Faster Relative Entropy Coding with Greedy Rejection Coding

Relative entropy coding (REC) algorithms encode a sample from a target distribution $Q$ using a proposal distribution $P$ using as few bits as possible. Unlike entropy coding, REC does not assume discrete distributions or require quantisation. As such, it can be naturally integrated into communication pipelines such as learnt compression and differentially private federated learning. Unfortunately, despite their practical benefits, REC algorithms have not seen widespread application, due to their prohibitively slow runtimes or restrictive assumptions. In this paper, we make progress towards addressing these issues. We introduce Greedy Rejection Coding (GRC), which generalises the rejection based-algorithm of Harsha et al. (2007) to arbitrary probability spaces and partitioning schemes. We first show that GRC terminates almost surely and returns unbiased samples from $Q$, after which we focus on two of its variants: GRCS and GRCD. We show that for continuous $Q$ and $P$ over $\mathbb{R}$ with unimodal density ratio $dQ/dP$, the expected runtime of GRCS is upper bounded by $βD_{KL}[Q || P] + O(1)$ where $β\approx 4.82$, and its expected codelength is optimal. This makes GRCS the first REC algorithm with guaranteed optimal runtime for this class of distributions, up to the multiplicative constant $β$. This significantly improves upon the previous state-of-the-art method, A* coding (Flamich et al., 2022). Under the same assumptions, we experimentally observe and conjecture that the expected runtime and codelength of GRCD are upper bounded by $D_{KL}[Q || P] + O(1)$. Finally, we evaluate GRC in a variational autoencoder-based compression pipeline on MNIST, and show that a modified ELBO and an index-compression method can further improve compression efficiency.

Paper

References (38)

Scroll for more · 26 remaining

Similar papers

Peer review

Reviewer tpKF6/10 · confidence 3/52023-07-02

Summary

The paper presents a coding scheme for greedy rejection sampling by Harsha, which builds upon the existing A* coding framework. The trick here is to add a partition operation between each reject/accept steps, which allows a more efficient scheme under certain assumptions. Overall, the work contributes meaningfully to the field of rejection sampling. Minor adjustments to the figure margins are needed for improved readability.

Strengths

The proposed method and analysis demonstrate the potential for achieving faster runtime under certain assumptions about the proposal and target distributions. The experiment shows the benefits of the method. The paper is relatively easy to follow.

Weaknesses

While the paper is generally well-written, the lack of sketch proof hinders the ease of understanding why the proposed coding scheme works or how the assumption matters. Including a sketch proof would greatly enhance the understanding of the presented approach. The differences between the present algorithm and prior work on A* coding must be explained more clearly. It would be beneficial if the author could comment on the complexity of integrating equation (2) with distributions beyond the Gaussian, thus expanding the scope of the analysis.

Questions

Can the authors explain the implication of assumption 2&3? Comparing the A* coding, I do not follow how GRCS and GRCD can outperform A* coding in terms of runtime complexity. Line 274: I do not see where in the appendix factor 2 (in the log term) does not appear in the coding length. Also, 11 bits in O(1) seem to be longer than the bound by Li and El Gamal but A* coding and GRC seem to be on par in Figure 4 (right). Can the author provide some additional justifications for this?

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

n/a

Reviewer BHnG5/10 · confidence 1/52023-07-07

Summary

The authors propose a greedy rejection coding based algorithm for performing compression. The authors claim that GRC an ERC based compression method is the first method which does not make unrealistic assumptions, is guaranteed to finish in fixed amount of time. They generalize the algorithm to arbitrary partition spaces and provides guarantees to finish in limited runtime. Disclaimer - The reviewer is not the best person to evaluate this paper. They have no experience in this area.

Strengths

* The paper is reasonably well write and is able to explain the challenges very well. * The proofs by and large look correct, however the reviewer can not comment on their novelty *The authors have some experiments.

Weaknesses

It would have been good to see how their method works at even higher dimension., can they compress say other models using their method.

Questions

Look at weakness sections

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have highlighted some limitations.

Reviewer ULav5/10 · confidence 1/52023-07-17

Summary

This paper presents a new algorithm, Greedy Rejection Coding (GRC), for Relative Entropy Coding (REC). REC is an alternative to entropy coding and quantization, which naturally interfaces with continuous variables and has applications in data compression. Current REC algorithms have faced challenges like slow runtimes or restrictive assumptions, hindering their widespread use. The new GRC method, along with its two variants GRCS and GRCD, generalizes a previous rejection sampling-based algorithm to arbitrary probability spaces and partitioning schemes. The authors demonstrate that GRC terminates almost surely, producing unbiased samples, and that GRCS has an optimal expected runtime and codelength for a wide range of one-dimensional problems. The effectiveness of GRC is showcased in a compression pipeline with variational autoencoders on the MNIST dataset, where it outperforms in compression efficiency.

Strengths

1. The paper introduces a new approach, Greedy Rejection Coding (GRC), to the field of rate-distortion compression (REC). The authors build on the rejection algorithm from Harsha et al., extending it to arbitrary probability spaces and partitioning processes. Theoretical results show that the GRC method is accurate and terminates successfully under certain general assumptions. 2. The authors validate their ideas with a proof-of-concept implementation using Variational Autoencoders (VAEs) on the MNIST dataset. They compare them against previous methods like A* coding. Their experiments validate the theoretical underpinnings of the algorithms and show improvements over prior work. 3. The paper deals with the challenge of coding overhead in compression problems. The authors propose a principled modification to the Evidence Lower Bound (ELBO) and entropy coding of GRCD's indices using a ζ distribution as innovative ways to improve compression efficiency.

Weaknesses

1. **Dependency on CDF:** A major limitation of the proposed Greedy Rejection Coding (GRC) approach is that it requires the ability to evaluate the cumulative distribution function (CDF) of the distribution Q. This could potentially limit the algorithm's practicality in situations where the CDF is difficult or computationally expensive to calculate. 2. **Applicability to Multivariate Distributions:** The paper mentions that the current approaches assume target-proposal pairs over the real line. This might limit the effectiveness of the proposed algorithm when dealing with multivariate distributions, as they are decomposed into univariate conditionals, resulting in additional coding overhead per dimension. 3. **Lack of Generalizability to Other Datasets:** The empirical validation of the approach is based on synthetic data and one real-world dataset (MNIST). How well the proposed approach would generalize to more complex datasets or tasks is unclear. 4. **Overhead Mitigation Measures:** The proposed methods to reduce the overhead, such as a modified ELBO and entropy coding using a ζ distribution, could add complexity to the algorithm. Moreover, it's not yet fully clear how much improvement they can bring in diverse scenarios. 5. **Efficiency Analysis:** While the paper discusses the runtime of the proposed algorithm compared to existing methods, it does not provide a detailed computational efficiency analysis. Therefore, it might be unclear how the proposed algorithm would perform under different computational resources.

Questions

Please see the Weaknesses.

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

This work does not pose any ethical concerns.

Reviewer t3AN6/10 · confidence 3/52023-07-26

Summary

This paper proposes new Greedy Rejection Coding methods that generalize to continuous probability space and arbitrary partitioning schemes. The new methods produce correct samples with the expected code length of $O(D_{KL}[Q||P])$ where $Q$ is the target distribution and $P$ is the proposal distribution. The paper then evaluates the new algorithms and verifies the improvements in runtime in practical experiments.

Strengths

- The proposed methods achieve the optimal runtime up to a constant order while guaranteeing the optimal expected codelength. - The experiment results also verify the theory. The runtime of proposed methods (in terms of the number of steps) stays constant when $D_\infty[Q||P]$ varies and is always smaller than the runtimes of previous methods AS* and AD*. - Comparisons to previous results, technical novelties, and limitations are thoroughly discussed. - The paper also provides plenty of high-level discussions and visualizations of their results.

Weaknesses

- I think some intuitions about why the modifications help improve the runtime and also why the authors decide to make that particular modification can be added to the main text to improve the readability of the paper.

Questions

- Can the author explain why the bound mentioned in line 274 is better than the bound in eq. 10? - What is the wall-clock runtime of the proposed methods? It's clear that the proposed methods significantly improve the runtime in the number of steps but I'm curious if some of the extra steps in the algorithms could cause overhead in terms of wall-clock runtime/step. - There's a typo in line 321: "GRC can become significant when the KL over Hence..."

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Authorsrebuttal2023-08-17

Request for the Reviewers to Respond to Our Rebuttal

We thank the reviewers again for the feedback they provided on our work. As the author-reviewer discussion period ends on 21 August, we kindly ask the reviewers to let us know if our rebuttal has addressed their concerns adequately as soon as possible so we may respond to any outstanding issues on time. Additionally, if our rebuttal has effectively addressed the reviewers’ concerns, we kindly invite them to acknowledge it and consider raising their scores. Thank you once again for your time and feedback. Thank you, Authors of paper 5337

Reviewer t3AN2023-08-19

Thanks for the response! I will keep my score.

Reviewer tpKF2023-08-21

Thank you for the response and clarifying the questions. I will keep my score to 6.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC