On the Limitations of Fractal Dimension as a Measure of Generalization

Bounding and predicting the generalization gap of overparameterized neural networks remains a central open problem in theoretical machine learning. There is a recent and growing body of literature that proposes the framework of fractals to model optimization trajectories of neural networks, motivating generalization bounds and measures based on the fractal dimension of the trajectory. Notably, the persistent homology dimension has been proposed to correlate with the generalization gap. This paper performs an empirical evaluation of these persistent homology-based generalization measures, with an in-depth statistical analysis. Our study reveals confounding effects in the observed correlation between generalization and topological measures due to the variation of hyperparameters. We also observe that fractal dimension fails to predict generalization of models trained from poor initializations. We lastly reveal the intriguing manifestation of model-wise double descent in these topological generalization measures. Our work forms a basis for a deeper investigation of the causal relationships between fractal geometry, topological data analysis, and neural network optimization.

Paper

References (48)

Scroll for more · 36 remaining

Similar papers

Peer review

Reviewer d6bB7/10 · confidence 4/52024-07-04

Summary

This study revisits recent claims about persistent homology dimension (in the limit a fractal dimension) and the generalization gap of supervised learning systems. The essential theoretical background is summarized, and then three main experiments are conducted, concerning 1) (partial) correlation of the concepts, 2) influence of poor or adversarial initialization, 3) double-descent phenomenon. Overall, the results indicate that PH only weakly correlates with the generalization gap, and that hyper-parameters may be a confounder since there appears to be a stronger correlation between PH and hyper-parameters.

Strengths

S1 - The considered topic is relevant to ML theory and the experiments were conducted with high quality. All formal statements are correct. S2 - The reported results, and the discussion thereof, are insightful and useful to the wider community. S3 - The discussion of limitations is wide and transparent. Section 5 is a role-model of how a paper's limitations should be presented.

Weaknesses

W1- Readability. The work is not self-contained and heavily depends on information from other papers. This makes it tedious to study this work and requires the reader to frequently switch back and forth between multiple papers. Although several passages are well-written, the paper is difficult to follow overall. This may be, in part, because the paper is not self-contained, but the frequent cross-references within this paper suggest that the structure may not be optimal. W2 - The hypothesis testing setup is not properly described. It is not true that the hyper-parameters must have a significant influence if we reject the null-hypothesis that there is a correlation between PH and the generalization gap. Lines 228-230 need to be re-written (and extended) so that the hypothesis test setup is clear and rigorous to avoid misinterpretations. W3 - The title and abstract are misleading need to be updated to reflect on this work's experimental focus. I was expecting a theoretical paper. For example, the phrase "This work performs an extended evaluation of these topological generalization measures" could also be understood as a formal evaluation.

Questions

Q1 - The considered hypothesis test of PH <-> generalization vs. PH <-> hyper-parameters actually feels like two hypothesis tests, where each considers the absence of the respective correlation separately. Please explicitly list the considered hypotheses and consider if such as split makes sense. Q2 - Please consider the improvement suggestions below. **Improvement suggestions** line 65: This is a slight abuse of notation. I would prefer if the composed loss function were explicitly declared as such, i.e. $\ell(w,z)=\mathcal{L}(h_w(x),y)$. Otherwise, the paper is not self-contained. line 80: It is stated that further details on persistent homology are given in the supplementary material, but actually it is in the appendix. This is good for a camera-ready version, but in the review phase it is a bit confusing. line 82: **x** is not a metric space, it is a set. A metric space is a tuple of a set and a metric. Please state what the metric is – presumably $\rho$?. typos in lines 89 & 106 Eq. (2): $w'$ is not defined here. Also, maybe use a different notation, since $\rho$ is later used for Spearman correlation. lines 144-154: I would appreciate if at least some of the selections were explained, e.g., why cross-entropy loss? line 172: This needs to be explained further. I understand that long trainings needs to stop at some point, but an interruption may influence the results if done poorly. line 233: "causally correlated" -> "correlated". lines 244-245 can be omitted since they have the same content as 242-243. In Table 2, it might be helpful to use colors or bold fonts to visually distinguish between significant and non-significant findings, i.e., p <= 0.05 or not. Definition of mutual information: Maybe use a different notation for probability like $\text{Pr}(\cdot)$ , since $p$ is already the $p$-value. line 276: Grammar mistake, "PH dimensions ... does" Section 4.3. This discussion is feels too short. Maybe condense lines 213-243 to make more space. line 290: Typo? Please rephrase. line 291: typo "preventted" -> "prevented"

Rating

7

Confidence

4

Soundness

4

Presentation

2

Contribution

3

Limitations

The limitations are properly addressed

Reviewer garn7/10 · confidence 4/52024-07-10

Summary

This work empirically studies the connection between the generalization gap and the fractal dimension of a neural network's optimization trajectory. A recent line of research has established such connections in a theoretical manner through upper generalization bounds. Specifically, three types of experiments are performed 1. An analysis of the correlation between fractal dimensions and empirical generalization gap in a typical training scenario. Here, the experiments from prior work [1] are replicated and analysed to greater detail. 2. An analysis of such correlations for models that do not generalize properly, obtained from training starting at an adversarial initialization. 3. A double descent behavior of fractal dimensions is observed when training models of varying width. In the first two experiments, only few statistical significance of the correlation between fractal dimensions and generalization is observed, i.e., less than between generalization and the L2 norm of parameters.

Strengths

- The empirical evidence presented in prior work is closely examined and questioned. While this might not be as appreciated as the derivation of novel theoretical results or of previously unobserved empirical findings, it is definitely necessary in order to establish a solid foundation future work can build upon. - There is substantial effort in identifying and transparently communicating the limitations of this work, much more than in all other submissions I have reviewed for NeurIPS 2024. - Multiple different correlation scores are computed, and the findings are tested for statistical significance. - The background section provides a good overview over the research on generalization bounds in terms of fractal dimensions.

Weaknesses

- In section 4.1, the experiments from [1] are carefully replicated. However, the observations do not match the ones from [1]. For example, for FCN-5 on CHD, a *negative* correlation score of $\rho = -0.7622$ is reported, whereas [1] reports a *positive* correlation of $\rho = 0.87$, which looks plausible from [1, Figure 2]. This inconsistency is surprising, but it is neither discussed nor mentioned. If the setting is indeed equal, then the correlation coefficients should be much more similar. I do trust both works in that they correctly report their results, but this contradiction needs to be resolved in some form. - This submission presents the empirical data, but is very cautious in interpreting it. On the one hand, it is good to let the data speak for itself, but on the other hand, the relevant part is how the observations relate to the larger context. For example, at the end of section 4.1 (Correlation Analysis), it is stated that "Using this method, we conclude that the correlation coefficients between generalization and the norm or the ratio of learning rate over batch size are indeed significantly higher than the correlation coefficients with PH dimensions [l.209]" What are the conclusions from this? Similar questions apply to the other sections. - It seems that the data is collected from only 10 random seeds (in each setting). For a purely empirical work that seems a bit low, especially if strong claims are deduced from it, e.g. "We demonstrate that fractal dimension fails to predict generalization of models trained from poor initializations" - Not a weakness of the submission (rather on my end), but worth mentioning: I do not have sufficient knowledge on testing for correlations or independence. Thus, I will not comment on the used statistical pipeline (setup and choice of correlation coefficients, null hypotheses, tests, etc) - typos: l.290 we were only work ... l.291 preventted ### References 1. Dupuis et al., Generalization Bounds using Data-Dependent Fractal Dimensions, ICML 2023. 2. Birdal et al., Intrinsic Dimension, Persistent Homology and Generalization in Neural Networks, NeurIPS 2021 3. Hofer et al., Connectivity-Optimized Representation Learning via Persistent Homology, ICML 2019 4. Chen et al., A Topological Regularizer for Classifiers via Persistent Homology, AISTATS 2019 5. Carrière et al., PersLay: a neural network layer for persistence diagrams and new graph topological signature, AISTATS 2020

Questions

- Your empirical observations contradict the theoretical predictions deduced from the generalization bounds in [1,2,..]. What do you think is the cause? Is it, because there is an error in the theory, are the assumptions in the theory unrealistic (not fulfilled in practice), is the fractal dimension not the decisive term in the generalization bounds, etc? - A limitation that is not stated in your discussion section, but most likely plays a role, is the uncertainty of the fractal dimensionality estimates. I wonder, how much difference there is, if you continue SGD for another 5k iterations and re-estimate the dimension from them. Are the estimates approximately equal? Is their order the same? Is there any rank correlation between estimates from the first 5k and the second 5k iterations at all? Note that even if the estimates are equal, it does not mean that the estimates are accurate, as they might be biased. - I guess, this was not your intention, but the remark "While efficient PH computation is a field in itself, integrating it with network training to make fractal dimension computations more feasible is an interesting direction, especially if we are able to derive a regularization term. [l312]" suggests that integrating persistent homology into neural network training hasn't been done before. This is not the case, and there are *many* works that do so, [3-5] to name a few. In fact, such a regularization term is even used by Birdal et al. [2] to encourage low ph dimensions. - I did not understand your test of conditional independence, cf l.257. Would you explain it to me?

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Yes, the authors adequately addressed the limitations!

Reviewer whkD6/10 · confidence 1/52024-07-12

Summary

This paper performs experiments to evaluate the correlation between the generalization gap and the persistent homology (PH) dimension, a measure of fractal dimension deriving from topological data analysis. The authors identify that the $\ell^2$ norm of the final parameter iterate correlates more strongly with the generalization gap than the fractal dimension. Moreover, the authors also demonstrate that the fractal dimension fails to predict the generalization of models trained from poor initializations.

Strengths

The authors provide extensive experiments and the conclusions are well supported. Moreover, this paper is well-organized and clearly written.

Weaknesses

1. The definition of the PH dimension is not easy to understand. 2. Some results need more explanation (see Questions).

Questions

1. In Table 1, the correlation coefficients could have different signs in the same line. Why would this happen? Does this come from different models and data? 2. In Table 2, how to explain the influence of the batch size on the p-value? 3. I cannot figure out how Figure 1 manifests double descent.

Rating

6

Confidence

1

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors discuss limitations in Section 5.

Reviewer rTx25/10 · confidence 4/52024-07-13

Summary

The paper investigates the effectiveness of using fractal dimensions, particularly the Hausdorff dimension and persistent homology dimension, as measures for predicting the generalization gap in neural networks. It demonstrates that fractal dimensions fail to predict generalization for models trained from poor initializations and shows that the $\ell_2$ norm of the final parameter iterate correlates more strongly with the generalization gap than fractal dimensions. Experimental results show that persistent homology dimension fails to predict the generalization of models with adversarial initialization, which generalize poorly.

Strengths

The paper critically evaluates the use of fractal dimensions, demonstrating that they fail to predict generalization for models trained from poor initializations. The authors conduct a rigorous statistical analysis, including partial correlation measures conditioned on network hyperparameters, revealing that fractal dimensions are not statistically significant predictors of the generalization gap when hyperparameters are considered. The study uncovers an interesting manifestation of model-wise double descent in persistent homology-based generalization measures, suggesting potential for further research.

Weaknesses

Lack of theorical results.

Questions

No questions.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Computational limitations and lack of Identifiable Patterns in the Correlation Failure

Reviewer d6bB2024-08-08

Acknowledgement of author rebuttal

I thank the authors for their rebuttal. Overall, I think they did a good job addressing the points I raised. I am also happy with the overall response, and how the authors handled the points raised by the other reviewers. The proposed changes for W2 seem sufficient. I agree with the authors that W1 is something that probably can't be easily fixed. Regarding the abstract, it might be sufficient to rewrite the fourth sentence as *This work performs an ~extended~ **empirical** evaluation of these topological generalization measures*. I will keep my positive rating.

Authorsrebuttal2024-08-09

Reply to Acknowledgement of author rebuttal

We are very grateful to the reviewer for their reply and are happy to see that they found our answers and the changes that we proposed satisfactory. We will implement these alongside the modification in the abstract suggested by the reviewer in a revised version of the paper. We are very pleased that the reviewer remains positive and maintains their acceptance rating. Thank you again for all you support and the time devoted to helping us improve our work!

Reviewer garn2024-08-09

I thank the authors for their rebuttal. The values in the updated Table 1 do not contradict the empirical results of Dupuis et al. [2023]: anymore. The previous **discrepancy to Dupuis et al. [2023]** was a major factor in my cautious rating and is now resolved. However, now the rank correlations to norms on CHD have a flipped sign. Would you comment on that? My second concern (limited **conclusions from the empirical findings**) has also been addressed satisfactorily. The sketched interpretations of results should certainly be added to the main part in an extended (but still nuanced) manner. As should be the two main conjectures of why the empirical findings seem to contradict the generalization bounds. Regarding the added **ph-dimension estimates**, my question has only partially been answered, in that rank correlations between ph-dimensions have been presented, which addresses the relative order of estimates. However, I was also asking whether the estimates are approximately equal at different numbers of SGD iterations, which concerns the actual values of the estimates. It also seems (to me) that comparing always to 5k iterations (i.e. 5k vs 10k and 5k vs 15k) instead of (15k vs 10k) would be more telling, as differences from 5k to 10k might propagate. Moreover, as the measured rank correlations are far from 1, I would be very interested in a variant of table 1, where 15k iterates are used. FCN-5 & CHD and FCN-5 & MNIST would be sufficient. All **minor concerns** have been sufficiently addressed. Regarding points raised by the **other reviewers**; I agree with Reviewer d6bB that title and abstract are misleading should to be updated to reflect on the work's experimental focus.

Authorsrebuttal2024-08-12

Reply to Official Comment by Reviewer garn (1)

We are very grateful to the reviewer for their reply and are happy to see that they found our answers to most of their concerns satisfactory. We now address the two pending points that they mention in their message. - **Flip of sign of the correlation with norm**: - Although there are challenges and limitations to the parameter norm as a measure of generalization [1], it is still a widely used and accepted measure of model complexity. We included this in our experiments as a simple, universally understood reference point for a high-complexity generalization measure. Future work could compare further complexity measures against these topological measures. - Classical ML theory would suggest a positive correlation between the norm and generalization gap. On regression we observe this positive correlation, but on the classification we observe the inverse. - It would appear that the norm correlates with the LB/ratio which itself is (weakly) positively correlated with gap on regression, and (strongly) negatively correlated with gap on classification. A contributing factor to the weak correlation between LB/ratio and gap on regression could be the limited range of learning rate values considered ($[10^{-3}, 10^{-2}]$). - Finally, we note that the norm correctly correlates with generalization gap in all experiments using adversarial initialization and in the double descent experiment (not with evaluation accuracy as observed in the Euclidean dimension). We can present similar plots like the ones included in the general answer in the final version of the paper. - We can also include relevant discussion when we present Table 1, we suggest the addition of the following paragraph: *Classical machine learning theory suggests a positive correlation between the norm and the generalization gap, which we observed in our regression experiments. However, in classification tasks, we found an inverse relationship, likely influenced by the learning rate (LR) and batch size (BS) ratio (LR/BS ratio), which may act as a confounding factor similar to effects seen with PH dimensions. Although the norm faces challenges and limitations as a generalization measure for deep networks [1], we included it in our comparisons as a simple, universally understood, and widely used and accepted reference point to compare with high-complexity generalization measures like the PH dimensions. We note that the norm correctly correlates with generalization gap in experiments with adversarial initialization and double descent, contrasting with PH dimensions (see Figures).* - **PH dimension estimates**: We apologize if the table presented was incomplete. We now present an updated version of the table, where: - We include comparisons between 5k and 15k. - We include a new column per case with the % difference, that is: $[(\text{PH dim at value 1}−\text{PH dim at value 2})/\text{PH dim at value 1}]∗100$. We have computed this difference per seed in each model, and present here the average and standard deviation as error bar. Looking at these new columns, one can note that the relative differences between values are small, and do not have a consistent sign, indicating a (small) fluctuating pattern that is likely due to randomness. | Model & Data | Measures | 5k vs 10k | | | 10k vs 15k | | | 5k vs 15k | | | |:------------------:|:----------:|:----------------:|:------:|:------:|:----------------:|:------:|:------:|:----------------:|:------:|:------:| | | | $\%$ difference | $\rho$ | $\tau$ | $\%$ difference | $\rho$ | $\tau$ | $\%$ difference | $\rho$ | $\tau$ | | FCN-5 & CHD | Euclidean | $0.15 \pm 1.74$ | 0.63 | 0.43 | $-0.41 \pm 1.05$ | 0.88 | 0.72 | $-0.25 \pm 1.72$ | 0.68 | 0.53 | | | Loss based | $-3.83 \pm 6.82$ | 0.75 | 0.56 | $-3.48 \pm 6.62$ | 0.62 | 0.49 | $-7.24 \pm 7.44$ | 0.72 | 0.55 | | FCN-5 & MNIST | Euclidean | $0.11 \pm 0.66$ | 0.65 | 0.46 | $0.18 \pm 0.60$ | 0.46 | 0.34 | $-0.07 \pm 0.74$ | 0.45 | 0.33 | | | Loss based | $1.53 \pm 1.35$ | 0.62 | 0.45 | $1.01 \pm 1.21$ | 0.53 | 0.38 | $2.62\pm 1.22$ | 0.65 | 0.48 | | AlexNet + CIFAR-10 | Euclidean | $-0.75 \pm 0.63$ | 0.86 | 0.68 | $0.02 \pm 0.40$ | 0.87 | 0.69 | $-0.72\pm 0.64$ | 0.92 | 0.78 | | | Loss based | $-1.05\pm 2.19$ | 0.61 | 0.46 | $-1.41 \pm 1.67$ | 0.23 | 0.16 | $-2.47 \pm 2.77$ | 0.14 | 0.11 | # References [1] Jiang, Y., Neyshabur, B., Mobahi, H., Krishnan, D., & Bengio, S. (2019). Fantastic generalization measures and where to find them. arXiv preprint arXiv:1912.02178.

Authorsrebuttal2024-08-12

Reply to Official Comment by Reviewer garn (2)

- **Table 1 using PH dimensions with 15k iterations**: In the table above we are computing PH dimension with 5k iterates, taking the weights of the 0-5k, the 5k-10k and the 10k-15k iterations after reaching 100% train accuracy. In response to the reviewer’s request for an updated version of Table 1 with PH dimension computed using all 15k iterations after reaching 100% training accuracy, we initiated these additional experiments. Unfortunately, we are limited by computational constraints during this discussion period and it will not be feasible to train and compute the dimension for a sufficient number of models by the end of the discussion period (with each model taking more than 1h to run, and a total of 720 models required for these new results). We sincerely apologize for this.

Reviewer garn2024-08-13

Thank you again for your response. Regarding the **correlation with weight norms**. I was wondering, why the updated evaluation method for table 1 had a very strong effect on the correlation with weight norms (a positive sign in the correlation coefficient, compared to the negative sign in the original table 1 in the submission), but only on one dataset (CHD). You make a good point however, that in theory, we would expect a *positive correlation* between norms and generalization gap, which is not seen in table 1, except on the CHD regression task. In this light, statements comparing weight norms and fractal dimensions as a measure of generalization need to be toned down, including (but not limited to) this sentence in the abstract: *"We further identify that the ℓ2 norm of the final parameter iterate, one of the simplest complexity measures in learning theory, correlates more strongly with the generalization gap than these notions of fractal dimension"* Regarding **fractal dimension estimates**. Thank you for the new table, please include this table in the appendix of the paper, as it shows (except for AlexNet on CIFAR-10) that the fluctuation in rank correlations between fractal dimension estimates only mildly depends on the number of SGD steps after which the dimension is estimated. To strengthen this points I encourage you to also add a column 5k vs 5k (which I expect to have similar values). Regarding **table 1 at 15k steps**. I was under the impression that you already had the necessary models available from the comparison of fractal dimension estimates. As this does not seem to be the case, and as the compute requirements appear to be excessive, I will leave it at that. --- Overall, the authors sufficiently addressed all of my (ant other reviewers') points. I will raise my score from 5 to **7: Accept**.

Authorsrebuttal2024-08-13

Reply to Reviewer garn

We are very grateful to the reviewer for their reply and are really pleased to see that they feel that we have sufficiently addressed the points raised by them and other reviewers. We would like to reiterate our appreciation for their careful review, which has been instrumental in improving the quality of our work, making the results and interpretations more nuanced and precise. Thank you for your continued support and for engaging with us in this meaningful discussion!

Authorsrebuttal2024-08-13

Reply to Official Comment by Reviewer whkD

We sincerely thank the reviewer for their response and are pleased to know that our answers have addressed their questions. We are also very grateful for the time and effort they dedicated to providing valuable feedback, which has contributed to improving our work.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC