper. Thank you again!

Reviewer UrxY2023-08-16

I thank the authors for their detailed responses to my reviews. After carefully reading their response, I decide to raise my score to 5.

Authorsrebuttal2023-08-16

Thanks

Thank you for the raised score. We will carefully revise our paper to include the discussions in the rebuttal. Thank you again!

Reviewer sAWs2023-08-18

Response

Q1-Q3. Noted. Q4. So perhaps I'm slightly confused here, but it seems like all of the gain is from subbing in a stronger model (zero-shot predictor), nothing from what is actually proposed in this paper (in fact it loses performance in two cases shown). From RHO-LOSS paper this seems to be an obvious perhaps even implied step? If my understanding is incorrect, then the experiment I asked for remains: RHO-LOSS with CLIP/zero-shot predictor vs RHO-LOSS and proposed model.

Authorsrebuttal2023-08-18

Thanks for your reply

We apologize for any confusion caused by our previous reply. We would like to provide further clarification regarding the method **Proposed - zero-shot predictor + validation model from RHO-Loss**. This method specifically refers to the variant of our method that utilizes the validation model of RHO-Loss to **take the place of the zero-shot predictor**. It is important to note that the key distinction between this method and RHO-Loss lies solely in the selection principle. The significant disparities in training speed and final performance serve as evidence that our selection principle is indeed effective. We acknowledge that including a baseline that combines RHO-Loss with a zero-shot predictor could be beneficial in double-checking this point. The corresponding results of RHO-Loss based on our codebase are listed below: |Method | Epochs to reach 40.0% ACC $\downarrow$ |Epochs to reach 52.5% ACC $\downarrow$ |Final ACC. (%) $\uparrow$ | | -------- | -------- |-------- |-------- | | RHO-Loss w/ zero-shot predictor (CLIP-RN50)| 59 | 92 | 58 | | RHO-Loss w/ zero-shot predictor (CLIP-ViT-B/16)| 53 | 86 | 60| | Proposed | 32 | 53| 63| Given these, we would like to ask the reviewer to re-evaluate our contributions. We welcome any further comments.

Authorsrebuttal2023-08-18

Thanks

Thank you very much! We will carefully revise our paper to incorporate the discussions and results in the rebuttal. Thank you again!

Reviewer sAWs2023-08-18

One other concern

Just a brief lookover again. Are these the results of a single seed? It seems like that is the case, as only in figure 1 is it reported that multiple runs are used. Multiple runs should be used in evaluating the results as these algorithms are not the most robust from my experience.

Authorsrebuttal2023-08-19

Regarding random runs

Sorry for the lack of clarity. The results in the tables of our paper correspond to the average over 3 random runs. The results in the rebuttal are from a single seed and we did not engage in cherry-picking. We have empirically observed that our performance remains relatively stable across different runs. We will be careful about this point when revising our paper. Thanks for your kind advice!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC