Theoretical and Practical Perspectives on what Influence Functions Do

Influence functions (IF) have been seen as a technique for explaining model predictions through the lens of the training data. Their utility is assumed to be in identifying training examples"responsible"for a prediction so that, for example, correcting a prediction is possible by intervening on those examples (removing or editing them) and retraining the model. However, recent empirical studies have shown that the existing methods of estimating IF predict the leave-one-out-and-retrain effect poorly. In order to understand the mismatch between the theoretical promise and the practical results, we analyse five assumptions made by IF methods which are problematic for modern-scale deep neural networks and which concern convexity, numeric stability, training trajectory and parameter divergence. This allows us to clarify what can be expected theoretically from IF. We show that while most assumptions can be addressed successfully, the parameter divergence poses a clear limitation on the predictive power of IF: influence fades over training time even with deterministic training. We illustrate this theoretical result with BERT and ResNet models. Another conclusion from the theoretical analysis is that IF are still useful for model debugging and correcting even though some of the assumptions made in prior work do not hold: using natural language processing and computer vision tasks, we verify that mis-predictions can be successfully corrected by taking only a few fine-tuning steps on influential examples.

Paper

Similar papers

Peer review

Reviewer qity7/10 · confidence 4/52023-07-05

Summary

This paper reexamines the assumptions used in the deduction of influence function (IF) methods in order to explain the failure of IF in predicting leave-some-out-retrain performance. The authors find that all five assumptions used in the previous deduction will be violated to different degrees in practice and propose a combination of HIF and Arnoldi-based methods to overcome these violations. They find that four of the assumptions can actually be bypassed or fixed, however, the remaining challenge named *parameter divergence* seems inherent. They show that the predictive ability of IF will gradually fade over training steps due to this phenomenon and show that accounting for this effect, it is better to interpret IFs as proxies for the effect of a few fine-tuning steps.

Strengths

* **Originality.** This paper shows an in-depth analysis of the assumptions on which IFs are based. Their analysis combines theoretical thinking well with experimental observations. * **Clarity.** The results are presented in a clean way and are a pleasure to read. * **Significance.** Although the reviewer is not an expert in the field of influence functions, the question this paper targets seems important and the paper sheds light on the real obstacles to solving this problem, as well as proposes possible ways to solve it.

Weaknesses

The reviewer thinks the paper may improve in the following aspects. * The authors argue that (1) using Arnoldi-based methods helps boost the accuracy of HIF, and (2) accounting for training trajectories can improve the estimation for IFs. The theoretical deduction of these propositions is very reasonable but it would be better to include ablation experiments or cite relevant literature to showcase the phenomenons. * The authors observe that IF methods perform poorly on Resnet but do not have any explanation for this phenomenon. It would be beneficial if the authors can investigate the reason behind the difference of IF methods performance on NLP and CV tasks.

Questions

The reviewer is interested in the following questions. * Why does the correlation between IFs and performances seem to increase before decreasing? * How would the authors suggest modifying the current IF estimation methods based on their observations? * What is the author's explanation of the rapidly increasing phase 1 in figure 1?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have adequately addressed the limitations and the reviewer has not noticed any potential negative societal impact.

Reviewer SKsY8/10 · confidence 4/52023-07-06

Summary

Influence functions allow one to measure the effect of up-weighting a training sample on the test loss (or functions of the model parameters) of a test example. This paper studies several of the assumptions needed for hessian-based influence functions to produce valid leave-out-one estimates of the effect on a test sample's loss. This assumptions include: convexity, stability of the hessian, and additivity of training trajectory. The paper shows that some of these assumptions are not as problematic, in practice, as previously assumed. They show that several of these assumptions can indeed be satisfied with changes to the formulation; for example, if the hessian becomes degenerate, one can approximate the hessian via arnoldi iteration. It turns out that the most problematic assumption results from a divergence between the initial set of parameters, and another obtained by re-training or even fine-tuning for longer time steps. The paper then bounds this parameter divergence using a discrete version of gronwall's lemma, and gives a set of takeaways for how to evaluate influence functions: the takeaway from this work is that influence scores only predict effect of a training sample over a small number of training steps. In the final portion of the paper, they evaluate how to use influence functions to correct errors.

Strengths

Overall, I really enjoyed reading this paper, and found the breakdown of the assumptions really comprehensive. **Quality/Clarity**\ I found the paper to be of very high quality. Each assumption is stated clearly and cleanly discussed. For example, I found the resolution of assumption #1 to be interesting. Here the authors show that for hessian-based influence functions one does not need to assume strict convexity but we can relax it to assume that the final gradient steps do not change the hessian by too much. While this might seem like a simple change, it helps provide clarity to this literature. The bigger takeway from this work is that influence scores only predict effect of a training sample over a small number of training steps. This finding is important since it requires rethinking how these approaches are currently used. Overall, this paper sets out key assumptions of influence functions, discusses how to remedy violations to these assumptions, and presents experiments to corroborate the theory presented. **Significance**\ One important takeaway is that, for deep models, it now makes sense that influence scores obtained via retraining have poor correlation with leave-one-out retraining scores. Secondly, another important takeaway is that one can correct errors simply be reweighting or fine-tuning on opponents or proponents. The findings from the paper should be important for practitioners and others using influence functions as part of their debugging toolbox. **Originality**\ This paper mostly considers a question that was also studied by Bae et. al. (Neurips 2022), but does so in a different way and using mostly new techniques. I found the use of the discrete version of gronwall's lemma to be quite nice. I am not sure if this is new, but taken together, this work advances the understanding of influence functions quite substantially.

Weaknesses

The main weakness I have with this work are minor and listed below: **Relation with Bae et. al. from Neurips 2022**: The main related work to this one is this paper. However, this work does not do enough to sufficiently contrast their analyses with theirs. I am mostly familiar with this previous work, and understand that they address exactly the same question as well, but arrive at a different conclusion. Their finding is that hessian-based influence functions approximate what they term the "proximal bregman response function", and that the ability of this quantity to match a sample loo estimate is mostly affected by the linearization error (i.e., due to the taylor expansion), and 2) hessian approximation. This work seems to not agree with theirs, i.e., you suggest that hessian approximation be handled with arnoldi, and that the key discrepancy comes from parameter divergence. It would be great if the authors could spend more time in the draft to contrast this work against that one. I think this is an important issue to address in the paper.

Questions

Some clarification questions: - For error correction, if I have k examples that I want to correct, then am I correcting each example independently or the taking gradient steps on all k examples at the same time. If it is independent, then isn't this too onerous? - One use of these influence scores is mislabelled training samples, how does assumption 5 affect that? When in training should the self-influence score be a useful metric for mislabeled label correction? - This is minor question, going from $\theta_\epsilon \approx \theta_0 + \epsilon^\top\nabla \theta_\epsilon$ to equation 1, where did the second $\nabla_\theta$ come from? Is this the chain rule? I am familiar with other ways to derive influence functions.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors discuss limitations.

Reviewer RiS87/10 · confidence 3/52023-07-06

Summary

This paper studies the limitations of Influence Functions (IF) based on assumptions about convexity, numerical stability, training trajectory and parameter divergence. The authors discuss each of those assumptions in detail and propose ideas to address the shortcoming of those limitations. The authors describe solutions for 4 assumptions made in the paper and discuss the limitations of the fifth assumption, parameter divergence, which is harder to address. Apart from that the authors also illustrate the theory of parameter divergence on the example of NLP and vision classifiers. They confirm theoretical findings of fading influence over time when retraining the models on vision and test classifiers. The authors also show how to perform error correction based on proponent-correction and opponent-tuning using a few fine tuning steps. They show that their approach outperforms baseline re-training procedures.

Strengths

+ The paper is well written, especially the introduction, problem statement and the assumptions are well framed. + The related work is well cited and based on the studies in the recent papers, this paper proposes to address the limitations highlighted in the literature such as the LSOR problem for influential examples. + The paper studies 5 important assumptions made by IF and proposes solutions that help to alleviate the limitations posed by those assumptions. It also discusses and proposes a solution for the temporal dependency which is not accounted for in the original TracIn paper.

Weaknesses

+ The main contributions of the paper seem to be around studying 5 different assumptions and proposing solutions for them. The main contribution is for assumptions 3 - 5 where the authors show that IF can predict \theta_{\epsilion, t} only for a limited period of time and propose incorporating temporal term into the tracin's formulation. These are important findings, however, it is not very clear whether it is a large enough contribution and whether it is completely novel since similar limitations of Tracin were also discussed in the Kelvin Guu, et.al. + It looks like all the proofs corresponding to theorems mentioned in the main paper are in the appendix. Since it is not required for reviewers to read appendices it might be good to consider moving some of the important proofs or parts of them into the main paper. + Section 4.2 might be difficult to follow for someone if they do not already know about Arnoldi. There is a brief description that Arnolodi approximates P1 using subspaces spanned by eigenvectors corresponding to top-k eigenvalues but I think that overall intuition of Arnoldi is not clear if someone is not familiar with that work. + Overall I think that Section 4.2 might be a bit difficult to follow (e.g. how is Theorem 1 applied to restricted variation). + It feels that it is somewhat difficult to follow how we arrive from Eq 7 to Eq 8. Perhaps adding more explanation in section 4.4 would be good. **Minor comments** + The definition of C^k function on line 179 is a bit unclear. C gets used again in line 234 (C\epsilion) which makes it a bit confusing in terms of the readability of the paper. + Figure 3: X-axis is not annotated

Questions

+ Section 6: Proponents-correction: when the set B is identified and relabeled, is it done for truly mislabeled examples ? How were those mislabeled examples sourced ? For experiments with SST and ResNet how were set Bs curated ? + Section 6: Which formulation of IF (TracIn or Influence Functions) was used in the experiments ? + Section 5.2: Is R(t) pearson correlation between losses before and after re-training averaged for the same batches? Does the `Step` indicate different checkpoint on Figure 2 ?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors discuss several limitations of their work and propose ideas to address those limitations in the future.

Reviewer 7Fho7/10 · confidence 2/52023-07-21

Summary

This paper verifies assumptions of Influence Functions that show discrepancies in theory and empirical results, convexity, numeric stability, training trajectory, and parameter divergence. Although the other assumptions are addressable, parameter divergence is not. Hence, the paper proposed a solution to it and also showed empirical results.

Strengths

- Based on their observation of the issue, they proposed a simple solution to it. - To the reviewer, this paper reads as an instructional paper, which can contribute to the community. - It verified the applicability of the IF assumption one by one on convexity, numeric stability, training trajectory, and parameter divergence.

Weaknesses

- Aside from its technical contribution, there’s room where its writing can be improved. In particular, the introduction can be better structured, and there are multiple repetitive places that can be concise. - In Fig 2 (1) BERT case, although the authors claimed that the predictive power degrades monotonically over time, the “fading” effect is actually not clear. There’s only one time of peak at an early step, and there's no decreasing trend in the later steps. There is no fading effect either in the Appendix (Fig. 5.)

Questions

Do the authors have other models to show the fading effect? As BERT’s case is unclear, only one empirical case of ResNet shows it.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes.

Reviewer qity2023-08-11

Response to Rebuttal

Extended results on ViT will definitely improve the paper. The reviewer has read the response and decided to keep the score.

Reviewer RiS82023-08-14

Rebuttal Response

Thank you authors for the clarification and detailed responses to my questions. I've increased the rating by 1 point.

Reviewer SKsY2023-08-15

Response to rebuttal

Thanks to the authors for responding to my questions. The response here w.r.t. Bae et. al. is important, and belongs in the paper. To me that is the most relevant related work to this one. I maintain my rating.

Reviewer 7Fho2023-08-16

Thanks for the response. I maintain my rating.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC