acceptodds
Under review as a conference paper at ICLR 2027

Learning to Acquire via Look-Ahead Loss Estimation for Active Learning

Abstract

A common strategy in (deep) active learning is to acquire samples that the current model finds most challenging, using signals such as uncertainty or loss. This introduces a myopic bias in acquisition, as it relies on the assumption that samples identified as challenging at acquisition time remain challenging throughout subsequent optimization within the same acquisition round. Across image, text, and audio classification tasks, we find substantial within-round temporal misalignment: agreement between current and future per-sample loss rankings degrades rapidly as training epochs increase, so samples that appear difficult under the current model may become less challenging after subsequent training epochs. In this work, we present AHEAD (Active learning with Horizon-aware Estimation of Anticipated Difficulty), a framework that learns sample difficulty from estimated future loss rather than observed current loss. Concretely, AHEAD predicts each sample's relative loss epochs ahead, providing a future-oriented supervision signal that avoids relying on outdated loss observations and improves robustness to rapidly evolving training dynamics. The predicted future difficulty serves as a consistent target for training a representation-based importance function that reweights labeled samples during training. Because its score depends only on sample representations, it can be directly applied to unlabeled data as a label-free acquisition function. The result is a single mechanism for emphasizing samples that remain difficult during subsequent optimization rather than those that are momentarily challenging. Across six datasets spanning three modalities, AHEAD consistently outperforms strong uncertainty-, diversity-, and learning-based active learning baselines.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Active Learning, Learning to Acquire, Look-Ahead Loss Prediction, Sample Reweighting, Future Difficulty Prediction

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_9hVdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AHEAD, an active learning framework motivated by changes in sample difficulty during training. It predicts relative losses several epochs ahead and uses these predictions to train a representation-based importance function, which both reweights labeled training samples and ranks unlabeled candidates for annotation. Experiments on six datasets spanning images, text, and audio report improved training performance and competitive active-learning results, with ablations examining the effect of look-ahead supervision.

Critical strengths
  1. AHEAD offers an interesting approach to loss-based active learning by accounting for how sample difficulty evolves during training. Its use of predicted future losses to supervise a shared importance function connects training-time reweighting with unlabeled sample selection, providing a coherent framework for incorporating anticipated difficulty into both stages.

  2. The experimental evaluation is broad, covering six datasets across image, text, and audio classification with different backbone architectures. AHEAD improves over uniform training and achieves competitive active-learning performance in several settings. The accompanying horizon and warm-up ablations further explore when look-ahead supervision is beneficial, adding useful analysis beyond the main comparisons.

Critical weaknesses
  1. AHEAD’s contribution is insufficiently distinguished from prior dynamics-aware and look-ahead acquisition methods. TiDAL and TOD already exploit training dynamics for acquisition, while LADA (NeurIPS 2021) anticipates augmentation effects and NTK-based look-ahead methods approximate candidate-induced model updates. These approaches differ from AHEAD’s prediction of relative loss at a specified horizon, but their omission leaves the methodological positioning incomplete. The paper needs to explain what this particular prediction target adds and provide targeted comparisons with the closest applicable alternatives. Contrasting current and future loss alone does not establish the added value over existing approaches that already consider training dynamics or subsequent operations.

  2. The experimental evidence does not adequately substantiate the claimed acquisition advantage. Table 1 and the Round 0 differences in Figure 2 indicate training-side benefits, yet the evaluation does not isolate these from improvements in sample selection. Beyond this confounding, the central temporal-misalignment claim is not directly quantified, and the reported RankNet accuracy does not establish an advantage over current-loss rankings on unseen candidates. The acquisition-batch-size ablation also leaves unclear how the total annotation budget is controlled. Finally, the claim of highest final accuracy on CIFAR-10 and CIFAR-100 conflicts with Figure 2, where AutoAL with priors has higher final means. Together, these issues weaken confidence in both the explanation and the extent of AHEAD’s reported acquisition gains.

Questions for authors
  1. Relation to prior work. How does AHEAD’s fixed-horizon loss target differ from the training-dynamics signals in TiDAL and TOD, and from the distinct look-ahead formulations in LADA and NTK-based acquisition? What evidence supports the added value of this particular target over the closest applicable alternatives?

  2. Separating acquisition from training benefits. Were the initial labeled sets and model initializations shared across methods in Figure 2, and what explains the Round 0 performance differences? Does AHEAD retain an advantage when selected subsets are evaluated under a common training procedure, or when random acquisition is combined with AHEAD’s reweighting?

  3. Validation of the future-difficulty signal. Can the authors quantify the claimed disagreement between current and future loss rankings? For Figure 3’s ranking accuracy, were predictions recorded before the corresponding future losses became available, and were the evaluated samples excluded from predictor training? A comparison with current-loss rankings, particularly for the highest-ranked candidates, would help establish whether forecasting provides additional acquisition-relevant information.

  4. Interpretation of the batch-size ablation. Does B=16B=16–512512 in Figure 3 denote the acquisition batch size or the training mini-batch size? If it denotes acquisition, what were the total annotation budget and number of rounds for each setting? Clarifying what was held fixed is necessary to distinguish an effect of acquisition granularity from differences in supervision or training.

  5. Consistency of the final-performance claims. How should the claim in Section 4.2.2 that AHEAD achieves the highest final accuracy on CIFAR-10 and CIFAR-100 be reconciled with Figure 2, where AutoAL with priors appears to have higher final means? Could the authors provide the corresponding numerical results with uncertainty estimates and confirm that the endpoints use matching annotation budgets?

1 comment

We sincerely thank you for your detailed review and constructive comments. We appreciate the opportunity to clarify the methodological distinctions and experimental settings.


Q1. Relation to prior training-dynamics and look-ahead acquisition methods.

Unlike prior methods that primarily use training dynamics to guide acquisition, AHEAD uses predicted future relative losses for both sample acquisition and training-time reweighting. Specifically, AHEAD jointly trains a look-ahead loss predictor, an importance function, and the backbone neural network, using predicted future losses to supervise sample reweighting during backbone training. The look-ahead mechanism in LADA considers future acquired samples, whereas AHEAD predicts sample difficulty at a fixed future training horizon. TOD constructs a target for the predictor to learn, whereas AHEAD uses observed future relative losses as delayed supervision. The distinction lies in both the prediction target and its use in acquisition and reweighting.


Q2. Separating acquisition and training benefits, and explaining round-0 differences.

AutoAL and DISTIL use different codebases, which may account for the small differences observed at round 0. Within each codebase, the round-0 results are similar across methods, except on CIFAR-10, where the original training schedule was insufficient for convergence. We reran this experiment with additional training epochs and trials, and the updated results are provided in Figure 1 of the anonymous repository: https://anonymous.4open.science/r/AnonymousRepo-D90D/.

We emphasize that AHEAD combines sample acquisition with training-time reweighting through joint training of the look-ahead loss predictor, importance function, and backbone neural network. Consequently, the reported comparisons evaluate the overall procedure rather than isolating the effect of acquisition alone. We agree that the suggested controls would help separate these contributions. AutoAL also reported that its backbone-only baseline (referred to as the uniform backbone baseline) performs worse than several competing methods. To examine this setting, we ran a variant of AHEAD in which the evaluation model was trained as a backbone network using the acquired labels, without AHEAD's training-time reweighting. For example, the round 5 result for CIFAR-100 for this variant is 58.7018 ± 0.7493, which is substantially lower than the corresponding AHEAD result.


Q3. Validation of future-loss predictions and their relevance to acquisition.

Ranking accuracy measures the predictor's ability to order samples by their future losses. It serves as a diagnostic of prediction quality rather than a direct measure of acquisition utility. We also report Pearson correlations between current and future losses for newly acquired samples before subsequent training. These results are provided in Table 1 of the linked repository.


Q4. Interpretation of batch size in Figure 3.

B denotes the training mini-batch size, rather than the acquisition batch size. In this ablation, we train the model using the full dataset to assess the effect of mini-batch size on the reweighting procedure. We will make this distinction explicit in the figure caption.


Q5. Consistency of final-accuracy claims with Figure 2.

We agree that the statement in Section 4.2.2 regarding the highest final accuracy on CIFAR-10 and CIFAR-100 is inaccurate. AHEAD does not achieve the highest final accuracy on both datasets. Our intended claim was that AHEAD achieves the best performance on most of the evaluated datasets, rather than all of them. We will revise the statement to be consistent with the reported results in Figure 2.

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

The paper introduces a novel active learning selection strategy. It is based on the idea of predicting the usefulness of the labeling of samples in future epochs rather than being reliant on their current status. This approach seems novel and useful. The approach is suitably placed into a broader context within the submission. The experimental validation uses 6 datasets from 3 settings (though only one each for text and audio) and a reasonable number of competitors.

Critical strengths

The idea to predict future sample importance is very intriguing and appears to be novel. Using relative ranking rather than true importance seems like a sensible approach that helps the solvability of the task. Overall, it seems like a suitable strategy to active learning. The results corroborate this suitability. My assessment is that the paper adds a worthwile approach that could open interesting future directions.

Critical weaknesses

(W1) The biggest concern is the unexplained swap of metrics for one dataset (EER vs Accuracy). It is unclear why this was done and it may have been to hide poor performance based on accuracy.

(W2) There are some presentation issues: The descriptions of the figures can be somewhat insufficient. For Figure 2, only the directionality of one of the two metrics is given. For Figure 3, the actual setting for the middle plot is underspecified. The font sizes in the figures are generally too small. Several aspects (∆ and BCEWithLogits) are not defined, τ is used before its introduction. The metrics are insufficiently explained and should be properly cited. All of these should be easily fixable by the authors.

Questions for authors

Please explain how your answers will be reflected in the paper (just educating reviewers in this private discussion room is not sufficient):

(Q1) Why do you swap to Equal Error Rate instead of Accuracy for ASVSpoof2019? (cf. W1)

(Q2) Please confirm how you will modify the paper to meet the points mentioned above in (W2).

2 comments

We sincerely thank you for your review and constructive suggestions.


Q1. Use of EER instead of accuracy for ASVspoof2019.

We report Equal Error Rate (EER) for ASVspoof2019 because it is a standard evaluation metric in audio deepfake detection, particularly for datasets with imbalanced class distributions. We will make this rationale explicit in the revised manuscript.


Q2. Figure presentation, experimental details, and notation.

We will clarify whether higher or lower values indicate better performance for each evaluation metric, specify the experimental settings for the middle panel of Figure 3, and increase font sizes across all figures. We will also define (Δ\Delta) and BCEWithLogits, and introduce (τ\tau) before its first occurrence in the revised manuscript.

Thank you, I'll retain my score (3/weak accept).

Comments on the paper

1 comment

We sincerely thank you for your thorough review and valuable comments, and we appreciate the opportunity to explain certain aspects of our work.


Q1. Separating the contributions of acquisition and training-time reweighting.

AHEAD is designed to improve active learning performance by combining look-ahead acquisition with training-time reweighting (i.e., by jointly training the look-ahead loss predictor, importance function, and backbone neural network). We acknowledge that comparing AHEAD with backbone-only baselines (i.e., referred to as "uniform backbone" baseline) does not isolate the contribution of acquisition from that of reweighting. The additional controls suggested by the reviewer would help distinguish these effects. AutoAL also reported that its backbone-only baseline performs worse than several competing methods. Concretely, we ran a variant of the AHEAD where the evaluation model was trained as a backbone network with the acquired labels, and we observed similar results as reported in AutoAL: For instance, the round 5 result on CIFAR-100 for this variant is 58.7018 ± 0.7493, which significantly underperforms AHEAD.

Furthermore, we reran the existing experiments with additional training epochs and trials. The updated results, which slightly differ from those in the submitted manuscript due to randomness, are provided in Figure 1 of the anonymous repository: https://anonymous.4open.science/r/AnonymousRepo-D90D/.


Q2. Experimental settings and the T=0 baseline in Table 1.

For Question 2, Table 1 reports results obtained using the full training dataset. We checked the implementation and repeated the experiments, confirming that 53.47% is the correct result. T=0 refers to the version without look-ahead prediction but with AHEAD architecture retained. ResNet training can exhibit variability across runs, which may explain why the average accuracy in this setting is lower than that of uniform training. Our look-ahead approach aims to mitigate this issue.


Q3–4. Temporal rank misalignment and the relevance of predicted future losses to acquisition.

For Questions 3 and 4, we report Pearson correlations between current and future losses for newly acquired samples, measured before subsequent training. The results are provided in Table 1 of the linked repository.


Q5. Potential feedback loop between the predictor and learner.

For Question 5, the predictor's target changes as the learner is updated. In our implementation, the prediction targets are held fixed within each training block using an EMA model, and the predicted weights are detached when updating the learner. These choices limit direct feedback during each update, although they do not entirely eliminate interactions over longer training horizons. The T=0 ablation in our original submission examines the contribution of look-ahead prediction.


Q6. Computational overhead.

The wall-clock time comparison is provided in Figure 2 of the linked repository.


Q7–8. Definition of hyperparameters and the trajectory regularizer.

For Questions 7 and 8, T refers to the look-ahead horizon in epochs, while E denotes the number of warm-up epochs. The occurrence of K in Section 3.4 is a notation error; it should be E. We will correct this in the revised manuscript. The trajectory regularizer mentioned in Appendix C was not used in the reported experiments. We will remove this inaccurate description.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.