Canonical normalizing flows for manifold learning

Manifold learning flows are a class of generative modelling techniques that assume a low-dimensional manifold description of the data. The embedding of such a manifold into the high-dimensional space of the data is achieved via learnable invertible transformations. Therefore, once the manifold is properly aligned via a reconstruction loss, the probability density is tractable on the manifold and maximum likelihood can be used to optimize the network parameters. Naturally, the lower-dimensional representation of the data requires an injective-mapping. Recent approaches were able to enforce that the density aligns with the modelled manifold, while efficiently calculating the density volume-change term when embedding to the higher-dimensional space. However, unless the injective-mapping is analytically predefined, the learned manifold is not necessarily an efficient representation of the data. Namely, the latent dimensions of such models frequently learn an entangled intrinsic basis, with degenerate information being stored in each dimension. Alternatively, if a locally orthogonal and/or sparse basis is to be learned, here coined canonical intrinsic basis, it can serve in learning a more compact latent space representation. Toward this end, we propose a canonical manifold learning flow method, where a novel optimization objective enforces the transformation matrix to have few prominent and non-degenerate basis functions. We demonstrate that by minimizing the off-diagonal manifold metric elements $\ell_1$-norm, we can achieve such a basis, which is simultaneously sparse and/or orthogonal. Canonical manifold flow yields a more efficient use of the latent space, automatically generating fewer prominent and distinct dimensions to represent data, and a better approximation of target distributions than other manifold flow methods in most experiments we conducted, resulting in lower FID scores.

Paper

Similar papers

Peer review

Reviewer iKjw6/10 · confidence 3/52023-06-14

Summary

The authors address the interesting question of how to disentangle the relevant manifold directions properly in the latent space of manifold learning normalizing flows (MLF). After a comprehensive theoretical background, the main motivation is presented in form of a toy example of a simple noisy line embedded in $\mathbb{R}^{2}$. When varying the latent dimension $z_{i}$ of a RNF and setting the remaining latent variables to $0$, the corresponding contour lines are partially aligned. Doing the same with the proposed method, CMF, leads to orthogonal contour lines and thus to a meaningful latent representation where one latent variable encodes the line and the other the (noisy) off-manifold direction. This intuition leads straightforwardly to enforcing the orthogonality of the flows Jacobian matrix. To do so, the authors propose a new cost function for learning manifold structures using an NF by adding a penalty term which ensures that the off-diagonal entries of the Gram matrix (Jacobian-transpose-Jacobian) are $0$ at the one hand, and are sparse at the other due to the l1-norm. The effecitiveness of the method is tested in experiments on toy-data, tabular data and 32x32x3 image data.

Strengths

The paper is overall well written. The theoretical background is exhaustive and well-explained. The example in 4.1 serves as a good illustrative motivation. The experiments are comprehensive.

Weaknesses

The contribution is somewhat marginal. Though citing [1], they don't mention that the corresponding authors propose a very similar penalty term. To my understanding, the only differences of the penalty term in [1] are a) that in [1] the diagonal entries are penalized to be $1$ whereas in this work the diagonal entries are unconstrained. b) [1] used the L2-norm whereas in this paper the L1-norm is used. Given the great similarity, the penalty term suggested by [1] should be included in the comparison. What do we gain by not constraining the flow to be an isometry? In addition, the authors use the M-flow for benchmarking although a better method for overcoming the limitations of M-flows (namely that the log-determinant term is not considered while training) was already proposed in [2]. At least citing and ideally comparing with [2] should be part of the paper. The same is true for comparing with [3]. It would be nice to see how the contour lines differ from the method introduced in [3]. Furthermore, I am not sure to understand the reasoning behind Definition 4.1. To my understanding, every differentiable manifold is a canonical manifold as one is free to choose the basis of the tangent space to be orthogonal. Thus, such a basis always exists which is the only requirement to be a canonical manifold. Definition 4.1. is simply an existence statement and not useful as such. I think what the authors wanted to define is a manifold equipped with a chart s.t. the chart induces an orthogonal basis of the tangent space. However, this is very much the definition of the principal component flow [3]. Please enlighten me in case I totally misunderstood something. I am also confused about the comparison with rectangular NF (RNF) in Figure 2. By definition for RNF we have that $d<D$. However, in Figure 2 the author uses $d=D$. This is a contradiction. If $d=D$, a standard NF can be used rather than an RNF. In addition, one claimed strength of the proposed CMF is the sparisity induced by the l1 norm. Thus, the relevant dimensions should be found automatically which elevates, in theory, the necessity to estimate $d$ a priori. Then, why not always set $d=D$ and then apply the proposed penalty term? The number of prominent latent dimensions should correspond to the true intrinsic dimensionality of the data. Finally, the notation for the manifold, data manifold, and learned manifold is confusing. In line 74 or line 90, $\mathcal{M}$ is a lower dimensional manifold. In line 83, $\mathcal{M}_{\theta}$ is introduced without explaination. In line 212, $\mathcal{M}$ is referred to as the learned manifold and in the same line, the data manifold is introduced without further explanation. In addition, a new notation for a canonical manifold is suggested. I find the various versions for a manifold confusing and don't see the added benefit of it (especially given that Definition 4.1. is not useful in my opinion). [1] Eike Cramer, Felix Rauh, Alexander Mitsos, Raúl Tempone, and Manuel Dahmen. Nonlinear 350 isometric manifold learning for injective normalizing flows. arXiv preprint arXiv:2203.03934, 351 2022. [2] C. Horvat and J.-P. Pfister. Denoising normalizing flow. In Advances in Neural Information Processing Systems, volume 34, 2021. [3] Edmond Cunningham, Adam D Cobb, and Susmit Jha. Principal component flows. In Interna410 tional Conference on Machine Learning, pages 4492–4519. PMLR, 2022.

Questions

1. Does Definition 4.1 really makes sense (see above) 2. Is the comparison with RNF in Figure 2 fair (see above)? By definition of RNF, it must be $d<D$. Please elaborate. 3. Why not always set $d=D$ and then apply the proposed penalty term (see above)? 4. Figure 1. Why is the density varying greatly when only varying one dimension? Shouldn't be uniform if the density is properly learned? 5. What is the quality in terms of density estimation? For instance, what is the KS statistics compared to the true density, or, at least, does it sum to 1? 6. Why not benchmark with PCA flow? It would be interesting to see if this method can overcome some of its limitations (see above). 7. Moebius band with RNF seems to be poorly learned. Why? 8. Why prominent latent dimensions are the ones with the greatest $G_{ii}$? Is this a mathematical statement/proposition or intuition? 9. line 305: in which case would an orthogonal basis not be an optimal representation? 10. What happens if we use the L2 rather than L1 norm? 11. How are prominent latent dimensions defined? Do they correspond to the true intrinsic dimensionality? Is there a clear cut-off in the magnitudes of $|G_{kk}|$ in e.g. MNIST? If not, a somewhat arbitrary threshold needs to be defined, isn't it? *Minor*: + line 70-72: the statement is only true for large $n$ + line 83: What is $\mathcal{M}_{\theta}$? + line 83: "to encorage $x\in \mathcal{M}_{\theta}$" do you mean $g_{\phi}(z)$ rather than $x$? + line 114: typo: "due to lack..." + line 145: noise is not perpendicular to line + line 217: mention reason (maybe in a footnote) + line 225: $\gamma$ not crucial but had an impact on training time? What is crucial for you? + line 244: bad wording "the expected the canonical manifold learning" + line 297: Inconsistent acronym: M-flow vs. Mflow + line 299: inferior lok-likelihood calculations of Mlow...where shown? + line 309/310: wrong statement: there are diffusion models with lower dimensional latent space. However, by default indeed the latent space has the same dimensionality

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

Major limitations are addressed in Section 6. However, some questions are unresolved such as: how to find/define the prominent latent dimensions without an intrinsic dimensionality estimator?

Reviewer 8w9Q5/10 · confidence 3/52023-07-05

Summary

The paper studies the current manifold learning methods. It compares canonical manifold learning flow (CMF) with other manifold learning methods and demonstrates that CMF can learn the orthogonal features existing in data. From synthetic data on Moebius, the paper shows the benefits of using CMF. Also, the paper shows the generated images from the CMF learned latent space is of higher quality than other methods' learned latent spaces. Lastly, the paper acknowledges some limitations of CMF, for example, high computation cost of the full Jacobian tranpose Jacobian.

Strengths

1. The paper is very clear on the advantages of CMF against other manifold learning methods. 2. The paper did various types of experiments to showcase the CMF's performance among the methods.

Weaknesses

1. Although being mentioned in the limitations, the paper does not provide results about the computation costs of CMF. There is no comparison of the computation time of CMF against other methods, also how the dimension affects the computation cost. It is unclear if CMF is scalable to higher-dimensional datasets. 2. The paper does not compare CMF with a lot of methods, including PCA and ICA or purely deep learning methods. For example, an autoencoder can also extract feature representations from data. How does CMF compare to these approaches?

Questions

1. What is the complexity of the RNF and CMF methods? 2. For table 2, are there explanations on why M-flow works the best and CMF is the worst on GAS? 3. Figure 3 is hard to interpret. What are the advantages of CMF in this plot? 4. What are the generative models used in the image experiments?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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 have discussed the limitations clearly in the paper. They acknowledge the high computation costs and the limitation that the method only applies when the manifold is homogeneous to Rd space.

Reviewer 7TiS7/10 · confidence 4/52023-07-07

Summary

The authors introduce a new method for regularising manifold learning flows. Essentially, it attempts to reduce the entanglement between the dimensions of the learned manifold by encouraging non-diagonal elements of the metric tensor to be small. Leveraging the already necessary computation of $J^\top J$, where $J$ is the Jacobian of the flow transformation, this is done efficiently by minimising the $\ell-1$ norm of the non-diagonal entries. Experiments on synthetic and real data show how this is effectively achieved, especially when compared to similar models that do not employ the regularisation scheme. Beyond achieving the desired effect, the proposed method achieves lower FID-like scores on real tabular data.

Strengths

The paper is very well written and the presented ideas are easy to follow. Experimentally, the authors confirm both in intuitive, synthetic examples and real data. The proposed regularisation scheme is also simple and computationally efficient, which is also a desirable feature. Overall, I consider it a strong contribution and see high potential of being widely adopted as a mechanism for regularising manifold learning flows. I also appreciate the limitations raised and discussed by the authors, indicating maturity in their analysis and raising important considerations related to the use of their work.

Weaknesses

Although I see the potential mentioned above, I do consider it could be viewed more widely as a weak point, given the relatively niche application. Although it does not affect the merit of the work, the principle might be too specific to manifold learning flows. I would consider more tabular data experiments are needed, with additional data sets, given the mixed results attained. More concretely, some analysis of what specifically differentiates the data sets enough to cause the difference in performance.

Questions

- Tied to the aforementioned weakness, what do the authors posit is the cause for the performance gap in GAS? - Related to the previous question: are the metric tensors found in tabular data also behaving similarly to the ones shown in Figure 3? - Is it clear during training when a pathological scenario is reached? How does it behave w.r.t. increasing/decreasing the value of $d$? - Are the learned manifold representations also useful as a feature extraction procedure for downstream tasks? Could this be also a scenario for evaluating the quality of the learned models? (As opposed to only raw performance in scores) Minor comments: - line 35: Caterini et al [7] has => Caterini et al. [7] have - line 44: Please avoid unnecessary adjectives such as "tedious" - line 62: why is the set starting with $x_0$?

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

I was pleasantly surprised with the discussion on limitations in the paper. I believe the authors addressed important issues and even highlighted which other approaches could have better raw performance.

Reviewer tk9x5/10 · confidence 4/52023-07-08

Summary

This paper studies the problem of learning a latent representation for data supported on a low-dimensional manifold. It proposes to promote orthogonality of the tangent vectors arising from a learned chart, on top of existing rectangular flow loss. Experiments are provided to demonstrate the effectiveness of the algorithm.

Strengths

Promoting orthogonality of the tangent vectors from a learned chart is interesting.

Weaknesses

## Presentation: 1. The paper seems insufficiently prepared and proofread. In particular, there are numerous evident typos even in the first paragraph of the introduction. Consequently, they weaken the credibility of the paper. 1. Lines 29, 33 the "D" in "R^D" uses a mathbb, whereas in lines 31, 32 the "d" or "D" does not. 1. Line 29 “fulfil” -> “fulfill” 1. In equation 1, the q_phi(x) should have been q_phi^{-1} (x) 1. Section 3 talks about general manifold learning, and is almost detached from the technical sections 2 and 4. 2. Line 43: A line is one-dimensional. What is a two-dimension line? 3. If one compares equations (5) and (6), it appears that we assume the learned map q_phi is a chart of manifold M. If that is the case, this should be made explicit. 4. From (8), it is unclear that where the introduced G term are evaluated. I suppose it is on points x_i’s?

Questions

## Significance Looks like the inverse of g_phi is in objective (6), so using line 75 it boils down to the inverses of g_eta and f_theta. How easy is it to invert g_eta and f_theta? The root of this question is, if one takes f_theta to be a square matrix, then inverting the square matrix is needed, leading to high cost if the dimension of the square matrix is large. If f_theta is a square matrix followed by say ReLU, then ReLU is not invertible.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

Yes

Reviewer iKjw2023-08-10

Answer

I appreciate the detailed answers of the authors. I am mainly satisfied with them, and they help me to further appreciate the conceptual contribution of the paper - in particular the usage of the L1 error and the corresponding advantages. I will adjust my score accordingly. Regarding my questions: *Q2*: I'm afraid I disagree. The authors of "Rectangular Normalizing Flows" define them only for $d<D$. If $d=D$, an RNF should result in a standard NF. Please correct me if I am wrong. However, using a standard NF on a hollow sphere or Moebius band, Figure 2, should lead to very different results (the density must, in fact, degenerate unless you add some noise, which you did not mention). *Q3*: I don't quite see why this is computationally more expensive. The only computational impact will have the computation of the metric tensor. However, this only requires a backward pass which is efficient to compute. Please elaborate/correct me. *Q8:* This intuition should be part of the discussion in my opinion. *Q11: " For the analysis plots, we order the dimensions according to their weights and strictly choose a pre-defined number of them where applicable. "* So you need, in the end, an ID estimation. If so, this should be also clearly discussed in the limitations.

Authorsrebuttal2023-08-16

We greatly appreciate the valuable insights offered in the review and the subsequent responses. Q2: Indeed, it is right that it would be equivalent to a standard NF. Moreover, as aptly noted, effectively learning these manifolds remains a challenge, even with the full-dimensional latent space. However, the RNF is based on CIF [5] which is an NF that already tries to solve the learning pathologies of complicated manifolds, "CIFs are not subject to the same topological limitations as normalising flows". Furthermore, in that regard, we estimate there is no practical limitation in setting $d=D$, as also seen empirically. Considering the above, the CMF method is showcased to outperform these previous methods, which also aligns with the theoretical intuition. We extend our appreciation to the reviewer for the thought-provoking discussion, and we'll certainly incorporate a comprehensive comment in the revised manuscript to reflect this exchange. Q3: The bottleneck for both methods is the calculation of the Jacobian-transpose-Jacobian (JTJ), as explained in Appendix 4 and RNF. The complexity of the approximation used is $\mathcal{O}(id^2) < \mathcal{O}(d^3) $ if $i << d$, $i$ refers to iterations of the iterative conjugate gradients method. M-Flow seeks to circumvent this computation entirely, while RNF aims for an approximation. Notably, as accurately noted, the JTJ yields the metric tensor directly, incurring no significant additional computational expense. When $d=D$, as previously discussed, this equates to a standard NF. In this regard, manifold flow methods were prompted to enhance computational efficiency by setting $d<<D$. Indeed, in alignment with the reviewer's observation, and as addressed here, the efficient approximate methods indicate that CMF could indeed be implemented for a full-dimensional flow. Further exploration of this potential could be undertaken in future studies. We will ensure to provide a diligent commentary in the revised manuscript. Q8 and Q11: We wish to clarify that this is solely necessary for Figure 4 and potentially for any future use that entails isolating prominent components. In this context, we sincerely appreciate the suggestion, which undoubtedly represents a substantial improvement over our current approach of analysis. Rest assured, we will incorporate these discussions into the updated version of the manuscript. [5] Relaxing Bijectivity Constraints with Continuously Indexed Normalising Flows Rob Cornish, Anthony L. Caterini, George Deligiannidis, Arnaud Doucet

Reviewer 8w9Q2023-08-19

Thank you for addressing my questions and concerns. I am more leaning to accept now and raised the score to 5.

Authorsrebuttal2023-08-20

We would like to express our gratitude to the reviewer for their evaluation.

Reviewer tk9x2023-08-20

Thank you for the response. 1.2: I apologize for being ignorant of the spelling - British English, as any other English, is perfectly fine. 1.1-1.3: My intention was not to be nit-picking. Everyone makes typos and mistakes, and that is what proofreading is for. Having obvious typos in early parts of the paper is a sign of a lack of proofreading, thus it lowers the credibility of other parts that are harder to check. Nonetheless, after reading the response and other reviews, I raised my rating.

Authorsrebuttal2023-08-20

We extend our appreciation to the reviewer for offering clarifications and investing the time to re-examine our work.

Reviewer 7TiS2023-08-20

I thank the authors for addressing all issues I raised. After reading the other reviews and rebuttals, I am increasing my score to 7 (Accept), as I no longer think there are any outstanding issues with the submission.

Authorsrebuttal2023-08-20

We would like to thank the reviewer for their valuable comments.

Authorsrebuttal2023-08-20

We sincerely appreciate the time and effort that all the reviewers have dedicated to assessing our work, and we express our gratitude to them.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC