acceptodds
Under review as a conference paper at ICLR 2027

WASSAL: Wasserstein Distance-Based Soft Subsetting for Active Learning

Abstract

We introduce Wasserstein Distance-Based Soft Subsetting for Active Learning (WASSAL), an approach that enhances active learning for image classification by explicitly using distributional information when selecting samples for expert labeling. Traditional AL methods often overlook this perspective and rely mainly on model uncertainty or geometric diversity. The key to our methodology is a Weighted Wasserstein adaptation formulation, which learns class-specific weights for the unlabeled samples so that the weighted samples are distributionally close to the labeled datapoints of a given class. WASSAL employs these learned weights in two ways: (i) samples with the least weights across classes are treated as most confusing from a distributional perspective and form the query set for expert labeling; (ii) samples with the highest weights are treated as typical/landmark examples of their respective classes and can be soft-labeled without expert consultation. We propose a novel weighted loss term involving these soft-labeled examples in the next training round. Hence WASSAL not only leverages expert knowledge on the few queried examples, but also utilizes information implicitly available in non-queried examples. Our formulation extends the Wasserstein core-set view of Mahmood et al. (2022) by making the matching class-conditional. Across five benchmarks, WASSAL and WASSALWITHSOFT lead every budget on Caltech-101 and STL-10, remain within a few points of the best baselines on CIFAR-10 and PneumoniaMNIST, and hold their own under the high-variance SVHN pool, showing that distributional soft subsetting is a practical addition to the active learning toolbox.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Active Learning, Image Classification, Wasserstein Distance, Soft Labeling, Distributional Matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pfHPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes WASSAL, an active learning method built on a class-conditional Wasserstein distance. For each class, the method learns weights over the unlabeled samples so that the weighted unlabeled distribution matches the labeled data distribution of that class. Samples with low weights across all classes are queried for expert labeling, while high-weight samples are assigned pseudo-labels and incorporated into training through an additional weighted loss.

Critical strengths
  1. The overall motivation is reasonable: using the distribution of the existing labeled examples to guide sample selection is an interesting direction.
  2. The paper compares against a reasonably broad set of baselines covering uncertainty-based, gradient-based, and representation-based active learning methods, and reports results both with and without the proposed weighted soft loss.
Critical weaknesses
  1. I am not convinced of the novelty of the Wasserstein based weighting formulation, as the proposed class conditional weighting appears closely related to conventional nearest neighbor or kernel based similarity. In Eq. (4), the unlabeled sample weights are optimized over the probability simplex without any additional constraints. For the unregularized Wasserstein objective, this optimization appears to reduce to assigning each labeled example to its nearest unlabeled point. Under the plan entropy regularization in Eq. (3), it seems to yield a corresponding distance based softmax assignment. This raises the question of what the Wasserstein optimization captures beyond simpler similarity based weighting. Since the paper does not compare against nearest neighbor or RBF kernel based weighting, it is difficult to assess whether the additional optimization is necessary.

  2. The interpretation of the learned weights is also not sufficiently justified. The paper assumes that samples with low weights across all classes are confusing or informative, whereas high weight samples are reliable representatives of their classes, but I do not think these interpretations follow directly from the optimization objective. A sample may receive low weights because it lies far from all existing labeled examples rather than because it is close to a decision boundary. Such samples could be outliers or examples from poorly represented regions, and it is unclear whether selecting them necessarily improves classification.

Similarly, the class wise weights are normalized across unlabeled samples rather than across classes for each sample, so they cannot be directly interpreted as class probabilities or classification confidence. This matters because high weight samples are used to generate pseudo labels, but the paper reports neither pseudo label accuracy nor any analysis of how the learned weights relate to pseudo label correctness. Without such evidence, it is difficult to judge whether the weights provide a reliable signal for either acquisition or pseudo labeling.

In addition, the paper does not clearly report how many pseudo labeled samples are used during training. Eq. (6) selects the "top few" samples for each class, but the selection size or ratio is not specified. Since the weighted soft loss is a main contribution, I think the authors should report the pseudo label coverage, the ratio of pseudo labeled to expert labeled samples, and how these quantities vary across datasets and acquisition rounds. Some analysis would also help clarify whether the improvement comes from reliable Wasserstein based pseudo labeling or simply from using additional unlabeled examples.

Questions for authors

Q1. Relation to similarity based weighting

Which regularized OT objective is used to solve Eq. (4) in practice (e.g., the plan-entropy form in Eq. (3), a KL penalty with respect to the product of the marginals, or a debiased Sinkhorn divergence), and with what regularization strength? Given that the optimal weights appear to reduce to a nearest neighbor assignment without regularization and to a distance based softmax assignment with plan-entropy regularization, could the authors compare WASSAL against nearest neighbor and RBF kernel weighting baselines that use the same selection and pseudo labeling rules?

Q2. Characteristics of the queried samples

Why should low weights across all classes indicate informative samples rather than outliers or samples far from all labeled examples? Could the authors analyze the queried samples, for example by reporting their distance to the nearest labeled example and their overlap with the samples selected by CoreSet?

Q3. Quantity and quality of pseudo labels

  • How many pseudo labeled samples are selected per class in Eq. (6), and what is the ratio of pseudo labeled to expert labeled samples across datasets and acquisition rounds?
  • Could the authors report pseudo label accuracy as a function of the learned weights or of the number of selected samples?
  • How are samples that receive high weights for multiple classes handled?

Q4. Source of the soft loss gains

How does WASSALWithSoft compare with standard confidence based pseudo labeling or self-training under the same annotation budget and training protocol? In Appendix B, are the pseudo labels for the other query rules also obtained from the Wasserstein weights in Eq. (4)? Finally, what explains the decrease in average accuracy on four of the five datasets when the soft loss is added?

1 comment

We thank the reviewer. The additional experiments clarify the OT objective, query behaviour, and source of improvement.

Q1. OT objective; NN/RBF comparisons. Our implementation uses debiased Sinkhorn divergence (ε=0.0025), with 10 projected Adam steps initialized from uniform weights. We agree that the exact LP yields 1-NN assignments, while entropic transport yields distance-softmax weights. We compared these alternatives using identical query, S2 selection, and loss rules (Eqs. 5–7).

Caltech-101, budget 100, 3 seeds:

WeightingAccuracyS2 label acc.S2/labeled
Sinkhorn24.04±1.9414.2%2.1
1-NN (exact LP)22.55±1.5116.0%0.55
RBF (median bandwidth)21.38±2.2316.3%0.97
No soft loss18.00±1.41––

Sinkhorn gains 1.49 points over 1-NN and 2.66 over RBF. Although NN/RBF yield slightly more accurate pseudo-labels, Sinkhorn selects a 2–4× larger soft subset covering all classes. We will clarify that the transport step constructs a large, class-covering training subset.

Q2. Query characteristics. At budget 100, queried points lie at the 87th percentile of nearest-labeled-point distance on Caltech-101 and 64th percentile on STL-10. They cover relatively distant regions but are not simply the most distant points. Overlap with CoreSet queries is 24% and 6%, respectively.

Without soft loss, the query rule performs comparably to Random (Caltech-101: 18.00 vs. 19.38; STL-10: 26.97 vs. 25.71). We agree that low weights should not be interpreted as proximity to decision boundaries. We will instead describe the rule as coverage-oriented.

Q3. Pseudo-label quantity and quality. S2 includes top-weight points covering up to 80% of each class's mass, capped at 30/class (Caltech-101) or 400/class (STL-10). Points selected by multiple classes are assigned to their highest-weight class only.

The S2/labeled ratio is 2.1 on Caltech-101 (2.4 in round 1; 1.5 in round 7) and 3.7 on STL-10.

Pseudo-label accuracy is 14.2% on Caltech-101 (chance 1%; model 10.6%) and 25.7% on STL-10 (chance 10%; model 23.8%). It increases across rounds: 10.6→15.6% and 20.0→29.7%, respectively.

Accuracy by weight quartile (low→high):

DatasetQ1Q2Q3Q4
STL-1023.325.926.527.3
Caltech-10127.911.99.57.5

Higher weights correlate with correctness on STL-10 but not Caltech-101. Hence within-class weight magnitude is not a reliable confidence measure across datasets. The class-level S2 selection carries the useful signal.

Q4. Source of soft-loss gains. We confirm that all +soft rows in Appendix B use the same Wasserstein S2; only the query rule changes.

Self-training control: With identical annotation budgets, confidence-based self-training scores 11.48 vs. 24.04 on Caltech-101 and 25.50 vs. 28.65 on STL-10. Transport selection is particularly useful when model predictions are weak.

Fixed-S2 reweighting: Replacing Wasserstein weights with confidence or uniform weights yields comparable performance (Reviewer 1 response). Thus S2 selection, rather than specific weight values, explains most of the improvement.

Loss-scaling control: On Caltech-101, scaling the hard loss improves accuracy from 18.00 to 22.26, versus 24.04 with transport pseudo-labels (+1.78). On STL-10, gradient clipping limits scaling effects (27.16 vs. 26.97), while WASSALwithSoft achieves 28.65. The improvement cannot be attributed solely to loss magnitude.

We acknowledge that soft loss lowers average accuracy on four of five datasets. Its benefit is most pronounced when transport pseudo-labels outperform model predictions, especially in the many-class, few-label regime. We will explicitly restrict our claims.

Revisions. We will (i) specify the Sinkhorn objective, ε, solver, nonzero filtering, S2 rule and β; (ii) correct the imbalance description and Caltech-101 pool size; (iii) report mean±std; (iv) add learning curves, AULC, balanced accuracy, β sensitivity and S2 statistics; (v) include 1-NN, RBF, self-training, fixed-S2 and scaling controls; and (vi) remove the decision-boundary interpretation, emphasizing transport-guided soft-example selection instead.

Reviewer_yZMnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes WASSAL, a Wasserstein distance-based active learning framework that aims to improve sample selection and learning efficiency, focusing on settings with limited labeled data. The method uses class-specific weights for unlabeled samples learned by minimizing the Wasserstein-2 distance between a weighted distribution of unlabeled samples and the empirical distribution of labeled samples belonging to each class. The weights represent how much each unlabeled sample contributes to the class-specific probability distribution that best matches the distribution of labeled examples of that same class. The low weights of a sample across all the classes represent the most confusing example and are sent to oracle/expert for annotation, while high weights mean the unlabeled sample looks like a typical example of that class in the labeled training set, and the method aims to learn from the signals of those examples with soft labels, i.e., using the class label(s) that it is most representative of, without expert annotation. The framework uses a weighted loss term, using the weights learned from the aforementioned optimization, for training with these representative samples to account for their lack of expert-verified labels. The method improves on prior class-agnostic Wasserstein-based work by making the distribution matching class-conditional and suggests the significance of using both confusing and typical examples.

Critical strengths
  1. The proposed method improves over its baseline, which it used as an inspiration.
  2. The paper raises an interesting question about the significance of using the unlabeled examples that are representative of the labeled set.
Critical weaknesses
  1. The method doesn’t improve over compared methods consistently enough to be significant. The paper attempts to justify this with a different data regime for one of the datasets. Since different methods outperform at different budgets across different datasets, this resembles the restatement of the “no free lunch” challenge that has lately been posed by the active learning literature rather than a novel contribution.

  2. The paper lacks a graph showing the performance across AL rounds; the reported metric is only from the final round. The paper mentions that the dual approach of including both confusing samples and samples representative of labeled data has the potential to reduce the number of AL rounds or expert queries. Such a demonstration could have provided some evidence for this assertion.

  3. No ablation study is presented for the soft-loss hyperparameter beta. This makes the choice of the given beta value arbitrary.

  4. Even though the paper starts the experimental setup by specifying that the selected dataset is class-imbalanced. The paper reports a construction of a balanced unlabeled pool apart from the PneumoniaMNIST (Table 1). This undermines the very setup the paper tries to present in the beginning.

  5. The number of samples (Values s1 - section 3.2.1 and s2 -section 3.3) used for oracle and for soft-labeling, along with their ablations, are not reported. The topk selection depends on the score estimator learning well-behaved scores; ill-learned weights/scores can select a few of the same samples in both directions (confusing and representative), or, in other words, if the learned weights are not diffuse enough, then there is no significant distinction between confusing and representative samples in the unlabeled pool.

  6. Only accuracy is reported. Though the initial seed dataset is balanced, the unlabeled pool for PneumoniaMNIST is imbalanced; the accuracy metric can prioritize performance in the majority class.

  7. No standard deviation value is reported. Since the initial pool is small, the sampled distribution may look very different across seeds, which sets the model in a varied learning path and may have significant variation in metrics across seeds and across AL training rounds.

Questions for authors

Please see the weakness section.

1 comment

We thank the reviewer. Our additional experiments clarify the method's scope, learning dynamics, and limitations.

W1. Consistency of gains. We narrow the claim to many classes with few labels per class. On Caltech-101 (102 classes; 8 labels/class), WASSALwithSoft leads at all five budgets: roughly +6 points over WASSAL without soft loss and +12.6 over confidence self-training (Reviewer 1 response). The gain comes primarily from transport-selected soft examples (S2); query selection (S1) alone performs similarly to Random. Results on CIFAR-10, STL-10, SVHN and PneumoniaMNIST are mixed. We will remove any claim of universal superiority.

W2. Accuracy across AL rounds. Caltech-101 mean accuracy, aggregated over 5 budgets × 3 seeds:

Methodr0r1r2r3r4r5r6r7
WASSALwithSoft1.717.121.122.323.023.123.524.3
WASSAL1.75.912.115.016.818.018.919.3
DCoM1.75.612.716.618.220.020.922.0
MaxHerding1.75.813.116.417.919.520.321.2
BADGE1.75.211.615.117.118.719.620.6
Random1.75.411.915.617.518.519.820.9

WASSALwithSoft crosses 20% accuracy after 2 rounds, versus 5 for DCoM and 7 for BADGE/Random. Area under the learning curve (AULC) is 19.52, compared with 14.72 for the best competing baseline (DCoM). This supports fewer AL rounds on Caltech-101, not universally: on STL-10, WASSAL has a higher AULC (19.07 vs 18.57). We will add round-wise plots for every dataset.

W3. β sensitivity. Caltech-101, budget 100, 3 seeds:

β00.10.3 (paper)13
Accuracy18.00±1.4124.30±1.8824.04±1.9423.72±0.8523.93±1.71

Every tested positive β gives approximately +6 points over β=0. Variation over β∈[0.1,3] is smaller than typical seed variability, so performance is relatively insensitive within this range; β=0.3 was not tuned to maximize these results. We will add the ablation.

W4. Class imbalance. We agree the wording was incorrect. The pools are class-balanced except PneumoniaMNIST (1000:3000). We will correct the text.

W5. Sizes and overlap of S1/S2; diffuse weights. At budget 100 (subset sizes averaged over rounds and seeds):

QuantityCaltech-101STL-10
S1 per round100100
S2 per round2,2631,702
S2 / labeled set2.13.7
Median nonzero weights/class1,216/~3,3803,139/~4,300
S1∩S200

Transport weights remain diffuse rather than collapsing to a single-neighbor assignment. S1 selects annotation queries; S2 selects soft-training examples. The sets are disjoint by construction, and no point belongs to S2 for multiple classes. We will state the S2 rule and these statistics explicitly.

W6. Balanced accuracy, PneumoniaMNIST. Mean per-class recall at the final round, averaged across 9 budgets × 3 seeds:

MethodBalanced acc.Overall acc.
UHerding80.6785.05
GradMatch-TSS80.5884.85
WASSAL80.3884.40
BADGE80.2484.68
Random78.5982.88
Wasserstein-IP77.4482.16
WASSALwithSoft76.9682.23

WASSAL ranks third among 13 methods, 0.29 points below the best; however, adding the soft loss reduces balanced accuracy on this binary, imbalanced task. This reinforces the need to restrict the claim to the many-class, few-label regime. We will include balanced accuracy in the revision.

W7. Standard deviations. Mean±std for Tables 2, 3 and 6 appear in our Reviewer 1 response (Part 2). All use 3 seeds except classical STL-10 baselines at budgets 100–200 (2 seeds). We will incorporate stds in the revised tables.

Planned changes: qualify the scope; add round-wise curves, β ablation, S1/S2 sizes and overlap, balanced accuracy, variability estimates; and correct the imbalance description.

Reviewer_7eY4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WASSAL learns per-class weights over the unlabeled pool so the reweighted pool matches that class's labeled examples in 2-Wasserstein distance (eq 4), solved with projected gradient descent through Sinkhorn. Points with low weight across all classes are queried; high-weight points are treated as class landmarks and enter the next round's loss as weight-scaled soft labels, beta = 0.3 (eq 7). It is framed as a class-conditional version of the pool-level Wasserstein core-set of Mahmood et al. (2022). Evaluation is ResNet-18 on CIFAR-10, SVHN, PneumoniaMNIST, STL-10 and Caltech-101 with pools near 4k, twelve baselines, mean final accuracy over three seeds: WASSALwithSoft leads every Caltech-101 budget, query-only WASSAL most STL-10 budgets, mixed on the rest.

Critical strengths

Caltech-101 (Tab 2): WASSALwithSoft tops all five budgets, 1.35 to 2.23 points clear of the best of twelve baselines, on 102 classes with 8 labels per class. Bigger than the seed wobble elsewhere in the paper, on the hardest split in the study.

Current baselines under a shared protocol (sec 4.1): Wasserstein-IP, MaxHerding, UHerding and DCoM, same features, budgets and training for everyone. Tab 8 then reports the soft loss hurting other rules, down to -5.49.

Critical weaknesses

Eq 4 is vacuous as written. w_j ranges over the whole simplex, so the row-marginal constraint never binds: any coupling with the right column marginals already has row sums lying in the simplex. The problem separates by column, and the minimiser is the 1-NN pushforward of the labeled set, supported on at most m points per class. On Caltech-101 that leaves 3774 of 4590 pool points at exactly zero weight, so eq 5 is a tie-break, not a selection rule. Tab 4 agrees: query-only WASSAL gets 19.32, under Random's 20.90.

So only the soft loss works, and it is just pseudo-labeling. Tab 4 has WASSALwithSoft below query-only WASSAL on four of five datasets. On Caltech, Tab 8 shows the same soft loss on CoreSet or LeastConf buys +4.91 and +4.44, most of the gap. No self-training baseline anywhere.

Questions for authors
  1. Clarification. Does your implementation actually optimise eq 4 as written? If it does, the minimiser is closed-form and supported on at most m points per class, and eq 5 is a tie-break. What escapes that: minibatch Sinkhorn, a non-converged iteration count, a constraint not stated in the paper? Please give eps, the iteration count, the batching, and the per-class count of non-zero weights.

  2. Clarification, no new runs. Per-seed accuracies, or mean and std, for Tab 2, Tab 3 and Tab 6; your reproducibility statement says the logs exist. Also, the two WASSAL rows in Tab 6 at budgets 175 and 200 look swapped (53.94 then 38.07, against 37.94 then 53.29).

  3. An experiment - Hold S2 fixed and rerun eq 7 with w_ji replaced by softmax confidence, and by uniform weights, on Caltech-101 and STL-10 over the same three seeds. Tab 8 varies the query rule but never the weight source.

1 comment

We thank the reviewer. The additional controls clarify what drives WASSAL's gains and where our claims should be narrowed.

Q1. Eq. 4 versus implementation. The reviewer is correct: the exact LP reduces to a 1-NN pushforward. Our code uses a regularized, early-stopped solver, which yields distributed weights. Median nonzero weights per class are 1,216/3,380 (Caltech-101; minimum 613) and 3,139/4,300 (STL-10; minimum 2,927). On Caltech-101 the median changes from 2,338 at round 1 to 773 at round 7. Thus Eq. 5 ranks nontrivial weights, not ties, and all queries follow that rule. We will revise Eq. 4 to match the objective and solver. An exact-LP/1-NN baseline scores 1.5 points below WASSAL (see Reviewer 3, Q1).

W2/Q3. Does the soft loss merely implement pseudo-labelling? We held S2 and β fixed, replacing only the weights in Eq. 7. We also tested confidence-based self-training with random queries and the same S2 size/rule by class.

Budget 100 (3 seeds)Caltech-101STL-10
WASSALwithSoft (transport weights)24.04±1.9428.65±0.35
Same S2, confidence weights23.90±1.7828.43±0.25
Same S2, uniform weights24.52±1.4128.35±0.17
Confidence self-training11.48±0.5825.50±0.50
WASSAL, no soft loss18.00±1.4126.97±0.99
Random, no soft loss19.38±2.2325.71±0.56

S2 selection, rather than its particular weights, drives the benefit. Compared with confidence self-training, the full method improves by 12.56 points on Caltech-101 and 3.15 on STL-10. Early on Caltech-101 (round 1), model pool accuracy is 1.3%; model pseudo-labels are 2.2% correct, versus 14.2% for transport pseudo-labels. With S2 fixed, uniform/confidence weights perform similarly to transport weights. We will therefore describe the contribution as transport-guided selection of unlabeled examples, not a special advantage of the loss weights.

Loss-magnitude control. On Caltech-101, soft loss contributes about 24× the hard-loss mass. Scaling hard loss ×24 without soft loss yields 22.26±3.00, versus 24.04±1.94 with transport pseudo-labels (+1.78). On STL-10, with gradient clipping, scaling hard loss ×2.4 gives 27.16 versus 26.97 unscaled; WASSALwithSoft yields 28.65. Increased gradient magnitude alone does not account for the STL-10 gain.

Q2. Variance in Tables 2, 3 and 6. We report mean±std (3 seeds; classical STL-10 baselines at budgets 100–200 have 2 seeds). Key comparisons across budgets are:

Dataset / method2550100175200
Caltech WASSALwithSoft20.69±2.1522.66±1.0924.85±0.8825.94±0.5027.40±1.04
Caltech best other baseline19.2321.3122.7024.0125.17
STL-10 WASSALwithSoft23.8023.7523.9725.2425.88
STL-10 WASSAL21.3423.8825.2026.6627.66
SVHN WASSALwithSoft24.81±1.0933.29±4.1235.28±2.8653.94±6.9938.07±3.12
SVHN WASSAL28.85±2.2335.76±4.1439.18±11.1537.94±8.0353.29±4.01

On Caltech-101, the gap to the strongest competing baseline exceeds the sum of reported stds at budgets 100, 175 and 200; this is descriptive, not a formal significance test. On STL-10 and SVHN, gains are not consistent, and we will qualify our claims accordingly.

Table 6, SVHN at 175/200. We rechecked the logs: the rows are not swapped. Per-seed accuracies: WASSALwithSoft 175 = 56.95/58.91/45.95; 200 = 41.64/36.74/35.85. WASSAL 175 = 28.85/44.08/40.90; 200 = 55.62/48.67/55.59. SVHN exhibits substantial seed variability (std up to 11 points) and nonmonotonic learning curves. We will report this explicitly.

Revision. We will correct Eq. 4, add the 1-NN LP baseline, fixed-S2 and loss-scaling controls, seed variability, and a narrower statement of the contribution. The budget-100 ablation and main-table runs yield different absolute values; their experimental settings will be distinguished explicitly in the revision.

Comments on the paper

1 comment

We thank the reviewers for careful reviews. We ran 57 new runs on Caltech-101 and STL-10 to answer the questions. The new results support one clear claim.

Main result. Class-conditional transport selects a soft subset of the unlabeled pool. Training on this soft subset gives large gains in the low-label, many-class regime:

  • +6.0 points on Caltech-101 over the same pipeline without the soft loss (24.04 vs. 18.00), and +1.7 points on STL-10.
  • +12.6 points over model-confidence self-training on Caltech-101, and +3.2 points on STL-10.
  • +1.5 and +2.7 points over 1-NN and RBF-kernel weights on Caltech-101.
  • Stable for any β in [0.1, 3].
  • On Caltech-101, it reaches 20% accuracy in 2 rounds. The best baseline needs 5 rounds.

Refined claim. The new results show which part of WASSAL carries the gain. The gain comes from transport-based selection of the soft subset. The query rule alone performs like Random. The loss is robust to how the selected points are weighted. We will refine the paper's claim to this component and remove broader claims.

Implementation details (added to Section 3).

  • Objective: debiased Sinkhorn divergence (GeomLoss, p = 2, ε = 0.0025, ε-scaling 0.7). The full pool goes in one batch.
  • Solver: projected Adam, lr 1e-3, 10 steps from uniform weights.
  • Query rule (Eq. 5): ranks points with a non-zero max weight.
  • Soft subset S2, per class: top points up to 80% of the class mass; cap 30 (Caltech-101), 100 (SVHN), 400 (CIFAR-10, STL-10), 500 (PneumoniaMNIST). Each point goes to its argmax class only. S1 ∩ S2 = ∅. β = 0.3.

Corrections. Table 1: the Caltech-101 pool has ~3,680 points (some classes have fewer than 58 images). Section 4.1: the pools are class-balanced, except PneumoniaMNIST (1000:3000).

About the new runs. They use a clean re-implementation of the pipeline. On Caltech-101 it reproduces Table 2 (WASSALwithSoft 24.0 vs. 24.9; WASSAL 18.0 vs. 18.9; Random 19.4 vs. 20.3). On STL-10 its absolute accuracy differs from Table 3. Thus we compare only rows inside each new table. We will rerun Table 3 with the released code.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.