acceptodds
Withdrawn from ICLR 2027

AutoPrune: Accelerating End-to-End Driving Training via Gradient-Preserving Adaptive Data Pruning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unified end-to-end architectures have reshaped autonomous driving, but their reliance on massive datasets leads to weeks-long training, severely hindering model iteration. A natural way to accelerate training is to remove simple or redundant data that provide limited learning value. However, existing data reduction methods for end-to-end driving often rely on static data features and hand-crafted selection rules, facing two fundamental limitations: (1) Gradient distortion. Pruning low-value data can alter the gradient composition and degrade model performance; (2) Training-dynamics misalignment. Evolving data values make fixed pruning decisions stale, while fewer iterations can misalign the optimization schedule with full-data progress. Therefore, we present AutoPrune, which reduces training-data usage to accelerate end-to-end driving training while mitigating pruning-induced performance degradation. First, Gradient-Preserving Data Pruning is introduced to compensate for the missing training gradients caused by pruning low-loss samples, so that the corrected aggregate gradient matches its full-data counterpart in expectation. Second, Training-Dynamics Adaptation uses loss-adaptive sample gating to update pruning decisions from the latest available loss records and pruning-aware learning-rate alignment to follow the corresponding full-data training progress. Experiments across multiple end-to-end driving models show that AutoPrune retains performance close to the corresponding full-data references at a 40% overall pruning ratio. Comparative and ablation studies further support the effectiveness of the proposed components.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Adaptive Data Pruning, Training Acceleration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DPp9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AutoPrune, an adaptive data pruning framework designed to accelerate end-to-end autonomous driving training without sacrificing planning performance. It contributes an expectation-preserving gradient compensation mechanism for pruned low-loss samples coupled with loss-adaptive sample gating and a pruning-aware learning rate alignment schedule, validated across multiple architectures (SparseDrive, VAD, DiffusionDrive, DistillDrive) on nuScenes where it achieves a 1.56× speedup at a 40% pruning ratio with no performance degradation.

Critical strengths

The method is evaluated across sparse, vectorized, diffusion-based, and distillation-based driving architectures and consistently outperforms various static and dynamic pruning baselines, demonstrating practical applicability and algorithmic robustness.

Critical weaknesses

Already stated in the Questions.

Questions for authors
  1. Given the large number of data-pruning methods, please further clarify the optimization perspectives of related pruning methods from the past three years and compare them with the proposed method, so as to better highlight the novelty of the pruning approach in this paper.
  2. Are there any simulation visualization results that support the conclusions of this paper?
  3. Are there experiments on other datasets, and do they exhibit behavior consistent with the conclusions of this paper?
Reviewer_A7dUOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AutoPrune, a data pruning method for end-to-end autonomous driving model training, aiming to reduce the amount of data used during training while preserving model performance as much as possible. The method consists of two main components. First, the authors introduce a gradient-preserving data pruning strategy that identifies low-value samples based on sample loss, randomly prunes a portion of these samples, and compensates for their gradient contributions using inverse probability weighting. Second, they design a dynamically adaptive training mechanism that updates the pruning range based on quantiles of sample losses and adjusts the learning rate schedule according to the number of pruned iterations, ensuring alignment with the original training progress. Experiments are conducted on the nuScenes dataset using models such as SparseDrive, VAD, DiffusionDrive, and DistillDrive. Results show that at a nominal 40% data pruning rate, AutoPrune maintains planning performance close to full-data training and achieves a 1.56× training speedup in the SparseDrive experiment. Overall, this work compresses training data for end-to-end autonomous driving models by integrating loss-based data selection, gradient reweighting, and learning rate scheduling adjustments.

Critical strengths

(1) The research problem has significant practical relevance. End-to-end autonomous driving models typically rely on large amounts of training data, resulting in high training overhead. To address this issue, this paper proposes a data pruning scheme that considers the potential changes in optimization signals caused by directly removing samples. Compared to methods that simply filter samples based on their importance, the authors further investigate the impact of data pruning on the training process, offering practical engineering value.

(2) The experiments include multiple models and ablation studies for comparison. The paper analyzes the impact of different pruning ratios on SparseDrive and compares it with various static and dynamic data selection methods. Additionally, the authors validate their approach on VAD, DiffusionDrive, and DistillDrive, and conduct ablation studies to examine the effects of gradient compensation, dynamic sample selection, and learning rate adjustment. These experiments provide some support for the effectiveness of the method.

Critical weaknesses

(1) The method has relatively limited innovation, as its core technical design is built upon existing approaches. This is my main concern regarding the paper. The core design of AutoPrune primarily includes loss-based sample selection, random pruning, inverse probability weighting, and learning rate scheduling adjustments. Among these, using loss to measure sample value, dynamically updating sample selection, and applying inverse probability weighting are all relatively common techniques in existing research. The gradient preservation strategy proposed in the paper essentially relies on inverse probability weighting to ensure that the weighted gradients of randomly retained samples match, in expectation, the full-data gradient. While this design is reasonable, the underlying derivation mainly stems from standard unbiased estimation principles and does not represent a significant theoretical breakthrough. Similarly, the training dynamic adaptation component primarily uses loss quantiles to determine the sample retention range and adjusts the learning rate based on the ratio between actual and original iteration numbers. These designs are rather straightforward and largely constitute combinations and engineering adaptations of existing techniques. Therefore, I believe the paper needs to further clarify what substantive new mechanisms AutoPrune introduces compared to existing dynamic data pruning and importance sampling methods, rather than merely integrating multiple established strategies into an end-to-end autonomous driving training framework.

(2) The theoretical analysis of gradient preservation is limited and insufficient to fully explain the observed training performance. Section 3.2 proves, via inverse probability weighting, that the pruned weighted aggregated gradient is consistent with the full-data gradient in expectation under fixed model parameters and sample partitioning. However, during actual training, model parameters are continuously updated and gradients are correlated. Particularly under mini-batch stochastic optimization and adaptive optimizers, expected consistency of a single aggregated gradient does not imply alignment in model update direction or final optimization trajectory. Moreover, Figure 2(b) primarily validates the gradient compensation effect by comparing the sum of gradient norms across different training methods. Yet, similar gradient norms do not necessarily indicate similar gradient directions, nor do they sufficiently demonstrate that gradient compensation restores the optimization behavior of full-data training. Although the authors mention this limitation in the appendix, given that Gradient-Preserving is one of the core contributions of this paper, I believe the current theoretical analysis and experimental evidence remain insufficient. Further investigation into the similarity of gradient directions before and after compensation, estimated variance, and their impact on optimization stability is recommended.

(3) The training acceleration benefits lack sufficient validation regarding actual computational costs. The paper presents training acceleration as one of its main contributions, reporting speedup ratios of 1.22x, 1.56x, and 2.17x in Table 1. Notably, these values exactly match the theoretical reduction in number of iterations calculated based on the proportion of reduced training data. For instance, if training consists of ten rounds—using full data in the first round and pruning 20%, 40%, and 60% respectively in the subsequent nine rounds—the theoretical speedups would precisely be approximately 1.22x, 1.56x, and 2.17x. Therefore, it remains unclear whether these speedup figures are derived from actual end-to-end training time measurements or merely computed based on reduced iteration counts. In practice, dynamic sample selection, loss tracking maintenance, data reorganization, and other operations may introduce additional overhead. The paper fails to adequately present real training time comparisons that include these costs, and lacks analysis on training duration and GPU resource consumption across different models. Furthermore, the primary acceleration experiments only involve the second stage of SparseDrive's training, rather than the complete two-stage process. Consequently, I believe the current experimental setup is insufficient to fully support the paper’s overall conclusion regarding end-to-end training acceleration.

(4) Experimental validation remains incomplete, and the fairness of comparison methods and reliability of results require further strengthening. Although the paper evaluates multiple end-to-end autonomous driving models, all experiments are based on the nuScenes dataset, and cross-architecture experiments use a randomly sampled 50% subset of the original training set. This somewhat limits the persuasiveness of the conclusions when generalizing to full-scale training and other data distributions. Furthermore, the paper primarily relies on open-loop planning metrics for evaluation, including trajectory L2 error and collision rate. For autonomous driving tasks, these metrics alone are insufficient to fully reflect a model's actual planning capability in complex interactive scenarios. On the other hand, Appendix C shows that some compared methods have undergone substantial reconfiguration. For example, RHO-Loss uses historical minimum sample loss instead of the reducible holdout loss calculated by an independent reference model as originally proposed. Such modifications may alter the original method's sample selection behavior, meaning the current comparisons might not adequately represent the performance of these methods under reasonable configurations. Finally, the paper does not explicitly report means, standard deviations, or confidence intervals from multiple independent training runs. Some performance differences are small, particularly in collision rate, so further replication experiments are needed to validate the stability of these improvements.

Questions for authors

Q1. Regarding the substantive innovation of the method. The main modules of AutoPrune are based on loss ranking, inverse probability gradient weighting, and learning rate scheduling mapping. These techniques have been widely applied in data selection and stochastic optimization research. Could the authors further clarify what the core methodological innovation of this paper is compared to existing dynamic data pruning approaches? In particular, what is the substantive difference between the gradient preservation mechanism proposed in this paper and conventional inverse probability importance weighting?

Q2. Regarding the calculation method for training speedup ratio. The speedup ratios of 1.22, 1.56, and 2.17 reported in Table 1 exactly match the theoretical reduction in training iterations calculated based on pruning ratios. Are these values derived from actual training time measurements or computed based on effective training iterations? Could you provide a comparison of actual wall-clock times, including data loading, sample selection, loss updates, and the entire training process, and clarify the overall acceleration gain achieved in the full two-stage training?

Q3. Regarding the effectiveness of gradient compensation. Section 3.2's theoretical analysis primarily ensures the consistency of aggregated gradients in expectation under a fixed model setting, while Figure 2(b) only shows the change in the sum of gradient norms. Could the authors further provide insights into the directional similarity between the compensated gradient and the full-data gradient, estimated variance, or differences in optimization trajectories? Such analyses would help determine whether gradient compensation genuinely improves the optimization process, rather than merely altering the magnitude scale of the gradients.

Q4. Regarding the fairness of comparative experiments and the stability of results. Some baseline methods in Appendix C employ different sample scoring mechanisms than the original implementation, particularly the redefined reducible loss in RHO-Loss. Could the authors further clarify whether these adaptations affect the performance of the baselines? Additionally, could they provide experimental results and statistical intervals across different random seeds to verify whether AutoPrune's improvements in planning error and collision rate are consistent?

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

The paper targets training-cost reduction for end-to-end autonomous driving via data pruning. It argues that pruning low-loss samples distorts the gradient composition and desynchronizes the LR schedule, and proposes AutoPrune: (i) uniform random removal of a fraction ρ of low-loss samples with inverse-probability reweighting 1/(1−ρ) on the retained ones; (ii) loss-quantile gating that recomputes the high-loss protection threshold every epoch; (iii) pruning-aware LR alignment that maps the reduced iteration budget onto the original full-data schedule. Experiments on nuScenes with SparseDrive-Small (stage 2), plus VAD/DiffusionDrive/DistillDrive under a reduced protocol, report parity with full data at 40% overall pruning (1.56× speedup) and ablations for the three components.

Critical strengths

(1) The problem is real and practically important, and the two diagnosed failure modes (gradient distortion, schedule misalignment) are sensible and rarely addressed together. (2) The algorithm is simple, cheap, and clearly specified (Algorithm A.1), with only two hyperparameters and a stated rationale for γ = 20%. (3) The paper compares against 7 adapted static/dynamic selectors and reports per-horizon metrics in the appendix; an anonymous reference implementation is provided.

Critical weaknesses

(1) The theoretical claim is a one-line Horvitz–Thompson identity conditioned on a fixed model snapshot and a fixed low-loss set; it says nothing about the trajectory, the variance amplification 1/(1−ρ), or the non-linear optimizer map, so it cannot support "gradient preservation" as a performance guarantee. (2) No compute-matched baseline (full data trained for 0.6× epochs); without it, parity at 40% pruning is not evidence of efficiency. (3) All main results are single-run; collision-rate differences correspond to a handful of events out of 6019 scenes, yet are discussed as percentage improvements. (4) The cross-architecture protocol silently uses a fixed 50% subset for 10 epochs, which is not the 40%-pruning-of-full-data setting claimed in the abstract. (5) Per-sample multi-task loss construction, gating cost, and stale-loss feedback are unspecified.

Questions for authors

The core mechanism and its theoretical justification are essentially InfoBatch's soft pruning with expectation rescaling; InfoBatch is neither cited nor compared, and the two genuine deltas — quantile gating and pruning-aware LR alignment — are never isolated, since the ablation changes the iteration budget together with the schedule. The empirical case is equally thin: one dataset, one seed per run, no compute-matched full-data baseline, collision-rate differences amounting to a handful of events out of 6019 scenes, and a cross-architecture protocol that trains on a fixed 50% subset for 10 epochs while the abstract describes a 40% full-data comparison. Until the authors compare against InfoBatch under a matched protocol, report ≥3 seeds with a compute-matched baseline, define the scalar per-sample loss, and quantify the stale-loss feedback loop, the full-data-parity claim is not supported.

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

Related papers

Discussion (0)

Sign in to comment.