Flow Factorized Representation Learning

A prominent goal of representation learning research is to achieve representations which are factorized in a useful manner with respect to the ground truth factors of variation. The fields of disentangled and equivariant representation learning have approached this ideal from a range of complimentary perspectives; however, to date, most approaches have proven to either be ill-specified or insufficiently flexible to effectively separate all realistic factors of interest in a learned latent space. In this work, we propose an alternative viewpoint on such structured representation learning which we call Flow Factorized Representation Learning, and demonstrate it to learn both more efficient and more usefully structured representations than existing frameworks. Specifically, we introduce a generative model which specifies a distinct set of latent probability paths that define different input transformations. Each latent flow is generated by the gradient field of a learned potential following dynamic optimal transport. Our novel setup brings new understandings to both \textit{disentanglement} and \textit{equivariance}. We show that our model achieves higher likelihoods on standard representation learning benchmarks while simultaneously being closer to approximately equivariant models. Furthermore, we demonstrate that the transformations learned by our model are flexibly composable and can also extrapolate to new data, implying a degree of robustness and generalizability approaching the ultimate goal of usefully factorized representation learning.

Paper

Similar papers

Peer review

Reviewer wDok6/10 · confidence 3/52023-07-05

Summary

The paper introduces a new generative model called Flow Factorized Representation Learning that learns a distribution over sequences of observed variables using which one can learn both disentangled and equivariant representations. The model is based on the idea of representing transformations in the input space as latent flow paths generated by the gradient fields of learned potentials. In this way, disentanglement can be read from the set of tangent directions, while equivariance can be read by analyzing whether the probabilistic path in the input space and the corresponding latent path result in the same latent representation. The paper includes detailed formal derivation of the model which is realized as a Variational Autoencoder. Experiments are performed on MNIST and 3DShapes datasets.

Strengths

- Solid model derivation - Strong experimental results showing empirical equivariance and disentanglement, though on basic datasets.

Weaknesses

W1) It is not easy to follow the derivations, and it is not clear how the training is performed at the end. It would be helpful to at least add a pseudocode of the algorithm. W2) It is not clear to me what exactly are the input sequences of observations in the experimental section. In other words, how big is T and how to you obtain the intermediate input datapoints between x0 and xT? W3) It is unclear what is the complexity of the model. W4) Experiments are performed on very simple datasets. It is even argued in line 33 that prior methods have not gone beyond toy settings but this work does not do that either.

Questions

1. What is the sequence length and how does this affect the results? Could you please add more details on how you obtain the input sequences? 2. How would this method generalize to more complex data? While the method demonstrates strong results on MNIST and 3DShapes, these datasets are rather basic. 3. In the Extrapolation case, how do you aggregate representations from each of the flows?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

- Limitations are discussed. - No code provided.

Reviewer znKv5/10 · confidence 3/52023-07-06

Summary

This paper proposes a flow factorized representation learning approach for disentanglement and equivariant representation learning in a learned latent space. This work introduces a generative model that defines a set of latent probability paths to specify different input transformations and each latent flow is generated following dynamic optimal transport. Results show that the proposed framework is able to learn both more efficient and more usefully structured representations and thus achieve improved performance on representation learning benchmarks.

Strengths

1. This work provides a new understanding of the disentanglement and equivariant, which are important problems for representation learning research. 2. This paper proposes flow factorized representation learning which is a new and reasonable viewpoint for structured representation learning. 3. Visualization results demonstrate that the proposed framework is able to switch transformations during the latent sample evolution and achieves both disentanglement and equivariance. 4. Experimental results show that the learned latent transformations have flexibility and can generalize to new data.

Weaknesses

1. The datasets used in this work are toy (e.g., MNIST, Shapes3D). Large-scale and real-world datasets are expected to evaluate the efficacy of the proposed approach, such as CIFAR, ImageNet. 2. Advanced baseline methods are expected to be compared including StyleGAN, StyleGAN2, and Diffusion models. 3. Can the learned disentangled representations be transferred across datasets, such as from Shapes3D to MNIST?

Questions

Please refer to detailed comments in weaknesses.

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

This works shows no potential negative societal impact.

Reviewer xnNs6/10 · confidence 2/52023-07-06

Summary

This paper proposes the Flow Factorized Representation to learn structured representations. It introduces a generative model that defines different input transformations with a set of latent flow paths. The latent evolution is generated by the gradient flow of learned potentials, which are parameterized by a set of MLPs. A PINN loss is introduced to ensure that the latent posterior flow follows the OT path. The proposed method is tested on the MNIST and Shapes3D datasets.

Strengths

- The paper is well-written, and the proposed method is technically sound. - The learned transformation are well factorized and demonstrate flexible composability.

Weaknesses

- The paper lacks a comparison in terms of disentanglement. Since the authors claim that the proposed method brings new understandings to both disentanglement and equivariance, it would be beneficial to compare it with state-of-the-art disentanglement methods. As traditional disentanglement metrics in VAE are unsuitable for this method, the VP metric [a] and the correlation metric [b] could be helpful. Additionally, a qualitative comparison would also be beneficial. - More state-of-the-art disentanglement methods should be included in the comparison, such as [c, d]. [a] Zhu, Xinqi, et al. "Learning disentangled representations with latent variation predictability." ECCV 2020. [b] Shen, Yujun, et al. "Interfacegan: Interpreting the disentangled face representation learned by gans." TPAMI 2020. [c] Estermann, Benjamin, et al. "DAVA: Disentangling Adversarial Variational Autoencoder." ICLR 2023. [d] Shao, Huajie, et al. "Controlvae: Controllable variational autoencoder." ICML, 2020.

Questions

- What value is $k$ set to in the experiments? If $k$ is larger than the actual factors of the datasets, how will the model perform?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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 xnNs2023-08-18

Response to the author rebuttal

Thanks to the authors for their efforts. I have read the rebuttal and the comments from the other reviewers, and most of my concerns have been addressed. It's good to see the added disentanglement metrics and the qualitative comparisons. I also recommend to include these in the revised paper. After consideration, I decided to raise my score from 5 to 6.

Authorsrebuttal2023-08-18

Thanks for raising the score!

Thanks for raising the score and the encouraging feedback! We would definitely include the extra experiments and the results of disentanglement metrics in the revised paper. We sincerely appreciate your constructive suggestions for improving our paper.

Reviewer ww476/10 · confidence 3/52023-07-08

Summary

The authors study disentangled and equivariant representation learning as the authors claim existing approaches are either ill-specified or can not effectively separate realistic factors of interest in the learned latent space. The key idea is to define a distinct set of tangent directions that associate different transformations. This is achieved by using flow factorized variational autoencoders, and encouraging each latent posterior flow to follow the optimal transport path, which is accomplished by using the PINN loss (equation 14) based on the Hamilton-Jacobi equation with an external driving force.

Strengths

The authors provide an alternative viewpoint at the intersection of disentangled and equivariant-based representation learning, and propose the flow factored representation learning which targets to define distinct sets of latent probability paths that explains different input transformations. The authors designed a sequential variational autoencoder with latent posterior flow following the OT path. This is achieved by making the gradient satisfy some PEDEs.  The graphical model, illustration and derivation overall look sound to me (I did not check Theorem 1). The authors also examine their proposed methods in two established datasets (MNIST and Shapes3D). In the two datasets, the proposed methods achieve clear benefits in terms of log-likelihood and equivariance error. As the proposed method is targeting for learning equivariant representations, it is expected that the method does achieve lowest equivariance error; Meanwhile the latent flow does perform the target transformation precisely during evaluation while leaving other trains unaffected according to the qualitative study.

Weaknesses

The methods study the intersection of disentanglement and equivariance, but quantitative results associated with disentanglement are missing. It would still be interesting to evaluate the proposed methods in terms of some major disentanglement metrics (e.g., MIG,DCI,Modularity) when compared with other existing approaches. Also, it would be a good test to benchmark the proposed methods in some other well-established datasets like dSprites, Natural, Falor3D and Issac3D, to examine equivariance error, some disentanglement metrics, reconstruction, or some other downstream tasks.

Questions

In equation 6,  will t=0 be problematic? In page 4 (right before equation 9), it’s unclear from the notation if you are defining one Neural potential for each value in the category, or you are using only one MLP.  The external force appeared abruptly in equation 13 without any introduction, and any details on neural parameterization (e.g., what does MLP looks like) are missing. Looking at equation 14, it’s unclear why an independent MLP (the f(z, t)) would affect the potentials (those u^k(z, t) which are also defined by MLP). According to table 1 and 2, the weakly supervised approach still outperforms other baseline methods (which is good), but are consistently worse than supervised ones (expected). It would be interesting to see if semi-supervised learning could improve over supervised one.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Do not identify any negative societal impact.

Reviewer tD2D6/10 · confidence 3/52023-07-26

Summary

This paper proposes a novel approach to unsupervised representation learning; More precisely, the authors focus on the learning of latent dynamics with the aim to design an equivariant mapping from the observational to the representational space; Authors leverage variational posterior approximation and dynamic optimal transport to do so and evaluate their method on MNIST (augmented with standard transformation - rotation, scaling, ...) and dSprites showing a significant log-likelihood and equivariance error improvement in comparison to SlowVAE and TVAE; In addition, authors highlight the flexibility of the approach by showcasing the ability of the trained VAE to generate coherent dynamics in the pixel space with superposed transformations and transformations shifts; Whilst the motivation towards learning equivariant models as well as the idea of learning disentangled latent dynamics through variational posterior approximation is not new, here authors propose a novel set of assumptions over the posterior ($q(z_t|z_{t-1},k)$) and use optimal transport to efficiently train the VAE;

Strengths

My expertise in optimal transport being limited, I will mostly focus my evaluation on the novelty of the proposed contributions, the soundness of the ELBO objective proposed as well as the assumptions considered; - I genuinely enjoyed reading this paper and found the contribution exciting. The manuscript is well-written, mathematical derivations are sound, and the experimental setting is clear; The fact that authors show a competitive performance of the model whilst considering only the information from element $x_0$, the first observations of a sequence, in their posterior factorization is exciting. This structural assumption sounds to me more intuitive and sound than the ones I encountered in previous works. - Whilst the datasets considered in this paper remain quite simple, as generally observed in this sub-field, the improvements in terms of log-likelihood and equivariance error (standard evaluation metrics) are significant and the model is general enough to allow for nice properties to arise (i.e., superposition of transformation, transformation shift);

Weaknesses

- The weakness of this work lies for me in the lack of diversity in the experimental setup considered as further detailed in the limitations sections.

Questions

- Could authors clarify how the number of transformations, $k$ was chosen in the experimental work and whether an ablation over this hyperparameter was performed? - Could authors discuss the computational complexity of the proposed approach and provide some insights on how this compares to the baselines considered?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

- As for most work in this sub-field (c.f., SlowVAE, TVAE), the experimental work is limited to simple datasets (i.e., MNIST, dSprites) and does not investigate how these methods perform on real-world images/videos;

Reviewer wDok2023-08-14

Thanks for the comments

Thanks for providing the thorough rebuttal. You addressed all my questions and I particularly appreciate the extra experiments on the two more complex dataset. I would recommend to add these to the main part of the paper if possible as they nicely highlight the strengths of the method.

Authorsrebuttal2023-08-14

Thanks for the encouraging feedbak; We would definitely include extra experiments.

Thanks for the encouraging feedback! We would definitely include the results and analysis of the extra experiments in the revised paper. We sincerely appreciate your thoughtful review and valuable comments, as well as the time you spent on our paper!

Reviewer tD2D2023-08-15

Rebuttal Acknowledgment

Thank you to the authors for their additional insights and results in particular for the positive results on more complex datasets which provide a more realistic view of the impact of the method; Could authors clarify how the baseline methods were tuned to obtain results shared in the rebuttal answer tables? After consideration, I decided to keep my score; Whilst the additional results provided by the authors increase the quality of the contribution, I believe the definition for a score of 6 (i.e., "Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.") matches my perspective on this work;

Authorsrebuttal2023-08-15

Thanks for the positive feedback!

Thanks for the positive feedback and the appreciation of the additional insights and results in the rebuttal. We respect your opinion on our paper. For the baseline methods, we take the official training codes of TVAE and PoFlow, and use the same VAE architectures for all the baseline methods. We also take the same MLP architectures of PoFlow for parameterizing the potentials $u$. The latent space is divided into 18 capsules each of 18 dimensions for TVAE, and the other two methods are defined with latent spaces of the same sizes ($18\times18=324$).

Reviewer ww472023-08-18

Thank you for your rebuttal.

Thank you to the authors for your effort and the additional good results. I believe including the latest results do make the paper with higher quality. I also read the opinions from other fellow reviewers. Some fellow reviewers mentioned comparing with powerful generative models like diffusion models, which I’m sure is urgent for the scope of this paper. After consideration, I prefer to keep my score as it is (as I’m hesitant to raise my score further to 7, which means a high impact and good-to-perfect evaluation).

Authorsrebuttal2023-08-18

Thanks for the positive feedback!

Thanks for the appreciations on our rebuttal. We would definitely include the latest experimental results in the revised paper. We respect your opinion/score on our paper, and we also agree that for better applicability the subfield of equivariant representation learning should try extending to more powerful generative models such as diffusion models. As discussed in the response to $\textcolor{green}{Reviewer\ znKv}$, despite the existing challenges in extending our framework, we would keep investigating this direction in future work!

Reviewer znKv2023-08-21

Response to the author rebuttal

Thanks to the authors for the rebuttal. Most of my concerns have been addressed. I decided to raise my score.

Authorsrebuttal2023-08-21

Thanks for raising the score!

Thanks for raising the score and the positive feedback! We appreciate your constructive suggestions and valuable comments.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC