acceptodds
Under review as a conference paper at ICLR 2027

What Does 5% Mean? The Coreset Regime Boundary Is Pinned in Samples per Class

Abstract

Coreset selection methods are usually evaluated at fixed fractions of the training set, yet the winning selection principle changes with budget. Where does this change occur, and which budget coordinate makes it stable? We measure the crossover budget m0 between a learnability probe (LFrac) and a coverage anchor (Herding), observing the reversal directly on four datasets including ImageNet-1k. Two controlled interventions distinguish the effects of pool size and visual bandwidth under fixed selection and training protocols. Shrinking the per-class pool of ImageNet-100 by 8.4× leaves m0 within 65–80 samples per class while its fraction reading drifts by roughly 10×; CIFAR-100 reproduces this approximate stability. Reducing effective pixel count by 49× on a fixed input canvas lowers full-data accuracy by 7.4 points, while m0 shifts systematically from about 66 to 80 within the same sampled budget bracket. The strongest invariance is therefore to pool size: the crossover behaves like an approximately fixed per-class sample requirement of the measured selection and learning system. This is more than unit conversion, because matching fractions can place methods in different budget regimes. Locating the boundary without downstream measurement remains open: the six dataset descriptors examined here provide no validated predictor, and naive subset mixtures on CIFAR-100 do not reliably beat the better pure endpoint. Oracle-tuned CCS improves on both endpoints near the transition, but a fixed setting transferred to ImageNet-100 loses at the smallest budget. These findings motivate reporting absolute samples per class and pool size alongside fraction in cross-dataset evaluations.

Primary area:
datasets and benchmarks
Keywords:
coreset selection, data pruning, samples per class, regime boundary, herding, learnability, ImageNet, evaluation protocol

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UjiSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies coreset selection and asks how the data budget and the selection score interact, and what role each plays. It compares two selectors. LFrac keeps the easy samples that a small proxy network learns early and stably, and Herding keeps a representative spread of samples. LFrac wins at small budgets and Herding at larger ones, and the budget where they swap places is called m₀. The authors argue that m_0 is better described in absolute samples per class than as a fraction of the dataset. With two controlled interventions, shrinking the per-class pool and degrading image resolution, and with negative results showing that m_0 cannot be predicted from dataset descriptors, that blending the selectors does not help, and that CCS helps only near m_0.

Critical strengths

(i)The paper is well written and motivated (ii)Reporting budgets as fractions is universal. Asking which coordinate is stable across datasets is a worthwhile question for the field. (iii)The pool-shrinking experiment is the strongest contribution and demonstrably interesting in Sec experiments. (iv)Highliting that simple descriptors, naive mixing, and an oracle-tuned CCS do not substitute for measuring the boundary is informative

Critical weaknesses

(i)The paper makes three kinds of claims : a phenomenon (the crossover exists), a coordinate (samples per class), and a set of negative results (no predictor, no mixing gain). It is not obvious if these three are the main contribution.

(ii)Intervention in image space" remains unclear as a motivation. Coreset method operates on a set of samples, yet the paper frames the visual-bandwidth experiment as an intervention on the image. The why is not motivated clearly. The experiment downsamples and upsamples back, so both selection and training see degraded inputs. The paper concedes that pixel count is not a calibrated measure of classification-relevant information and that the result does not falsify hypothesis (ii), that the boundary tracks per-image information. The experiment therefore cannot separate the hypotheses it was meant to separate. Its role in the argument, and why it counts as an intervention rather than a robustness check, is never made clear.

(iii)All conclusions concern LFrac vs. Herding with ResNet-18. The paper acknowledges that the numerical value of m_0 depends on the selector pair and feature source, and that the evidence does not identify its cause. Yet the title and the Sec 5 recommendations are stated in general terms. The abstract's "winning selection principle changes with budget" is a statement about principles, but two implementations are tested.

Questions for authors

(i)Which of the three contributions (phenomenon, coordinate, negative results) is the primary one? If forced to drop one, which would it be? (ii)Why is the resolution experiment described as an intervention in image space, and which of hypotheses (i)-(iii) does it actually distinguish? (iii)Does m_0 remain in the same samples-per-class band if LFrac is replaced by a confidence-, forgetting-, or EL2N-based easy-sample selector, or Herding by another coverage method?

3 comments
Reviewer_UjiSOpenReview ↗

Response

I thank the authors for their response.

Within-dataset vs. cross-dataset The boundary is set in samples per class. The pool-shrinking design varies N within a dataset. It shows that m0m_0 does not scale with the pool, but does it show that samples per class is the right unit ? The authors response own numbers (about 46 on ImageNet-1k, 65-80 on ImageNet-100, 60-66 for EL2N-easy on CIFAR-100) put m0m_0 at different values per dataset and pair. You argue that 5% is 25 per class on CIFAR-100 and about 64 on ImageNet-1k, with the crossover at about 46. Does this argument require m0m_0 to be shared across datasets, or only a dataset-specific m0m_0 that fractions fail to track?

Pool-independent instrument for EL2N EL2N-easy uses a proxy trained once on the full pool and reused for every shrunken pool. LFrac trains its proxy per pool. Does the stable band hold under the per-pool protocol for EL2N too, or does fixing the proxy test an easier setting?

Pool-independent instrument for EL2N EL2N-easy uses a proxy trained once on the full pool and reused for every shrunken pool. LFrac trains its proxy per pool. Does the stable band hold under the per-pool protocol for EL2N too, or does fixing the proxy test an easier setting?

Random vs. Herding On the smallest ImageNet-100 pool, m0m_0 lies above the largest tested budget. Is the slope of 0.08 computed with that point treated as censored, or is a value assigned to it? How would the slope change under each treatment?

Score definition For LFrac with K = 5, sis_i takes only six values. How are ties broken in per-class top-k selection, and does this affect the small-budget side of the crossover? The rebuttal raises ties as a reason to exclude forgetting counts, but does not address them for LFrac.

** Margin and the EL2N substitute** The rebuttal treats confidence as a close relative of correctness rate. Since EL2N uses softmax distance to the one-hot label, is it a margin-aware score? If so, did m0m_0 agree between LFrac and EL2N-easy on the same dataset and pool, so the two scores can be compared directly?

Held-out prediction and mechanism Can the stated unit predict m0m_0 on a held-out dataset or selector pair?

AuthorsOpenReview ↗

Response to follow-up questions (1/2)

Thank you for the follow-up. We answer the six points in order. The per-pool EL2N numbers in point 2 come from a small set of new runs (3 seeds, ResNet-18, protocol as in the paper).

1. Within-dataset vs. cross-dataset: the argument needs only a dataset-specific m0. The claim is that for a given dataset, selector pair and learner, the boundary is set in samples per class and fractions fail to track it. It does not require m0 to be shared across datasets, and the values you list (about 46 on ImageNet-1k, 65 to 80 on ImageNet-100) already differ. The 5% example uses each dataset's own boundary. The same 5% is 25 per class on CIFAR-100, on the easy-first side of its boundary, and about 64 per class on ImageNet-1k, past its boundary of about 46. Comparing methods at a matched fraction across datasets can compare them in different regimes. Inside one dataset the pool-shrinking experiment shows the same effect: on ImageNet-100 the boundary sits at about 5% of the full pool and at about half of the smallest pool. On whether samples per class is the right unit: the pool-shrinking design rules out fractions, and samples per class is the reading that stays in a band when the pool shrinks. A boundary stated per class carries over when the pool of the same dataset changes size, and one stated as a fraction does not. The paper claims no more than that.

2. Pool-independent instrument for EL2N: under the per-pool protocol the band is wider, and m0 still does not shrink in proportion to the pool. First a clarification. In the fixed rows of our first-round table, LFrac and EL2N-easy both use a full-pool proxy, so that comparison was under one protocol. A full-pool proxy gives LFrac 3 to 5 points higher accuracy on the small pools than a per-pool proxy, which strengthens the easy side and moves m0 up, away from the fraction prediction. The per-pool protocol is the stricter test, and we ran it. EL2N-easy now uses one 10-epoch EL2N proxy per pool and keeps the lowest m scores per class. Paired with Herding on CIFAR-100, N = 500 / 250 / 150 / 100 per class:

Herding extractorm0 (EL2N-easy, per-pool proxy)Fraction predictionSlope
per-pool (paper protocol)45 / 69 / >70 / 3945 / 22 / 13 / 9+0.12 (three uncensored pools)
fixed full-pool61 / 61 / 54 / 4761 / 31 / 18 / 12+0.16

Two curves cross more than once in the transition region, N = 500 in the first row (first 45, last 59) and N = 150 in the second (first 54, last 68). The others cross once or are censored. The per-pool band is wider than under the fixed proxy for the reason given to Reviewer ghPW in our first response: per-pool Herding redraws its extractor on each call and sits about one point from its competitor in the transition region, so its crossing is read with noise. With a fixed extractor all four pools fall between 47 and 61. Under both pairings the smallest pool lands at about 4 times the fraction prediction. The position and width of the band depend on the instrument, and no instrument makes m0 shrink in proportion to the pool. For comparison, per-pool LFrac paired with fixed Herding gives 59 / 50 / 37 / 49 on the same pools, a similar band from a different easy-first score.

3. Class-balanced random in place of Herding: the slope of 0.08 uses the three uncensored pools and assigns no value to the fourth. On those pools (N = 1267, 600, 300) m0 is 125, 105 and 111. On the N = 150 pool the grid ends at 90 per class and LFrac still leads class-balanced random by 0.01 points there, so the crossing is at or just above 90. Assigning 90 to that pool, the value most favourable to the fraction hypothesis, gives a slope of 0.13. Assigning 125, the largest value seen on the other pools, gives a slope of −0.01. The fraction hypothesis needs about 15 per class on that pool, and the difference is still positive at 90. Either treatment leaves the slope far from 1.

AuthorsOpenReview ↗

Response to follow-up questions (2/2)

4. Score definition: ties are broken by sort order, and they do not move the small-budget side. One correction: the shrinking-pool arms use K = 10, so the score takes 11 values (K = 5 is the window-sensitivity arm of Table 1 and Appendix A). Ties are broken by the order of the sorting routine, which is deterministic and reproducible but has no principle behind it, as Appendix C states. Ties are common at small budgets. At 25 per class on the full ImageNet-100 pool about 92% of the selected samples share the cut-point score, almost all of them correct in every epoch. At 20 per class on the full CIFAR-100 pool about two thirds do. The share falls as the pool shrinks, to about one third at 25 per class on the smallest ImageNet-100 pool, and LFrac leads Herding by 5 to 6 points at that budget on both pools. At small budgets LFrac selects a sort-ordered slice of the samples the proxy learned early. Three pieces of evidence say the slice does not matter. First, EL2N is continuous, with 49971 distinct values among the 50000 CIFAR-100 training samples, so it has almost no ties. Swapping it in for LFrac under the fixed instrument keeps the easy-first lead at small budgets, with m0 within 3 of LFrac on the two largest pools (numbers in point 5). Second, the selected members are interchangeable: LFrac selections from different proxy seeds overlap by only about 0.15 Jaccard, since the tied samples near the cut change from seed to seed, and downstream accuracy stays within seed noise (Appendix D). Which tied samples are taken does not change the result. The early-learned composition does. Third, the K = 40 arm of §4.1 has 41 score values and far fewer ties, and it gives the same crossing reading as K = 10, about 85 on the full CIFAR-100 pool with the Herding features of that section (Appendix A).

On forgetting: both scores have ties at the easy end, and the difference is the size of the tied block. For LFrac on ImageNet-100 the samples correct in every epoch are about a tenth of the median class, so a small budget taken from that block is still a draw from the easiest tenth. For forgetting computed from the same 10-epoch proxy, two thirds of the samples in the median class have zero forgetting events. A small budget taken from that block is close to a random draw from most of the class. Forgetting was excluded for lack of ordering power at those budgets.

5. Margin and the EL2N substitute: EL2N is margin-aware, and its m0 agrees with LFrac's on the two largest pools and comes earlier on the smallest, where EL2N-easy is the less accurate selector. EL2N is the L2 distance between the softmax vector and the one-hot label. It falls as the true-class probability rises and grows with the mass on wrong classes, so it carries margin information and measures a different quantity from LFrac's count of correct epochs. On the same dataset, pool and fixed instrument:

N per class500250150100
m0, LFrac5963>70>70
m0, EL2N-easy6166>7060
LFrac minus EL2N-easy accuracy, mean over budgets−0.4+0.4+1.7+2.9

The claim does not require the two scores to give the same m0, because the value of the boundary depends on the selector pair. The table shows how. On the two largest pools the two selectors are within 0.4 points in accuracy and their crossings with Herding agree within 3. On N = 150 both are still ahead of Herding at the largest budget, so that pool cannot order the two crossings. On the smallest pool LFrac is 2 to 4 points more accurate than EL2N-easy across the budget grid. Both are compared with the same Herding runs, so LFrac stays ahead of Herding for longer and its crossing comes later. A stronger easy side moves the boundary up, as in our general response. Under either score the smallest pool crosses far above the 12 per class that a fixed fraction predicts.

6. Held-out prediction: the unit predicts how m0 responds to pool size, not its value. The prediction is that on a new dataset or selector pair, shrinking the pool does not pull m0 down in proportion, so the fraction reading drifts. The paper contains one out-of-sample test: ImageNet-100 was run first and CIFAR-100 was a later replication under the same protocol, with the predicted outcome. Each pairing added in the first round (fixed instruments, EL2N-easy, class-balanced random, VGG-16-BN) was a further test. In each a proportional drop was a possible outcome, and it did not occur. The value of m0 is not predictable at present. That is the negative result of Appendix E, and the paper says the boundary has to be measured per dataset.

The new per-pool EL2N numbers are now in the updated PDF (Appendix I, Table 8). The questions led to runs we had not done, and we will answer anything further during the remaining discussion period.

Reviewer_ghPWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper investigates the evaluation coordinate for coreset selection budgets. While literature almost universally compares coreset methods across datasets at fixed fractions ff of the training pool, the authors argue that the regime boundary between "easy-first" selection (measured via LFrac, the fraction of epochs a sample is correctly classified over a short KK-epoch proxy run) and "coverage/diversity" selection (measured via Herding) is actually pinned in absolute samples per class (m≈f⋅Nm \approx f \cdot N), not fraction ff. Defining the crossover budget m0m_0 as the first zero-crossing of Δ(m)=acc(LFrac)−acc(Herding)\Delta(m) = \text{acc}(\text{LFrac}) - \text{acc}(\text{Herding}) with a ResNet-18 learner, they first locate m0m_0 across six datasets (directly observing the flip on ImageNet-1k at m0≈46m_0 \approx 46, ImageNet-100 at m0≈66m_0 \approx 66, CIFAR-100 at m0≈85m_0 \approx 85, and CIFAR-10 at m0≈222m_0 \approx 222, while CUB-200 and TinyImageNet remain right-censored). They then perform two controlled data interventions: (1) shrinking the per-class candidate pool NN on ImageNet-100 by 8.4×8.4\times (1267→1501267 \to 150) and on CIFAR-100 by 5×5\times (500→100500 \to 100), showing that m0m_0 stays within a relatively narrow absolute band (65–8065\text{--}80 on ImageNet-100; 36–4336\text{--}43 on CIFAR-100) while the crossover fraction f0=m0/Nf_0 = m_0/N drifts by 10×10\times and 4×4\times, respectively; and (2) downsampling/upsampling ImageNet-100 images (2242→322224^2 \to 32^2, a 49×49\times pixel reduction), which drops full-data accuracy by 7.47.4 points but shifts m0m_0 only modestly from 6666 to 8080. Finally, Section 4.4 reports negative results showing that six simple dataset descriptors fail to monotonically order m0m_0 across datasets, naive subset mixtures of LFrac and Herding/EL2N on CIFAR-100 exhibit anti-synergy, and Coverage-Centric Coreset Selection (CCS) only beats the pure endpoints inside the transition zone around m0m_0 when its cutoff hyperparameter β\beta is oracle-tuned.

Critical strengths
  • Convincing, well-controlled falsification of the fixed-fraction convention (Section 4.2): The nested pool-shrinking experiment on ImageNet-100 (N∈{1267,600,300,150}N \in \{1267, 600, 300, 150\}) and CIFAR-100 (N∈{500,250,150,100}N \in \{500, 250, 150, 100\}) is cleanly designed and directly answers a practical methodological question. Showing that the log-log slope γ\gamma of m0m_0 against NN is near 00 (−0.07-0.07 to +0.10+0.10) rather than 11, supported by the candidate-overlap check (Appendix F.2b) and fixed-mm random baseline check, makes a rock-solid case that comparing coresets at fixed ff across datasets with different per-class pool sizes NN confounds distinct regime behaviors.

  • Thorough empirical hygiene and honest negative results: Backed by 1,847 runs across multiple seeds and proxy windows K∈{5,10,20,30,40}K \in \{5, 10, 20, 30, 40\}, the authors do not oversell their findings. Explicitly reporting that naive LFrac+Herding/EL2N mixtures fail to beat pure endpoints and that transferring a fixed CCS cutoff (β=0.4\beta=0.4) from CIFAR-100 to ImageNet-100 fails at low absolute budgets (m=25m=25, losing by 3.43%3.43\%) is genuinely informative for practitioners.

Critical weaknesses
  • Confounding of proxy training steps with pool size NN, and fragile "instrument sensitivity" of m0m_0:

    • Inconsistent m0m_0 on the exact same dataset (8585 vs. 43.243.2 on CIFAR-100): In Section 4.1 (Figure 1), full-pool CIFAR-100 (N=500N=500) has a crossover of m0=85m_0 = 85. Yet in Section 4.2 (Table 2), full-pool CIFAR-100 (N=500N=500) has m0=43.2m_0 = 43.2, nearly a 2x shift on the exact same dataset! The authors attribute this in lines 302-307 to changing the "instrument" (different Herding feature extractor and budget grid), and in Appendix F.4 we see that the CIFAR-100 gap curve oscillates in sign multiple times between m=40m=40 and m=60m=60 (e.g., at N=500N=500: +2.14+2.14 at m=40m=40, −1.17-1.17 at m=45m=45, +1.58+1.58 at m=55m=55, −2.02-2.02 at m=60m=60). If changing the feature extractor training duration or grid resolution shifts m0m_0 by 2×2\times (43→8543 \to 85) and induces multiple sign oscillations, the claim that m0m_0 is a tightly "pinned" boundary is overstated, especially when the Herding extractor itself is trained from scratch on the shrunken pool NN (meaning a smaller pool N=100N=100 produces a weaker feature extractor for Herding than N=500N=500).

    • Fixed epochs vs. fixed gradient steps across NN and mm: In all experiments, both the KK-epoch proxy and the 90/200-epoch downstream ResNet-18 are trained for a fixed number of epochs at a fixed batch size. Consequently, when the pool NN shrinks by 8.4×8.4\times (in the proxy) or when mm is small (in the downstream learner), the network receives 8.4×8.4\times fewer gradient updates. It is unclear how much of LFrac's advantage at small mm is a fundamental property of the data distribution versus an optimization artifact of under-training ResNet-18 (too few gradient steps at 90/200 epochs when mm is tiny, where easy/low-gradient-variance samples converge faster).

  • Single downstream architecture (ResNet-18) and limited selector scope:

    • Every experiment in the paper uses a single downstream architecture (ResNet-18 trained from scratch) and defines the regime boundary almost exclusively through a single pair of selectors (LFrac vs. Herding). However, modern coreset selection and data pruning frequently use stronger self-supervised/pretrained feature extractors (e.g., DINOv2/CLIP) for coverage or evaluate across different model scales (e.g., ResNet-50, ViT) and training-dynamics scores (e.g., GraNd, Forgetting, Moderate Coreset, D2 Pruning). Notice that on CUB-200, the only dataset where the authors used frozen DINOv2 features for Herding, and on TinyImageNet, no crossover is observed at all up to f=0.70f=0.70.
Questions for authors
  • Decoupling Herding feature quality from pool shrinking (NN): In Section 3.3 and 4.2, when you shrink the pool NN from 1267→1501267 \to 150 on ImageNet-100 (or 500→100500 \to 100 on CIFAR-100), you re-train the Herding feature extractor from scratch on the shrunken pool. Doesn't shrinking NN degrade the representation quality of the extractor used by Herding, potentially penalizing Herding at small NN? What happens to the pool-shrinking curve if Herding uses a fixed, pool-independent feature extractor (e.g., frozen DINOv2 or the full-pool extractor) across all NN?

  • Fixed epochs vs. fixed training iterations: When m=25m=25 on ImageNet-100 (2,5002{,}500 total images, batch size 256, ∼10\sim 10 steps/epoch), 90 epochs is only ∼900\sim 900 gradient steps. If you train the downstream ResNet-18 for a fixed number of gradient steps (matching the step budget of larger mm) rather than a fixed 90/200 epochs, does LFrac still beat Herding below m0m_0, or does Herding catch up once given enough optimization steps to fit harder, more diverse examples?

  • Resolving the CIFAR-100 discrepancy (m0=85m_0 = 85 in §4.1 vs. m0=43.2m_0 = 43.2 in §4.2): Could you clarify the exact difference in the ResNet-18 Herding feature extractor between §4.1 (K=40K=40 proxy features, giving m0≈85m_0 \approx 85 even when LFrac uses K=10K=10) and §4.2 ("native R18 features", giving m0=43.2m_0 = 43.2)? Why does the Herding feature source shift m0m_0 by 2×2\times on the same dataset while changing image resolution by 49×49\times in §4.3 only shifts m0m_0 by 1.2×1.2\times?

  • Generalization beyond ResNet-18 and LFrac/Herding: Does the same crossover m0∈[65,80]m_0 \in [65, 80] on ImageNet-100 hold if the downstream learner is changed (e.g., ResNet-50 or a ViT/ConvNeXt), or if Herding is replaced by another coverage/diversity method (e.g., Moderate Coreset or k-CenterGreedy) and LFrac by Forgetting or GraNd/EL2N?

Reviewer_byV4OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper considers two coreset selection methods LFrac and Herding. Using these two methods the paper setups up the problem statement as to what is the right unit to test coreset budget is. The paper defines a budget m0m_0 as the per-class budget at which LFrac no longer outperforms Herding. The paper presents results on vision datasets, specifically CIFAR-10, CIFAR-100, ImageNet-100 and ImageNet-1K. The paper concludes by recommending the use of samples per class along with the fraction.

Critical strengths

The recommendation of using more metrics such as samples per class to report the performance of coresets is supported by simplified experiments.

Critical weaknesses
  • My first concern is the claim

Coreset selection methods are usually evaluated at fixed fractions of the training set, yet the winning selection principle changes with budget.

The paper relies on this statement but does not clearly demonstrate that methods change for vision or language settings with newer state of the art coreset selection methods. Assuming this statement is true the other issue is that Budget-adaptive selection [1] already targets and addresses this issues. Other literature such as BOSS [2] DUAL [3] are also missed.
  • The choice of Lfrac and Herding is not justified and the budget is likley to depend on the choice of selection methods (the paper already states this). This makes the paper incomplete in terms of analysis.

  • The paper evaluates on vision and 1 architecture and hence is hard to generalize the conclusions provided. Assuming they do generalize the claims are that more metrics need to be reported. The lack of novelty, the lack of experimental rigor, the fact that this core premise has been tackled by specific methods a few years ago and that is not presented in the paper makes me question the soundness of this paper.

[1] Choi, H., Ki, N., & Chung, H. W. BWS: best window selection based on sample scores for data pruning across broad ranges ICML 2024
[2] Acharya, Abhinab, et al. "Balancing feature similarity and label variability for optimal size-aware one-shot subset selection." ICML 2024
[3] Cho, Yeseul, et al. "Lightweight dataset pruning without full training via example difficulty and prediction uncertainty." ICML 2025.

Questions for authors
  1. Why Lfrac and Herding?
  2. How is this paper's observations different from BWS [1]?
4 comments
AuthorsOpenReview ↗

Follow-up to Reviewer byV4

Dear Reviewer byV4,

We are writing in case our response was missed. It takes up your two questions directly.

On BWS, BOSS and DUAL: these methods establish that the best selection changes with the budget, and the revised related work now says so. None of them changes the size of the pool, so their designs do not test whether that change follows the fraction or the number of samples per class. Our pool-shrinking experiment tests exactly that question.

On the choice of LFrac and Herding: Appendix Table 7 now reports the seven-method benchmark behind the choice. The new Section 4.3 repeats the pool-shrinking test with a second easy-first score, with class-balanced random selection, and with a second learner. The discussion with the other reviewers has since added further instrument pairings (Appendix I, Table 8). In none of them does the boundary shrink in proportion to the pool.

If any concern remains, we would be glad to answer it before the discussion closes on November 18. If the response and the revision resolve your concerns, we would be grateful if you could update your assessment.

AuthorsOpenReview ↗

Revised PDF with new content marked in blue

Dear Reviewer byV4,

We understand that this is a busy period, and we appreciate the time you have given to our paper. We have uploaded a version of the revised PDF in which the new sections and paragraphs are marked in blue, so that the parts answering your review are quick to locate.

  1. Budget-adaptive selection (BWS, BOSS, DUAL): the blue paragraph in Section 2.
  2. The choice of LFrac and Herding: Appendix H with Table 7, the seven-method benchmark behind the choice.
  3. The single selector pair and the single architecture: the new Section 4.3 with Table 3, which repeats the pool-shrinking test with a second easy-first score, with class-balanced random selection, and with VGG-16-BN.

If a concern remains, we would be glad to answer it. If these parts resolve your concerns, we would be grateful if you would consider updating your assessment before the discussion closes.

Sorry for the delayed response, some of my queries were answered I will raise my score to 2.

AuthorsOpenReview ↗

Response to Reviewer byV4

Dear Reviewer byV4,

Thank you for reading the response and the revision, and for raising your score. We appreciate it.

We would also be glad to answer any remaining question before the discussion closes.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.