acceptodds
Under review as a conference paper at ICLR 2027

How One Training Example Changes Others: Exact Route Attribution for BatchNorm

Abstract

In models with batch normalization, an example affects learning through both its own loss and the shared statistics used by other examples. Training-data attribution measures how individual examples influence a model, but a total influence score does not explain how that influence arises. We introduce Edge-Game Replay (EGR) to separate these contributions as they propagate through training. EGR intervenes on loss participation, normalization statistics, later batch membership, and running-buffer updates, and uses exact Shapley values over complete training replays to allocate the final evaluation-loss difference. We find that shared statistics and loss participation make contributions of comparable magnitude. Across 16 CIFAR-10 ResNet runs, adding statistic contributions to the loss contribution reduces mean absolute prediction error by 66.9% after 393 updates and raises Kendall correlation from 0.391 to 0.822 on a disjoint evaluation set. The resulting score ranks deletions more accurately than TracInCP, last-layer influence functions, and single-model TRAK at all three comparison horizons. Eight CIFAR-100 WideResNet runs also show improved two-update prediction. The decomposition also guides the choice among loss masking, exclusion from normalization statistics, and example deletion.

Primary area:
interpretability and explainable AI
Keywords:
batch normalization, training data attribution, training dynamics, Shapley values, data deletion

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_SH2XOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Edge-Game Replay (EGR), a framework that replays the training to quantify a sample's influence on the model. This process separates the training into paths such as supervised loss and the effect on shared batch statistics, using Shapley values to measure the contribution of each of these paths to the final loss. The paper claims that this helps quantify influence in models that use batch normalization.

Critical strengths

The paper proposes a novel framework to quantify a data sample's influence. The empirical gains shown in the paper are substantial.

Critical weaknesses

Computationally, this framework is expensive. So the practical usage of this may be limited. Another big weakness is that this framework only applies to architectures with BatchNorm, so it can not be applied as is to more modern architectures. Additionally, while the paper does cite recent trajectory-based data attribution methods such as SOURCE [1] and Simfluence [2], it does not show empirical comparisons of performance.

  1. Bae, Juhan, et al. "Training data attribution via approximate unrolling." NeurIPS (2024).
  2. Guu, Kelvin, et al. "Simfluence: Modeling the influence of individual training examples by simulating training runs." (2023).
Questions for authors
  1. How can this framework be extended to more modern architectures that use LayerNorm, for example?
  2. Can there be some approximations done to improve the computational overhead?
  3. How does this method compare to SOURCE and Simfluence?
Reviewer_YjN5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how individual training examples influence model learning through shared BatchNorm statistics, beyond their direct loss contributions. The authors propose Edge-Game Replay (EGR), a framework that decomposes the effect of removing a training example into distinct computational routes, including loss participation, normalization statistics, running-buffer updates, and subsequent batch membership. EGR uses counterfactual training replays and exact Shapley values to quantify the contribution of each route to the final evaluation loss. Experiments demonstrate that BatchNorm-mediated contributions can be comparable in magnitude to direct loss contributions, and incorporating these effects improves the prediction of sample deletion outcomes over several existing attribution methods.

Critical strengths
  1. Interesting and novel research perspective. The paper investigates an interesting and relatively underexplored aspect of training-data attribution: how individual training examples influence others through shared BatchNorm statistics. By decomposing sample influence into distinct computational routes, the work provides a novel perspective on understanding how training examples affect model learning beyond their direct loss contributions.
Critical weaknesses
  1. The experiments primarily examine route contributions through training continuation from pretrained checkpoints, rather than interventions applied throughout training from random initialization. It remains unclear whether the observed importance of BatchNorm-mediated influence and the benefits of selective interventions persist in full training runs.
  2. Limited experimental scale and scalability. The experiments are restricted to relatively small datasets and models, primarily CIFAR-10/100 with ResNet and WideResNet architectures. The largest model contains only 36.5M parameters and is evaluated with just two continuation updates. Moreover, EGR requires 16 or 32 full training replays per example, raising concerns about its computational scalability to larger models and longer training trajectories.
  3. Limited applicability beyond BatchNorm. EGR specifically targets cross-example dependencies introduced by BatchNorm, limiting its applicability to architectures using other normalization techniques. In particular, modern architectures such as Transformers and LLMs predominantly employ LayerNorm or RMSNorm, which do not introduce the same cross-example dependencies. Consequently, the proposed decomposition has limited direct applicability to these widely used architectures, and its generalization to other forms of sample interactions remains unexplored.
Questions for authors

See Weaknesses.

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

The paper introduces Edge-Game Replay to decompose a training example’s deletion effect in BatchNorm networks into contributions from loss participation, normalization statistics, running-buffer updates, and later batch membership. Matched training replays evaluate route interventions, and Shapley values allocate the endpoint loss difference. Experiments examine deletion-effect prediction, mechanism controls, route partitions, and selective training interventions.

Critical strengths

The paper addresses a meaningful distinction between an example’s supervision and its contribution to shared normalization statistics. Explicit route interventions and matched replays provide a useful framework for studying their downstream effects.

The mechanism experiments are carefully controlled. Normalization controls, score-permutation tests, batch-size analyses, and alternative route partitions support the importance of statistic contributions. The numerical-replay checks and transparent reporting strengthen the study.

Critical weaknesses

My main reservation concerns the additional value of attribution-guided intervention selection. The intervention study demonstrates gains over ordinary training, but the best reported restricted EGR policy achieves 85.09% mean accuracy versus 85.01% for four-action direct search, with runtimes of approximately 419 and 87 seconds. Recalculating the released endpoints for the 16 paired models gives a mean accuracy difference of +0.088 percentage points, with a descriptive, unadjusted paired-t 95% interval of approximately [−0.040, +0.216].

The authors already provide these comparisons and supporting data. My concern is that the additional training benefit of the attribution-based selection rule remains insufficiently established relative to its cost. I appreciate the mechanism analysis, but am not yet convinced that the combined contribution is substantial enough for acceptance, particularly the claimed value of using the decomposition to guide interventions.

The mechanism analysis is informative, but the additional value of attribution-guided intervention remains insufficiently established. A convincing clarification or analysis of this point could change my recommendation.

Questions for authors

Given that all 16 coalition losses are computed before Shapley signs restrict eligible actions, what additional benefit does this filtering provide over direct search among the same available actions? Can the existing runs identify when this restriction improves held-out outcomes?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.