acceptodds
Under review as a conference paper at ICLR 2027

Regularized Transport Weights: a Principled Approach for Active Domain Adaptation

Abstract

Distribution shift, where there is a mismatch between training and test distributions, is a fundamental challenge in modern machine learning. In this context, we consider the setting of active domain adaptation, where the learner has access to a labeled source dataset , and aims to classify an unlabeled target dataset . The learner is allowed to query the labels of a few points in to mitigate the distribution shift between and , as in active learning. We propose a principled, task-agnostic method that jointly selects target queries and reweighs the source and queried examples. Our method tries to effectively use the existing source labels in order to balance coverage of the target distribution while mitigating the variance in the labels. We achieve this by minimizing a natural Regularized Transport Weighting (RTW) objective between the augmented dataset and the target . We provide theoretical guarantees for our method, including a generalization bound and an approximation guarantee for our algorithm. Experiments on standard benchmarks show that our method is competitive with the state-of-the-art methods, and outperforms prior methods on some standard benchmarks– with only a few samples, we are able to train a classifier that significantly improves upon the accuracy achieved on the source dataset alone.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
learning theory, active learning, domain adaptation, distribution shift

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_67odOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a task-agnostic method for active domain adaptation. The proposed RTW (Regularized Transport Weighting) jointly selects target samples to label and learns training weights for source and queried samples. Its objective combines transport cost with a penalty on weight concentration. The authors give theoretical evidence for the proposed method with generalization bounds and a greedy approximation guarantee. Also, they evaluate the method on OfficeHome-RSUT and mini-DomainNet.

Critical strengths
  1. The authors propose a single objective that connects query selection and training weights. They give theoretical evidence by a generalization analysis and a greedy approximation.
  2. The proposed method shows strong results at small query budgets on mini-DomainNet.
Critical weaknesses
  1. The impact of query selection of RTW itself is unclear. Comparisons that control for training weight are helpful to assess how much the proposed query selection contributes. For example, a comparison with random or K-medoids selection using the same RTW weighting helps to clarify this.
  2. The effect of the weight regularizer is also not directly shown. A comparison with RTW without the regularizer (λ = 0) and an analysis of sensitivity to λ will help clarify its contribution. The K-medoids baselines use different training weights, so they do not provide this comparison.
Questions for authors

How does RTW’s query selection time compare with K-medoids?

3 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for the opportunity to address any questions they may have. Below we answer some of the specific concerns raised.

  1. Thanks for the question about the effect of query selection vs reweighting. In this work, we focus on implementing the RTW method with the same fixed choice of λ\lambda for both the query selection, and the regularized reweighting. We chose to focus on this setting because this is the method for which we have theoretical guarantees. As a heuristic, it is very reasonable to consider two different choices of λ\lambda used for query selection and the final reweighting. For example, kk-medoids followed by our reweighting would correspond to λ=0\lambda = 0 at the query selection stage, and increasing λ\lambda at the end for reweighting. We leave the evaluation of these hybrid methods to future work. The kk-medoids method does not admit a theoretical guarantee by itself when the label is not deterministic. As a bad example, consider the case when the target distribution has some target mass pp at a point or tight cluster, that is not covered by the source. The kk-medoids method will query one example from this cluster, and then consider the entire cluster labeled. However, if there is some variance in the labels in the cluster, the single query is not enough to estimate the average label at that point with high accuracy. Our RTW method accounts for this uncertainty during selection, and trades-off being close to the target feature distribution (in a geometric notion of distance), with increasing the effective sample size of the reweighted distribution.

  2. Thanks for the suggestion to compare with RTW with λ=0\lambda=0. Since the kk-medoids baseline in our plots uses 50:50 balanced weights, we additionally ran RTW with λ=0\lambda = 0 (or equivalently, kk-medoids selection with nearest-neighbor weights) on both benchmarks, and we will include these curves in the revision. First, on mini-DomainNet, nearest neighbor weights completely removes the accuracy drop of kk-medoid when using 50:50 weights, and achieves a much higher accuracy overall. This confirms that transport-based weighting itself substantially improves over heuristic domain balancing. Second, comparing λ>0\lambda > 0 with λ=0\lambda = 0 matches the bias–variance trade-off intuition in Theorem 3.1: on OfficeHome-RSUT (the smaller dataset), λ=0\lambda = 0 would concentrate weight on fewer source samples, and the λ=10\lambda = 10 regularization consistently improves over λ=0\lambda = 0 across most tasks and budgets (by up to ~3 percentage points). On mini-DomainNet where more samples are present, λ=0\lambda = 0 and λ=1\lambda = 1 perform very similarly overall.

  3. Thanks for the question about the runtime of our method. Our method RTW does require more computation per candidate than kk-medoids. This is because it requires a regularized transport subproblem, as opposed to kk-medoids that only needs to compute nearest neighbor distances. In our experiments we implement a series of optimizations that maintain our theoretical guarantees. Subsampling the candidates, dual screening, warm starts, and batched GPU optimizations help us bring down the computational cost in our implementation, while preserving the theoretical guarantees. It is possible that there are further optimizations that can bring down the computational cost even further, or efficient and effective heuristics that combine aspects of RTW with other methods such as kk-medoids. This would be an interesting direction for future work. Moreover, we focus on settings where acquiring labels is costly, such as astronomy surveys, and therefore prioritize selecting queries that yield the most information.

Thank you for your rebuttal. Changing λ affects both which samples are selected and how they are weighted. Could you compare RTW and K-medoids for sample selection, using the same positive λ to compute the final RTW weights in both cases?

Following up on the runtime question, how long does each method take to select samples on the same hardware and with the same query budget?

AuthorsOpenReview ↗

Response to Reviewer 67od

We thank the reviewer for engaging meaningfully with our work. Below, we have attached the requested results, comparing k-medoids selection with RTW reweighting to RTW selection with RTW reweighting.

We would also like to mention that the reason we did not include this comparison in our original submission, is that we cannot establish theoretical guarantees for the algorithm that chooses the queries via kk-medoids, and then reweights the chosen queries with the RTW weights. This is because the kk-medoids selection strategy does not account for the potential variance in the labels yy. The main contribution of our work is to establish a theoretical framework, that allows us to derive the RTW weighting in the first place. Once we have this, it is possible that it could be incorporated into a variety of efficient practical heuristics. Evaluating these various methods in practical settings is an interesting direction that we do not pursue.

For the running time comparison, at each greedy round for OfficeHome-RSUT, RTW takes ~10-15 seconds to identify the point to query, whereas k-medoids takes less than a second.

In addition to this, we also present some synthetic experiments, which concretely demonstrate situations where k-medoids can fail (even if it uses RTW reweighting).

OfficeHome-RSUT

Held-out target accuracy (%) at 100 queries, with λ = 10.

TaskK-medoids + RTWRTW + RTW
Clipart → Product71.79 ± 2.0272.85 ± 1.35
Clipart → Real World72.35 ± 0.9271.39 ± 1.55
Product → Clipart44.51 ± 1.8143.53 ± 2.41
Product → Real World77.69 ± 0.9376.73 ± 2.61
Real World → Clipart44.90 ± 2.1545.29 ± 2.47
Real World → Product78.04 ± 2.5977.78 ± 1.55

mini-DomainNet

Held-out target accuracy (%) at 100 queries, with λ = 1.

TaskK-medoids + RTWRTW + RTW
Clipart → Real68.55 ± 0.2867.70 ± 0.75
Painting → Real73.32 ± 0.2373.04 ± 0.30
Sketch → Real64.31 ± 0.3463.40 ± 0.62

Synthetic Experiments

As mentioned before, our main focus of our work is principled approach involving reweighting, and an algorithm with rigorous guarantees. We now present simple synthetic experiments that demonstrate the failure point of the k-mediod based query strategy (even if we later reweight using our RTW objective with λ=1\lambda=1). We also compare it to our greedy RTW method with λ=1\lambda=1 (which comes with a guarantee).

For each problem, we ran 30 paired trials in 40 dimensions. Both methods used the same 200 labeled source examples and selected K=30K=30 queries from the same unlabeled target pool using features alone. We used λ=1\lambda=1 for RTW selection and computed the final training weights for both query sets using RTW with λ=1\lambda=1. We used the same model and training settings within each comparison.

Noisy midpoint regression. The source distribution is a mixture of three clusters: component AA with mixing weight 0.1 is a standard Gaussian with center on the x-axis at -12 while the other two clusters BB and CC are point masses with centers at 00 and +12+12 on the x-axis. The target is a mixture of 10%10\% from AA, 50%50\% from CC, and 40%40\% from BB. The Gaussian component has response zero, while the observations for cluster CC and BB have responses with means −1-1 and +1+1 respectively with variance 1/41/4. We used a pool of 1,000 target examples, fit weighted affine ridge regression with penalty α=10\alpha=10, and evaluated exact target population mean squared error (MSE).

The experiment results are given below. RTW selection reduced mean MSE from 0.943 to 0.776 under the common weighting rule, a 17.8% reduction. RTW selected six and five observations at the two noisy locations, respectively. K-medoids selected one at each, which already gives zero distance at that location. We saw a similar result with λ=10\lambda=10 for RTW selection and both final weight computations (0.753 versus 0.939 MSE). While we described a specific setting for concreteness, these failure models for k-mediod query selection are natural in settings with very hetergeneous populations with clusters of different radii e.g., a mixture of Gaussians with different covariances (it becomes worse in high dimensions), as k-mediod picks at most 11 (or very few points) from the tighter clusters even if there is variation in the labels. We will describe this synthetic experiment as an illustrative example in the final version.

Regression: target population MSE at 30 queries.

λK-medoids + RTWRTW + RTWRelative MSE reduction
10.943 ± 0.0440.776 ± 0.01217.8%
100.939 ± 0.0440.753 ± 0.00819.8%
Reviewer_pfAgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Regularized Transport Weighting (RTW), a task-agnostic method for active domain adaptation. Given labeled source data and an unlabeled target dataset, RTW jointly selects a small batch of target samples for labeling and reweights the source and queried samples. Its objective balances the transport cost between the weighted labeled data and the target distribution, while regularization is used to prevent excessive concentration of sample weights. Experiments across multiple datasets show that RTW is competitive with existing active domain adaptation methods and improves over source-only training in several settings.

Critical strengths
  1. The paper connects a transport-based objective to a generalization bound and establishes a supermodularity property that supports an approximation guarantee for greedy selection. It makes a theoretical contribution to a certain extent
  2. Promising results across multiple domain adaptation settings. The experiments cover several transfer tasks and compare RTW with different query-selection baselines. The results suggest that the proposed joint objective can be effective in practice.
Critical weaknesses
  1. The relationship between the approximation guarantee and the practical implementation is unclear. The theoretical guarantee is established for a greedy selection procedure, whereas the practical algorithm introduces computational shortcuts, including candidate subsampling and dual-based screening. The relationship between these steps and the theoretical guarantee requires further clarification.
  2. The choice of the regularization parameter requires further justification. The experiments use different values of (\lambda) for the two benchmark datasets. Since (\lambda) controls the trade-off between transport cost and weight concentration, its value may substantially affect both the selected queries and the resulting sample weights. The paper should clarify how this parameter is selected, what information is available during selection, and how sensitive the results are to this choice.
  3. Performance improvements are inconsistent across tasks. RTW performs strongly on several transfer tasks, but it does not outperform every baseline in every setting. Some gains are modest, whereas others are more substantial. Further analysis is needed to characterize and explain these differences.
  4. The evaluation relies on fixed pretrained representations and a linear classifier. This is a reasonable choice for isolating the effects of query selection and reweighting. However, it remains unclear how the method behaves when the learned representations are less suitable for the target domain or when the feature extractor is adapted during training. Since the transport objective directly depends on the geometry of the feature space, the choice and quality of the representation may substantially influence the method's behavior.
Questions for authors
  1. Preservation of the approximation guarantee. Do candidate subsampling and dual-based screening preserve the approximation guarantee established for the greedy algorithm? If the guarantee applies only to the idealized algorithm, could the authors compare the practical implementation with exact greedy selection on smaller instances to quantify the effects of these approximations?
  2. Selection and sensitivity of (\lambda). How is (\lambda) selected in the experiments, and what information is available when making this choice? Could the authors provide a sensitivity analysis to show how different values of (\lambda) affect query selection, sample weights, and final performance?
  3. Variation in performance across experimental settings. What factors explain the differences in performance across transfer tasks, datasets, and query budgets? In particular, can the authors identify the conditions under which RTW is more or less effective than the baselines?
  4. Consistency across transfer tasks and feature representations. Could the authors provide an overall comparison across transfer tasks and query budgets, together with a discussion of settings in which RTW does not outperform the baselines? Furthermore, how sensitive is RTW to the feature representation used to define the transport cost? Do the conclusions remain consistent when different pretrained feature extractors or domain-adapted representations are used?
2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for the opportunity to address any questions they may have. As mentioned in the general response, we view our primary contribution as theoretical. Our conceptual contributions are identifying a principled, statistically motivated objective (RTW) and providing a simple method with rigorous guarantees by establishing the objective’s supermodularity. We hope this main takeaway comes across on a direct reading of the introduction, and we will make sure to highlight it even more prominently in the revision.

Below we answer some of the specific concerns raised.

  1. Guarantees in practice We would like to point out that the practical implementation of our algorithm inherits the theoretical guarantees of the algorithm described in our paper. The distinction was made to make the exposition clearer. In particular, the dual pruning is a technique that makes the algorithm faster without affecting the guarantee at all while subsampling is known to maintain the 1−1e−ϵ1-\frac{1}{e}-\epsilon guarantee (with a small ϵ\epsilon loss), for details see this AAAI 2015 paper. We will clarify this in the final version.

  2. Choice of λ\lambda. From limited experiments, the performance of our algorithm doesn't seem very sensitive to the selection of λ\lambda as long as it is within values between 11 and 5050. We did not tune or optimize λ\lambda since our paper is primaily theoretical, we chose to evaluate and report λ=1\lambda=1 and λ=10\lambda=10 which were chosen arbitrarily.

  3. Variation in Performance and consistency. While the performance of RTW varies across tasks, crucially, across the evaluated settings, it is never substantially worse than other baselines. RTW performs particularly well in the low query budget setting. One possible explanation for this is that reweighting and careful query selection are more useful when we have very few labeled target samples. Similarly, RTW may be particularly effective when a subset of the source samples is representative of the target distribution. For example, in mini-DomainNet tasks, reweighting the source alone improves performance substantially.

AuthorsOpenReview ↗

Any remaining concerns?

We wanted to briefly follow up on our response above where we provided detailed answers to all the questions raised in your review. Could you let us know if these clarifications addressed your concerns, or if any concerns remain? We would welcome the chance to clarify anything further while the discussion window is still active. Thanks!

Reviewer_M4zrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Regularized Transport Weighting (RTW), a task-agnostic approach for active domain adaptation that jointly optimizes target sample selection and training sample reweighting. The paper addresses the challenge of learning under distribution shift when labeled source data are available but only a limited number of target-domain samples can be labeled. The proposed method formulates a regularized optimal transport objective that balances two factors: minimizing the distributional discrepancy between the weighted training data and the target domain, and controlling statistical variance by preventing excessive concentration of sample weights. The authors establish generalization bounds for the resulting weighted empirical risk minimization and prove the supermodularity of the proposed objective, enabling a greedy query selection algorithm with a theoretical approximation guarantee.

The method is evaluated on OfficeHome-RSUT and mini-DomainNet, using frozen ResNet-50 features and weighted logistic regression classifiers. Experimental results demonstrate competitive performance against existing importance-weighting and active domain adaptation methods, with improvements on several adaptation tasks, particularly under limited target-label budgets. The paper concludes that jointly optimizing sample selection and regularized reweighting provides a theoretically grounded and practically effective approach to active domain adaptation.

Critical strengths
  1. Meaningful theoretical contribution: The paper introduces a principled regularized optimal transport objective that jointly addresses target sample selection and training sample reweighting. In particular, the supermodularity result for the fixed-regularization objective provides a nontrivial theoretical contribution and supports a greedy algorithm with an approximation guarantee under exact selection. This strengthens the paper's methodological significance, despite the remaining gap between the theoretical assumptions and practical implementation.
  2. Promising performance in low-label-budget settings: The proposed method achieves substantial improvements over competitive active domain adaptation and importance-weighting baselines on several benchmark tasks, particularly on mini-DomainNet with limited target labels. These results demonstrate the potential practical value of the proposed framework, although its consistency and generalizability require further validation.
Critical weaknesses
  1. Gap between theoretical guarantees and practical implementation. The theoretical guarantees do not fully align with the implemented method. Theorem 4.1 assumes exact greedy selection, whereas the experiments use randomly sampled candidate subsets. Furthermore, the regularization parameter required by Theorem 3.2 depends on the optimal weights, while the practical algorithm uses fixed values. These discrepancies weaken the connection between the theoretical contributions and the empirical results.
  2. Limited experimental generalization and inconsistent performance improvements. Although the method is claimed to be task-agnostic, experiments are restricted to two image classification benchmarks using frozen ResNet-50 features and linear classifiers, without validation on different downstream tasks or model architectures. Moreover, RTW does not consistently outperform existing methods across all adaptation settings, and some improvements are small relative to the reported standard deviations. This raises concerns about the robustness, generalizability, and statistical reliability of the empirical advantages.
Questions for authors
  1. Theoretical Guarantee versus Practical Implementation: Theorem 4.1 establishes an approximation guarantee for exact greedy selection, while the experimental implementation evaluates randomly sampled candidate subsets. Could the authors clarify whether the theoretical guarantee still applies to the implemented algorithm and how candidate sampling affects the approximation quality?
  2. Regularization Parameter and Theoretical Assumptions: Theorem 3.2 specifies a regularization parameter dependent on the optimal sample weights, whereas the experiments use fixed values of (\lambda=1) and (\lambda=10). How were these values selected, and to what extent do the reported generalization guarantees apply to these practical settings?
  3. Applicability of the Generalization Bounds: The theoretical analysis requires a transport cost that bounds conditional-risk differences and assumes bounded losses, while the experiments use distances between pretrained ResNet-50 features and logistic regression classifiers. Could the authors clarify how these theoretical assumptions are justified in the experimental setting?
  4. Performance Consistency and Statistical Reliability: RTW does not consistently outperform existing methods across all adaptation tasks, and some improvements are relatively small compared with the reported standard deviations. Could the authors clarify whether the improvements are statistically significant across the five evaluation splits and explain the performance differences between adaptation tasks?
  5. Task-Agnostic Generalization: Although RTW is presented as a task-agnostic framework, all experiments focus on image classification using frozen ResNet-50 features and linear classifiers. Does the task-agnostic claim refer primarily to label-independent sample selection, or is there existing evidence that the selected samples remain effective across different downstream tasks, model architectures, or feature representations?
3 comments
AuthorsOpenReview ↗

Response to the Reviewer

We thank the reviewer for the opportunity to address any questions they may have. As mentioned in the general response, we view our primary contribution as theoretical. Our conceptual contributions are identifying a principled, statistically motivated objective (RTW) and providing a simple method with rigorous guarantees by establishing the objective’s supermodularity. We hope this main takeaway comes across on a direct reading of the introduction, and we will make sure to highlight it even more prominently in the revision.

Below we answer some of the specific concerns raised.

  1. Scope and contribution We would like to emphasize that we believe that our paper should be viewed as a foundational, theory contribution and not an extensive experimental study. Our contribution is a principled formulation of joint query selection and sample reweighting with approximation and generalization guarantees. In our paper, task agnostic means that the selection of queries does not take into account the labels but only the features provided by the fixed representation. This allows for the reuse of the queried points. The goal of task-agnosticity is not for the method to be competitive with all labeling methods for every task, and indeed this might be too much to expect for a single algorithm. Instead, one main contribution is that such an algorithm has uniform guarantees accross tasks. This is valuable for applications where the same labeled queries might be used for multiple tasks, e.g., in astronomical surveys, as we mentioned in the introduction.

  2. Theory and implementation We would like to point out that the practical implementation of our algorithm inherits the theoretical guarantees of the algorithm described in our paper. The distinction was made to make the exposition clearer. In particular, the dual pruning is a technique that makes the algorithm faster without affecting the guarantee at all while subsampling is known to maintain the 1−1e−ϵ1-\frac{1}{e}-\epsilon guarantee (with a small ϵ\epsilon loss), for details see this AAAI 2015 paper. We will clarify this in the final version.

  3. Fixed regularization From limited experiments, the performance of our algorithm doesn't seem very sensitive to the selection of λ\lambda as long as it is within values between 11 and 5050. Theorem 4.1 holds for every fixed λ≥0\lambda \ge 0, and Theorem 3.1 applies to any feature-only queries and weights (including fixed λ\lambda), while Theorem 3.2 gives the sharper oracle rate tying excess risk directly to 2RTWλ2\text{RTW}_\lambda We intentionally did not tune or optimize λ\lambda,and chose to evaluate and report λ=1\lambda=1 and λ=10\lambda=10.

  4. Statistical assumptions. In our experiments, embeddings are ℓ2\ell_2-normalized and the logistic regression head is ℓ2\ell_2-regularized, which keeps weights and logits bounded. The cross-entropy loss is therefore both bounded and Lipschitz in feature space. Our transport-cost condition hence reduces to a natural smoothness assumption on the representation: on the data distribution, images close in feature space have similar conditional label distributions Y∣xY \mid x. Lipschitzness in representation space is also standard in prior active learning and domain adaptation analyses. (e.g., de Mathelin et al., 2022 and Mahmood et al., 2022)

  5. Performance comparisons. While the performance of RTW varies across tasks, crucially, across the evaluated settings, it is never substantially worse than other baselines. RTW performs particularly well in the low query budget setting. One possible explanation for this is that reweighting and careful query selection are more useful when we have very few labeled target samples. Similarly, RTW may be particularly effective when a subset of the source samples is representative of the target distribution. For example, in mini-DomainNet tasks, reweighting the source alone improves performance substantially.

Thank you for the response and additional experiments. I appreciate the theoretical contributions, but I still have two main concerns. First, the statistical significance of the reported improvements has not been directly addressed. A paired significance analysis across the five random splits would help clarify whether these gains are reliable. Second, the new ablation shows that K-medoids selection with RTW reweighting outperforms RTW selection on 7 out of 9 tasks, which raises questions about the actual benefit of the proposed query selection strategy. Overall, I feel these concerns remain unresolved.

AuthorsOpenReview ↗

Any remaining concerns?

Thanks for your reply. Following your review which had five specific questions (Q1–Q5), we took care to provide detailed answers for these five points. Could you share which of your questions remain unresolved after our response, or what specific concern continues to hold back your rating? Any feedback would give us a welcome opportunity to clarify before the discussion period ends.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for the feedback.

  1. The primary contribution of our paper is conceptual, in the derivation of the RTW objective. This is a principled, statistically motivated objective that informs how to select the target queries, and how to reweight the training samples. Moreover, since this objective is supermodular, it comes with a simple algorithm that has rigorous guarantees. This is possible due to our new theoretical framing of the question, that makes it clear how to balance the competing objectives of reducing the bias (distance to target distribution) and variance (related to the effective sample size) of the reweighted training distribution. We believe that a main contribution of this work is this theoretical framing that allows us to derive principled objectives and algorithms. We include experiments to validate the theory that this framework can improve over previous ad hoc methods for choosing the samples and reweighting. In particular our new RTW method interpolates between methods proposed in prior work, such as kk-medoids (when λ=0\lambda = 0 and S=∅S = \emptyset) (de Mathelin et al, 2022) and uniform-weight Wasserstein querying (when λ=∞\lambda = \infty) (Mahmood et al, 2022). Our takeaway is that there is much to gain in this problem by desiging theoretically principled methods. It is possible that the broad methods that we design can be optimized for certain settings, and engineerined to perform even better on benchmarks, and we leave this to future work. The focus of this work is on the foundations, and we think that the submission should be primarily evaluated from this point of view.

  2. In response to concerns about a theory-implementation gap, we would like to clarify that the accelerations that we perform in the implementation of our algorithm maintain our theoretical guarantees. In particular, the dual pruning is a technique that makes the algorithm faster without affecting the guarantee at all while subsampling is known to maintain the 1−1e−ϵ1-\frac{1}{e}-\epsilon guarantee (with a small ϵ\epsilon loss), for details see this AAAI 2015 paper. We will clarify this in the final version.

  3. One of the reviews raised the interesting question of which aspect of our method is more important: query selection, or reweighting of the samples (training data + queries)? This is an interesting question to explore in future work. We would like to point out that both the query selection method and the reweighting method are new conceptual contributions of this work. It is possible that to consider different heuristics by doing query selection with parameter λ1\lambda_1 and reweighting using a (potentially) different λ2\lambda_2; or more generally either of these ideas could be combined with other methods in the literature. But of course they may not admit a theoretical guarantee, and we do not explore this in this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.