Gacs-Korner Common Information Variational Autoencoder

We propose a notion of common information that allows one to quantify and separate the information that is shared between two random variables from the information that is unique to each. Our notion of common information is defined by an optimization problem over a family of functions and recovers the Gács-Körner common information as a special case. Importantly, our notion can be approximated empirically using samples from the underlying data distribution. We then provide a method to partition and quantify the common and unique information using a simple modification of a traditional variational auto-encoder. Empirically, we demonstrate that our formulation allows us to learn semantically meaningful common and unique factors of variation even on high-dimensional data such as images and videos. Moreover, on datasets where ground-truth latent factors are known, we show that we can accurately quantify the common information between the random variables.

Paper

Similar papers

Peer review

Reviewer cgd17/10 · confidence 3/52023-06-30

Summary

The authors present a concept of common information that assesses and distinguishes shared information from unique information between two random variables. This notion, which is a variant of the Gacs-Korner common knowledge, is more optimizable and experimentally approximable when employing sample data.

Strengths

I believe this paper made a good contributions by proposing to relax the requirement that the representation must be a deterministic function and instead allows it to be a stochastic mapping. Based on the results, it has been shown that such an approach is more beneficial for quantifying and interpreting the latent representation.

Weaknesses

- The authors experimentally demonstrating that adding unsupervised viewpoints improves disentanglement; which is likely due to limitations in single-sample analysis; active interaction, not passive observation, fosters better learning of environment representations. but the observation has not been sufficiently explained or explored in the current study.

Questions

- Could you describe how the multi-view VAE model's unique latent variables capture the individual factors that cause variation, and how this contributes to the efficiency of inferring common and unique components from data?

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors address the limitation adequately.

Reviewer iAUJ5/10 · confidence 4/52023-07-07

Summary

The basis of the proposed work is a practical means to decompose the information contained in two random variables into common and unique components. With a couple tweaks to a vanilla VAE setup, data that has been paired such that certain factors of variation have a constant value within the pair can be used to train latent spaces that separate the components of the information. The common component of the information -- the Gacs-Korner common information -- is the primary goal here, and it’s generally difficult to acquire because it requires finding a random variable that is simultaneously a deterministic function of both input variables. Except for highly constrained joint distributions, such a variable generally does not exist (beyond the vacuous solution). The current work proposes a relaxation where the common component can be a stochastic function of the inputs, and this facilitates optimization as well as makes the method applicable to cases without cleanly shared information (at least in theory; it’s not shown here). As a route to Gacs-Korner common information, the paper offers a principled and pragmatic methodology that (as far as I can tell) offers something original and of value. The example scenarios and applied metrics, however, paint the work more as a route to partially disentangled latent spaces given grouped data, and in that regard it’s lacking novelty and missing relevant comparisons. The following papers leverage the same weak supervision (paired/grouped data with a subset of factors held constant) to learn latent spaces that separate common from unique information: - “Multi-Level Variational Autoencoder: Learning Disentangled Representations from Grouped Observations”, Bouchacourt 2018 - “Disentangling Factors of Variation with Cycle-Consistent Variational Auto-Encoders”, Jha 2018 - “Unsupervised Robust Disentangling of Latent Characteristics for Image Synthesis”, Esser 2019 - “NestedVAE: Isolating Common Factors via Weak Supervision” Vowels 2020 And one that does so without a generative model: - “Learning Disentangled Representations via Mutual Information Estimation”, Sanchez 2020 In summary, the method is an interesting contribution but the experimental results are not very effective.

Strengths

The principled route to extracting the common information in a pragmatic methodology is great. Though it is not demonstrated in this work, the authors motivate the work in terms of multi-modality data, which could be a rich area of application of the method.

Weaknesses

See the summary for the primary weakness. I can imagine a couple routes to make the experimental results more strongly support the method. First, the current experiments could include direct comparisons to some of the attached methods with a demonstration of what GKVAE brings. However, if GKVAE is not more effective in the scenarios currently in the manuscript, qualitatively different experiments could help show its merit. Paired data streams from different modalities could highlight the strengths of GKVAE, or perhaps an example where the content of the common information allows for insight about the relationship between the variables (ie, through inspection of the learned latent variable). Or some example where GK information is desirable. On a much more minor scale: I doubt the majority of readers will have a better intuition for “usable information” than for standard mutual information; it seems a weakness to me to rely on that rather than any of a number of mutual information estimators or bounds that should work fine on the low-entropy examples of this work.

Questions

- As common information is a lower bound to mutual information (L73), how can a lower bound to MI be used as a lower bound for common information (L194)? - Was $\lambda_c$ shown to be necessary given the random swapping of the common representation (L225)?

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Yes

Area Chair XDqo2023-08-18

Acknowledging the rebuttal

Dear reviewer, Thank you for your time and effort. The authors have tried to address you comments in their rebuttal. What do you think about their response? Could you please acknowledge the rebuttal as well as the other reviews. Thanks, The AC

Reviewer hfAm5/10 · confidence 3/52023-07-07

Summary

The paper proposes the Gracs-Korner Common information to measure common information between two random variables and propose a variational relaxation to compute the practical loss. The paper then derives an objective that eases disentanglement requirement.

Strengths

1. The paper proposes the Gracs-Korner Common information to measure common information between two random variables and propose a variational relaxation to compute the practical loss. To this point, the paper is novel to me. Empirical evidence demonstrates the advantages of the proposed method.

Weaknesses

1. The proposed method is only compared with basic VAE where as the advantage of the proposed method is not significant. 2. The proposed method seems to be loosely connected with Gacs-Korner Common Information. The derivation is a minor variation of a vanilla VAE. In this regard, it seems the impact of the work is limited, where the modification introduces additional encoder architecture but only brings limited empirical advantage. The final loss further decompose the distance between prior and posterior inference into “common components-prior distances” and the “unique components-prior distances". The driven force that separates the DNN to learn the decomposition remains unclear.

Questions

Please see weakness above.

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

3 good

Limitations

Yes.

Area Chair XDqo2023-08-18

Acknowledging the rebuttal

Dear reviewer, Thank you for your time and effort. The authors have tried to address you comments in their rebuttal. What do you think about their response? Could you please acknowledge the rebuttal as well as the other reviews. Best, The AC

Reviewer uZkW5/10 · confidence 4/52023-07-15

Summary

This paper proposes a new way to partition common and unique information in multi-view data, by leveraging and optimizing the objective of common information defined by Gacs and Korner. Authors extended the deterministic setup to stochastic scenario and used a variational autoencoder to realize the objective. Authors also carefully designed experiments in both static and time series data to validate the effectiveness of their architecture.

Strengths

1. The introduction of the Gacs-Korner common information to partition common and unique information in realistic multi-view data is novel. 2. A straightforward application of Gacs-Korner common information is hard. The stochastic relaxation made by authors makes sense to me. It is also good that authors constructed different datasets to validate the effectiveness of their method.

Weaknesses

I found the biggest issue of this manuscript is that some points are very unclear or need more explanations. Below I listed a few of them: 1) Can you elaborate more on the difference between common information and mutual information. I can understand that mutual information has no clear interpretation in terms of information decomposition. It would be much better to describe more differences. For example, in Eq.~(15), what would be the mutual information between $x_1$ and $x_2$; and why mutual information should be much larger than common information $z_c$? Are there more illustrative examples? 2) In terms of implementation, I appreciate that authors relax the constraint of $Z=f(X_1)=g(X_2)$ with a conditional divergence minimization term $D(p(z|x_1);p(z|x_2))$. How to implement this term in your VAE objective? 3) Also in implementation, it requires that $\beta_u > \beta_c$. How to balance the trade-off between $\lambda_c$ and $\beta$? In Eq.~(13), there should be two $\beta_u$ corresponding to two views, i.e., $\beta_{u1} KL(q_{\phi_{u1}}||...)+ \beta_{u2} KL(q_{\phi_{u2}}||...)$? 4) Proposition 3.1 requires an invertible mapping from $z$ to $x$. How is this condition reflected in your implementation?

Questions

Please refer to weaknesses above.

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Authors did not discuss potential limitations and negative societal impacts.

Reviewer uZkW2023-08-17

Thanks for your rebuttal. I do not have more concerns. It is good that authors use to two illustrative examples to demonstrate the merits of CI over MI. I suggest including this discussion also in the main text. My score remains the same, since I agree with Reviewer iAUJ that there are growing number of literatures on multi-view (disentangled) representation learning, which may also define their own ways to separate unique and common information. I understand that those literatures may suffer from a rigorous definition on what is common and what is unique. However, a comparison to state-of-the-art can enhance the quality of this paper.

Reviewer iAUJ2023-08-18

Thank you for the effort put into the rebuttal and the clarity on the role of $\lambda_c$ provided by the additional experiments. After consideration of the rebuttal and the other reviews, I am still of the opinion that the premise of this work is interesting and valuable but the experimental support is ineffective. I strongly recommend more relevant baselines on the current experiments and/or use cases where the distinction between common and shared/mutual information is more meaningful -- even a simple synthetic one as suggested in the above response would be helpful. GK common information is far enough off the beaten path that the onus lies heavily on the authors to demonstrate why the goal is worthwhile, especially in the midst of a variety of other methods that accomplish qualitatively similar separation of information.

Authorsrebuttal2023-08-22

Thank you for your comment and feedback! In line with the use cases we described in the above response, we plan to include a supporting synthetic experiment to better highlight the use cases and the difference between Gacs-Korner common information and mutual in the camera-ready version.

Reviewer hfAm2023-08-20

thanks for your reply

Thanks for the reply! I will maintain my score. Best

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC