acceptodds
Under review as a conference paper at ICLR 2027

Groups Don't Move as One: A Bregman Clustering View of Linearized Group Influence

10 views

Abstract

Influence functions provide an efficient alternative to retraining for assessing the effects of removing or reweighting training data. For groups, however, the standard first-order estimate reduces individual contributions to a sum that can conceal substantial differences among members, including opposing effects that cancel. We establish a Bregman centroid formulation of the linearized group influence objective, interpreting this aggregate as a single representative and using Bregman information (BI) to quantify the variation it omits. Extending this formulation to multiple clusters reveals distinct contributions whose sum recovers the original estimate. We further propose an estimator that jointly optimizes the weights of these cluster contributions, rather than simply summing them. Experiments uncover cluster-specific influence patterns and a positive association between target-specific BI and approximation error. The proposed estimator achieves higher than standard group influence in predicting target changes under the nonlinear proximal Bregman response function. Simplified clustering and correction variants reduce computational cost. The centroid view thus explains what additive group influence leaves out and provides a basis for using that information to improve estimation.

Primary area:
interpretability and explainable AI
Keywords:
influence functions, group influence functions, proximal Bregman response function, Bregman divergence, clustering with Bregman divergence, Bregman information

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 4 reviews, above 27% of the others.
Reviewer_KRQNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the influence function, and it claims that the standard first-order estimate reduces individual contributions to a sum, where the substantial differences among members are hidden. Therefore, they establish a Bregman centroid formulation of the linearized group influence function, and utilize the Bregman information to quantify the variation. The proposed estimator achieves higher R2 than standard group influence in predicting target changes.

Critical strengths

N/A

Critical weaknesses
  1. I strongly doubt this submission was finished by AI agent, since I find it is really hard to follow and understand.
  2. The experiments are too simple and toy. Now this paper is still focused on MNIST with MLP, while I believe recent papers of the influence function have conducted the empirical experiments on LLMs and even VLMs. Therefore, I do not think the experiments of this paper is solid and convincing.
Questions for authors

Refer to the weakness.

3 comments

We thank the reviewer for their feedback.


W2. The experiments are too simple and toy. Now this paper is still focused on MNIST with MLP, while I believe recent papers of the influence function have conducted the empirical experiments on LLMs and even VLMs. Therefore, I do not think the experiments of this paper is solid and convincing.

A2. Our experiments are a controlled mechanism study of how group structure affects influence prediction. We vary the partition and component weighting while holding the source model and nonlinear PBRF reference fixed. The study covers Dogfish logistic regression, MNIST MLP, and CIFAR-10 ResNet9. These results do not establish effectiveness at LLM/VLM scale.


W1. I strongly doubt this submission was finished by AI agent, since I find it is really hard to follow and understand.

A1. As disclosed in the manuscript, we used generative AI to assist our research and writing, and reviewed and verified the AI-assisted work. We made the final decisions on methodology, interpretation, and presentation and take responsibility for the final content.

Thanks, i would like to keep my score

We would like to clarify our earlier response regarding AI assistance. The authors formulated the research questions, developed the core methodology, and wrote the initial drafts of the manuscript, establishing their structure and flow, and used AI to help polish the language. Our previous response emphasized verification and responsibility but did not clearly articulate the authors’ role in developing the research itself.

We also take the concern about clarity seriously. Could the reviewer point to specific passages, sections, or paragraphs that were difficult to follow? Concrete examples would help us understand what made the impression that the submission was generated by an AI agent.

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

The paper studies heterogeneity hidden by additive group influence. It rewrites the quadratic objective underlying damped-GGN influence as a Bregman centroid problem, introduces dispersion measures, and partitions a removal group into components whose contributions exactly recover ordinary GI. It then proposes Joint E2, which reweights these fixed component directions using a quadratic approximation that accounts for the removed group’s curvature. Experiments show improved prediction of nonlinear PBRF responses on logistic regression, an MNIST MLP, and CIFAR-10 ResNet9. Euclidean clustering and a cheaper within-component curvature approximation reduce selected computational costs. The contribution combines a geometric interpretation with a reduced-space curvature correction.

Critical strengths
  1. An important problem to study. Group influence is an important research question and I think it's currently under studied.

  2. An interesitng proposal. Overall, I think using the clustering is interesting for group influence to reduce the complexity.

Critical weaknesses
  1. Limited validation of practical usefulness. PBRF is a justified evaluation target, but improved PBRF prediction does not establish improved prediction of ordinary deletion-and-retraining effects. It'll be useful to have a comparison with the usual correlation-style metric (such as LDS) where we use retraining as the baseline instead of PBRF.

I think it's a question of "approximating the exact effect" (i.e., the classical TDA literature aims to approximate the exact LOO effect) or "exact computation of the approximate effect" (i.e., PBRF, where exact computation might be possible but is itself an approximation).

  1. The method seems to be heuristics. It's still unclear to me why clustering responses should yield a better approximation.
Questions for authors

See weaknesses.

3 comments

We thank the reviewer for their valuable insights and thoughtful comments.


W1. Limited validation of practical usefulness. PBRF is a justified evaluation target, but improved PBRF prediction does not establish improved prediction of ordinary deletion-and-retraining effects. It'll be useful to have a comparison with the usual correlation-style metric (such as LDS) where we use retraining as the baseline instead of PBRF. I think it's a question of "approximating the exact effect" (i.e., the classical TDA literature aims to approximate the exact LOO effect) or "exact computation of the approximate effect" (i.e., PBRF, where exact computation might be possible but is itself an approximation).

A1. Our correction, applied to the retraining objective, improves prediction on Dogfish, most clearly at 1%1\% and 10%10\% removal. MNIST shows no consistent gain.

The table reports median query-wise Spearman / raw R2R^2 against actual leave-group-out retraining.

Model / removalGI (Group Influence)K1 E2Euclidean Joint K2
Dogfish 1%1\%0.8930.893 / 0.2360.2360.8970.897 / 0.7290.7290.9820.982 / 0.8810.881
Dogfish 10%10\%0.8060.806 / 0.06040.06040.8090.809 / 0.3580.3580.8870.887 / 0.4580.458
Dogfish 30%30\%0.7680.768 / −0.0547-0.05470.7660.766 / 0.02020.02020.7890.789 / 0.06580.0658
MNIST 1%1\%0.03500.0350 / −0.171-0.1710.03550.0355 / −0.171-0.1710.03470.0347 / −0.171-0.171
MNIST 10%10\%0.1610.161 / −0.152-0.1520.1620.162 / −0.152-0.1520.1650.165 / −0.155-0.155
MNIST 30%30\%0.2150.215 / −0.146-0.1460.2160.216 / −0.147-0.1470.2150.215 / −0.146-0.146

W2. The method seems to be heuristics. It's still unclear to me why clustering responses should yield a better approximation.

A2. Our intuition is that clustering captures response directions that are useful for correcting group influence. Under the same K=2K=2 Joint E2 correction, clustering outperforms the tested random partitions.

We compare two controls: cluster-size-matched random preserves the two Bregman cluster sizes and shuffles membership; equal-size random splits each group into random halves. Entries are median raw R2R^2 against PBRF; ranges span three random-partition seeds.

Model / removalK1 E2Bregman K2Euclidean K2Cluster-size-matched random K2Equal-size random K2
Dogfish 1%1\%0.9960.9960.9990.9990.9990.9990.996–0.9970.996\text{–}0.9970.9980.998
Dogfish 10%10\%0.9580.9580.9890.9890.9890.9890.961–0.9620.961\text{–}0.9620.971–0.9720.971\text{–}0.972
Dogfish 30%30\%0.8920.8920.9380.9380.9380.9380.899–0.9030.899\text{–}0.9030.904–0.9090.904\text{–}0.909
MNIST 1%1\%0.9980.9980.9990.9990.9990.9990.9980.9980.9990.999
MNIST 10%10\%0.9920.9920.9930.9930.9940.9940.9920.9920.9920.992
MNIST 30%30\%0.9520.9520.9560.9560.9580.9580.9520.9520.952–0.9530.952\text{–}0.953

Thanks for the response. I appreciate A1, but to me, the results on such simple dataset is itself a limitation (which I don't think carries that much weight to begin with as a weakness). I care much more about W2. I still do not think the provided answer provides useful insight into the design philosophy or provides an explanation of the underlying mechanism. I think a deeper insight into the design rationale and an (either experimentally or theoretically rigorous) explanation is needed.

Thank you for clarifying your main concern.

Our work starts from the intuition that a simple sum can obscure a group’s internal response structure. To understand this, we examined the objective underlying the influence estimate. It seeks a single representative that minimizes the total Bregman divergence from the individual response points.

Yet this single representative can conceal a counterintuitive fact: a very small subgroup can carry substantial influence. In our MNIST clustering, omitting the contribution of the smaller component (median: 2 of 6,000 examples) from GI reduces Spearman from 0.997 to 0.794 and R2R^2 from 0.978 to 0.630.

We explore this component structure as a basis for correcting group influence. Clustering constructs multiple representatives by minimizing total Bregman divergence. E2 then optimizes one weight per component under a second-order objective that incorporates the removed group’s curvature. This objective can favor different corrections along different response directions. Under clustering, similar responses share a component weight, whereas random partitions can tie different patterns to the same weight.

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

The paper establishes a Bregman centroid formulation of the linearized group influence objective and proposes an estimator that jointly optimizes the weights of cluster contributions. Experiments on logistic regression, an MLP, and ResNet9 show an empirical association between projected Bregman information and approximation error, and improved prediction of nonlinear PBRF responses over standard group influence.

Critical strengths
  1. The paper offers a novel and insightful perspective on the Bregman centroid of linearized group influence, revealing within-group variation and opposing contributions hidden by standard GI.
  2. Systematic experiments across three models support the effectiveness of component reweighting for predicting nonlinear PBRF responses.
Critical weaknesses
  1. These experiments do not establish whether the gains come from clustering or from additional component directions.
  2. Evaluation uses only the nonlinear PBRF response. A comparison with actual data removal and retraining would help assess whether the reported gains extend to retraining effects.
Questions for authors
  1. Why is the evaluation limited to nonlinear PBRF responses? Could a comparison experiment be conducted with actual data removal and retraining, to see whether the reported gains extend to that setting?
  2. Moving from single-component E2 to two components expands the response space from a line to a plane when the directions are linearly independent. Would random partitions or alternative response directions yield similar gains, or is clustering essential?
  3. What explains the varying performance of Intra E2, which approaches Joint E2 in some settings but underperforms single-component E2 in others?
3 comments

We appreciate the reviewer's valuable comments and feedback.


W1. These experiments do not establish whether the gains come from clustering or from additional component directions.

Q2. Moving from single-component E2 to two components expands the response space from a line to a plane when the directions are linearly independent. Would random partitions or alternative response directions yield similar gains, or is clustering essential?

A1. Clustering outperforms the tested random partitions at the same K=2K=2, so adding a second coefficient alone does not explain the gains.

We compare Joint E2 with two controls: cluster-size-matched random preserves the two Bregman cluster sizes and shuffles membership; equal-size random splits each group into equal halves. Entries are median raw R2R^2 against PBRF, with ranges across three random-partition seeds.

Model / removalK1 E2Bregman K2Euclidean K2Cluster-size-matched random K2Equal-size random K2
Dogfish 1%1\%0.9960.9960.9990.9990.9990.9990.996–0.9970.996\text{–}0.9970.9980.998
Dogfish 10%10\%0.9580.9580.9890.9890.9890.9890.961–0.9620.961\text{–}0.9620.971–0.9720.971\text{–}0.972
Dogfish 30%30\%0.8920.8920.9380.9380.9380.9380.899–0.9030.899\text{–}0.9030.904–0.9090.904\text{–}0.909
MNIST 1%1\%0.9980.9980.9990.9990.9990.9990.9980.9980.9990.999
MNIST 10%10\%0.9920.9920.9930.9930.9940.9940.9920.9920.9920.992
MNIST 30%30\%0.9520.9520.9560.9560.9580.9580.9520.9520.952–0.9530.952\text{–}0.953

The gains are smaller on MNIST and can disappear after rounding. Since the size-matched control preserves the imbalance, the benefit also depends on which examples form each component.


W2. Evaluation uses only the nonlinear PBRF response. A comparison with actual data removal and retraining would help assess whether the reported gains extend to retraining effects.

Q1. Why is the evaluation limited to nonlinear PBRF responses? Could a comparison experiment be conducted with actual data removal and retraining, to see whether the reported gains extend to that setting?

A2. The gains extend to ordinary retraining on Dogfish, most clearly at 1%1\% and 10%10\% removal, but are not consistent on MNIST.

PBRF is the primary reference because its local response yields the damped-GGN estimate studied here [1]. For LGO, we adapt the correction to the retained training objective. The table reports median query-wise Spearman / raw R2R^2 against retraining.

Model / removalGI (Group Influence)K1 E2Euclidean Joint K2
Dogfish 1%1\%0.8930.893 / 0.2360.2360.8970.897 / 0.7290.7290.9820.982 / 0.8810.881
Dogfish 10%10\%0.8060.806 / 0.06040.06040.8090.809 / 0.3580.3580.8870.887 / 0.4580.458
Dogfish 30%30\%0.7680.768 / −0.0547-0.05470.7660.766 / 0.02020.02020.7890.789 / 0.06580.0658
MNIST 1%1\%0.03500.0350 / −0.171-0.1710.03550.0355 / −0.171-0.1710.03470.0347 / −0.171-0.171
MNIST 10%10\%0.1610.161 / −0.152-0.1520.1620.162 / −0.152-0.1520.1650.165 / −0.155-0.155
MNIST 30%30\%0.2150.215 / −0.146-0.1460.2160.216 / −0.147-0.1470.2150.215 / −0.146-0.146

[1] Bae et al. (2022). If Influence Functions are the Answer, Then What is the Question? NeurIPS.


Q3. What explains the varying performance of Intra E2, which approaches Joint E2 in some settings but underperforms single-component E2 in others?

A3. Intra trades accuracy for cheaper edit-curvature computation; the omitted terms matter more at MNIST 30%30\% removal. Intra evaluates each direction only on its own component's examples, whereas Joint uses the whole removed group.

To test this explanation, we restore each direction's curvature over the whole group while still omitting cross-direction edit terms (full-group diagonal edit). All K2 variants below use the same Euclidean partitions and directions; entries are median raw R2R^2 against PBRF.

MNIST removalK1 E2Intra K2Full-group diagonal edit K2Joint K2
1%1\%0.9980.9981.001.001.001.001.001.00
10%10\%0.9930.9930.9940.9940.9940.9940.9950.995
30%30\%0.9640.9640.9620.9620.9680.9680.9680.968

At 30%30\%, Intra's raw R2R^2 decrease from K1 is significant (p=2.34×10−4p=2.34\times10^{-4}), while restoring the same-direction terms brings accuracy close to Joint. This supports their omission as a contributor to the loss. Intra remains close to Joint at smaller removals and still improves Spearman over K1 at 30%30\%.

Thank you for your detailed reply. My previous concerns have been solved and I would like to keep my positive score.

We are glad our response resolved your concerns, and we will incorporate the discussed points into the final version.

Reviewer_M6sGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the author studies the effect of removing training data without retraining it. Baseline/standard methods add up the effect of each data point, which reduces the effect of individual groups that can conceal substantial differences among members; e.g. two subgroups pull in opposite directions and cancel each other’s effects. Also, the authors argue that variance of these differences correlates with inaccuracy of the standard group estimate. They then split each group into k clusters whose contributions add back to the original estimate, and propose "E2", which re-weights these cluster contributions under a leave-group-out curvature correction. Evaluation uses the nonlinear PBRF as reference on three small models with 100 random removal groups and 100 queries, reporting Spearman and raw R2.

Critical strengths

-> Report the raw R-squared and the paired significance tests across the shared queries and you'll see just how marginal improvements are. The baseline GI is already 0.978 on MNIST, and the fast Intra variant shows a significant drop in R-squared at 30% removal.

-> Measuring value accuracy and statistical significance against a self-referential PBRF proxy rather than actual retraining only tells you how well the estimator fits.

-> The theoretical formulation is algebraically trivial in that the decomposition holds for any arbitrary partition as linear sums commute and the projected Bregman information is just a re-arrangment of half the sample variance of the individual influence scores.

Critical weaknesses

-> The PBRF ground truth has the estimator's large ridge penalty, curvature matrix and GI initialisation, all of which lead to inflated linear agreement and importantly omits the true leave-group-out retraining even on small convex models.

-> All of the key protocol choices (number of components, ridge penalty, number of clustering restarts, solver budgets) and the original model training schedules (fixed number of epochs, L-BFGS budgets) have no validation protocol. The only mention of a 4000 example validation split for MNIST is not explained in any way and it appears that all of the protocol choices were simply hand tuned on the same 100 groups and test queries that were ultimately used for the final reporting.

-> The proposed multi-component estimator is nothing more than a subspace-restricted Newton step which is in fact outperformed by its own full-space version). The paper completely omits the runtime costs of Full E2, the end-to-end speedups and the comparisons against existing group influence corrections.

-> Subgroup discovery is only performed on uniform random subsets which leads to extremely outlying splits (5,997 vs 3) and unrelated rather than opposing profiles (median cosine of negative 0.03). There is no counterfactual subgroup removal and no random partition baselines.

-> No source code is included with the submission so we can't check the custom solver budgets, the matrix free curvature approximations, the timing boundaries or the experimental results.

Questions for authors

-> See Weaknesses:

And

-> What were the ridge values (0.1, 0.3, 0.01) based on? On what data?

-> Please report leave-group-out retraining as the reference.

-> Were the number of clusters, the ridge value, the number of clustering restarts and the solver settings all chosen on a separate set of groups and queries or on the same 100 groups and 100 queries that the results are reported on?

-> The ground truth is computed with the same curvature matrix and the same ridge value as the method itself, and the ridge is much larger than the regulariser used in training. How were these values chosen?

-> How well converged is the ground truth? Please report the final gradient norms for each solver in Table 4.

3 comments

We sincerely thank the reviewer for the thoughtful suggestions and feedback.


S1. Report the raw R-squared and the paired significance tests across the shared queries and you'll see just how marginal improvements are. The baseline GI is already 0.978 on MNIST, and the fast Intra variant shows a significant drop in R-squared at 30% removal.

A1. Euclidean K2 significantly improves raw R2R^2 over K1 in 1212 of 1313 comparisons (p<0.001p<0.001), with smaller gains on MNIST.

SettingGIK1 E2Euclidean Joint K2Euclidean Intra K2
Dogfish 30%30\%0.8620.8620.8910.8910.9450.9450.9450.945
MNIST 10%10\%0.9780.9780.9930.9930.9950.9950.9940.994
ResNet9 1%1\%0.8910.8910.9650.9650.9750.9750.9740.974

MNIST's already strong GI baseline leaves less room for improvement. Intra improves over GI and K1 in all three main settings while reducing edit-curvature time by 46.6%46.6\% on MNIST 10%10\%.

At MNIST 30%30\%, Intra falls below K1 (p=2.34×10−4p=2.34\times10^{-4}), while Joint still improves, suggesting that the curvature terms omitted for efficiency become more consequential at this larger removal ratio.


S2. Measuring value accuracy and statistical significance against a self-referential PBRF proxy rather than actual retraining only tells you how well the estimator fits.

W1 (1). ... omits the true leave-group-out retraining even on small convex models.

Q2. Please report leave-group-out retraining as the reference.

A2. Better PBRF prediction does not by itself establish better retraining prediction. In our LGO tests, an ERM-adapted correction improves Dogfish, most clearly at 1%1\% and 10%10\% removal, but MNIST shows no consistent gain.

PBRF is our primary reference because its local response yields the damped-GGN estimate studied here. Bae et al. [1] establish this connection for neural networks. This motivates evaluating PBRF separately from ordinary retraining. The nonlinear PBRF reference is optimized separately from the E2 quadratic.

For LGO, we adapt the correction to the retained training objective. The table reports median query-wise Spearman / raw R2R^2 against retraining.

Model / removalGIK1 E2Euclidean Joint K2
Dogfish 1%1\%0.8930.893 / 0.2360.2360.8970.897 / 0.7290.7290.9820.982 / 0.8810.881
Dogfish 10%10\%0.8060.806 / 0.06040.06040.8090.809 / 0.3580.3580.8870.887 / 0.4580.458
Dogfish 30%30\%0.7680.768 / −0.0547-0.05470.7660.766 / 0.02020.02020.7890.789 / 0.06580.0658
MNIST 1%1\%0.03500.0350 / −0.171-0.1710.03550.0355 / −0.171-0.1710.03470.0347 / −0.171-0.171
MNIST 10%10\%0.1610.161 / −0.152-0.1520.1620.162 / −0.152-0.1520.1650.165 / −0.155-0.155
MNIST 30%30\%0.2150.215 / −0.146-0.1460.2160.216 / −0.147-0.1470.2150.215 / −0.146-0.146

[1] Bae et al. (2022). If Influence Functions are the Answer, Then What is the Question? NeurIPS.


S3. The theoretical formulation is algebraically trivial in that the decomposition holds for any arbitrary partition as linear sums commute and the projected Bregman information is just a re-arrangment of half the sample variance of the individual influence scores.

A3. We agree that the GI decomposition holds for any partition and that projected BI is a scaled variance of individual influence scores.

Our contribution is to derive this structure from the influence objective and use it in E2. We prove that the linearized group PBRF objective equals a Bregman centroid objective up to a parameter-independent constant (Eq. (8)). The minimum of its centroid-loss term gives BI; applying the same construction to the query-projected response points yields projected BI. Clustering the response points in their original geometry (Eq. (13)) supplies components whose relative weights E2 optimizes to correct GI (Eq. (17)).

W1 (2). The PBRF ground truth has the estimator's large ridge penalty, curvature matrix and GI initialisation, all of which lead to inflated linear agreement and importantly omits the true leave-group-out retraining even on small convex models.

A4. The gains over GI persist at smaller ridge values, except at MNIST λ=0.001\lambda=0.001. Entries are median R2R^2 against PBRF recomputed at each λ\lambda.

SettingTotal λ\lambdaGIK1 E2Bregman Joint K2Euclidean Intra K2
Dogfish 30%30\%0.00020.0002−0.0482-0.04820.1440.1440.1720.1720.1790.179
Dogfish 30%30\%0.0010.0010.3190.3190.4910.4910.6030.6030.5980.598
Dogfish 30%30\%0.010.010.8620.8620.8910.8910.9450.9450.9450.945
Dogfish 30%30\%0.10.10.9940.9940.9970.9970.9980.9980.9980.998
MNIST 10%10\%0.0010.0010.3360.336−0.765-0.765−0.148-0.148−0.119-0.119
MNIST 10%10\%0.010.010.8020.8020.8360.8360.8890.8890.8950.895
MNIST 10%10\%0.10010.10010.9780.9780.9930.9930.9950.9950.9940.994
MNIST 10%10\%0.30.30.9940.9940.9980.9980.9980.9980.9980.998

At MNIST λ=0.001\lambda=0.001, K1/K2 underperform GI, possibly because weak damping allows larger parameter changes, making the local approximation less reliable.


W2. All of the key protocol choices (number of components, ridge penalty, number of clustering restarts, solver budgets) and the original model training schedules (fixed number of epochs, L-BFGS budgets) have no validation protocol. The only mention of a 4000 example validation split for MNIST is not explained in any way and it appears that all of the protocol choices were simply hand tuned on the same 100 groups and test queries that were ultimately used for the final reporting. Q3. Were the number of clusters, the ridge value, the number of clustering restarts and the solver settings all chosen on a separate set of groups and queries or on the same 100 groups and 100 queries that the results are reported on?

A5. The calibration samples were randomly drawn. Q1 describes the model-specific ridge-selection procedures.

We chose K=2K=2 by elbow inspection of the training-side within-component Bregman dispersion (Eq. (13)). MNIST's 4,0004{,}000 validation examples monitor source-model loss and accuracy and are disjoint from the final queries. MNIST damping, KK, restart counts, and solver budgets were not adjusted using prediction scores on the final 100100 groups or queries.

Clustering uses ten restarts for Dogfish and two for MNIST/ResNet9, selecting the partition with the lowest objective value (Eq. (13)). Training and solver budgets are reported in Appendix C.


W3. The proposed multi-component estimator is nothing more than a subspace-restricted Newton step which is in fact outperformed by its own full-space version). The paper completely omits the runtime costs of Full E2, the end-to-end speedups and the comparisons against existing group influence corrections.

A6. We agree that Joint E2 is a subspace-restricted quadratic correction. Full E2 is more accurate and faster than K2 in our 100100-group comparison.

The table reports median query-wise raw R2R^2 against PBRF and total attribution time on Dogfish 30%30\% (RTX3090, FP64). Times include one shared preparation and exclude training and reference optimization.

MethodMedian raw R2R^2Total time for 100100 groups (s)
GI0.8620.8620.2110.211
K1 E20.8910.8910.2270.227
Euclidean Joint K20.9450.9450.4770.477
Euclidean Intra K20.9450.9450.4650.465
Full E20.9900.9900.3550.355
Basu-PBRF0.9670.9670.2510.251
RIF-PBRF0.9600.9600.2490.249

Basu-PBRF and RIF-PBRF are PBRF-adapted baselines without clustering.

K2 is faster for smaller workloads, with Full E2 becoming faster between 4848 and 6464 groups. The following median total times (seconds) come from a separate sweep on another RTX3090, with one shared preparation per workload.

Method / number of groups1616323248486464
Euclidean Joint K20.2580.2580.3120.3120.3670.3670.4280.428
Euclidean Intra K20.2570.2570.3080.3080.3610.3610.4180.418
Full E20.3380.3380.3540.3540.3700.3700.3860.386

Our contribution is the objective-derived component structure, which supports both interpretation and correction of GI. Full E2 provides an aggregate response; our decomposition additionally identifies and reweights the contributing subsets. W4 illustrates small but influential components and components dominated by mislabeled examples.

W4. Subgroup discovery is only performed on uniform random subsets which leads to extremely outlying splits (5,997 vs 3) and unrelated rather than opposing profiles (median cosine of negative 0.03). There is no counterfactual subgroup removal and no random partition baselines.

A7. The original splits are indeed imbalanced, but the small component remains important for influence prediction. On MNIST 10%10\%, omitting its contribution from Bregman Joint K2 predictions lowers median R2R^2 against the full-group PBRF reference from 0.9950.995 to 0.6360.636.

The label-flip experiment further shows that clustering can isolate components dominated by mislabeled examples. We train an MNIST MLP after relabeling 20%20\% of digit-33 examples as 88, then apply Euclidean gradient clustering to 3,0003{,}000 examples labeled 88. The table shows one example group:

KKComponentsGroup share (%)Clean 88 (%)Flipped 33 (%)
221197.297.284.384.315.715.7
22112.772.7700100100
32323385.985.993.293.26.836.83
3232292914.114.113.713.786.386.3

Clean/flipped percentages are within-row fractions. At K=32K=32, the rows pool three clean-majority and 2929 flipped-majority components.


W5. No source code is included with the submission so we can't check the custom solver budgets, the matrix free curvature approximations, the timing boundaries or the experimental results.

A8. The source code is available in our anonymous repository.


Q1. What were the ridge values (0.1, 0.3, 0.01) based on? On what data?

Q4. The ground truth is computed with the same curvature matrix and the same ridge value as the method itself, and the ridge is much larger than the regulariser used in training. How were these values chosen?

A9. Ridge was selected by model-specific calibration or numerical criteria, then fixed across the main method comparisons.

  • Dogfish: λ=0.01\lambda=0.01 was selected by BI–error Spearman correlation on 1212 groups and 2020 queries, with six additional groups for checking.
  • MNIST: damping 0.10.1 was retained from the fixed numerical configuration, with inverse solves checked by relative residuals. The total ridge is 0.10010.1001 after including training-regularization curvature.
  • ResNet9: 0.30.3 was the smallest tested damping reaching relative residual 0.010.01 within 2020 CG iterations on a 5,0005{,}000-example probe. The total ridge is 0.30050.3005.

Using the same ridge is consistent because influence approximates the local response of the PBRF reference. At the original model, the unperturbed PBRF has Hessian GG, the same damped GGN used by influence.


Q5. How well converged is the ground truth? Please report the final gradient norms for each solver in Table 4.

A10. The final PBRF gradient norms are small across all three main settings. The table reports the median and maximum gradient ℓ2\ell_2 norms across all 100100 groups per setting.

SettingSolver protocol in Table 44Median gradient ℓ2\ell_2Maximum gradient ℓ2\ell_2
Dogfish 30%30\%Newton–PCG, L-BFGS fallback1.00×10−111.00 \times 10^{-11}8.64×10−118.64 \times 10^{-11}
MNIST 10%10\%Strong-Wolfe L-BFGS4.15×10−64.15 \times 10^{-6}2.49×10−52.49 \times 10^{-5}
ResNet9 1%1\%Strong-Wolfe L-BFGS, FP32 warm / FP64 refine3.64×10−53.64 \times 10^{-5}1.38×10−41.38 \times 10^{-4}

Longer optimization produced stable query responses on the tested MNIST and ResNet9 groups. The MNIST initialization check also showed little change.

Comments on the paper

1 comment
Area_Chair_rogUOpenReview ↗

Encouraging Reviewer Feedback on Author Responses

Dear Reviewers,

Thank you all for your time and effort in reviewing and responding. Just a gentle reminder to reviewers to take a look at the authors' responses and share any updated thoughts on whether your main concerns have been addressed and, if not, to indicate which points are not well addressed. The author–reviewer discussion period ends on November 18 (AoE), so I would greatly appreciate your input before then. Thank you again for your time and contributions!

Kind Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.